The Roads We Build: AI as Social Infrastructure

By Kato Mivule, D.Sc. | A Review of Sloane (2026)’s “Predicted”

Abstract

In this essay, I review Sloane’s (2026) Predicted: How AI Is Restructuring Social Life and consider what insights it offers for Human-Centric AI (HCAI) and pragmatic approaches to AI adoption and ethics. Sloane steps outside the doom vs hype debate and examines prediction itself, framing AI as social infrastructure that people build, use, and reshape. I find her analysis strongest on accountability, transparency, and fairness, and her remedies sound in direction, especially using AI as a transparent tool for narrow tasks. Her ideas still need practical tools, like the right to challenge AI decisions, to give people real participatory power. Yet still, her thesis is even more clear and timely; AI isn’t fate, it’s social infrastructure we can reshape.

Image Source: Amazon, “Predicted”

HCAI Working Framework: The Human-Centric AI (HCAI) framework holds that AI should strengthen human abilities, not replace them. It judges AI against eight core principles: transparency, human agency, fairness, privacy, safety, accountability, inclusiveness, and well-being. The framework does not treat these principles as empty slogans. It asks two practical questions of any AI system in use: do people still keep real control over it, and does it meet a genuine, proven need in society [2]?

Pragmatic Approach: Pragmatism draws on Charles Peirce’s Pragmatic Maxim, which holds that the meaning of an idea lies in the practical effects we can expect it to have. To understand an idea, we ask what difference it would make in practice. Peirce also held that knowledge is never final. Truth is what careful, shared inquiry would settle on over time, so every conclusion stays open to correction as new evidence emerges [3]. For AI, this means judging systems and principles by the effects they actually have once in use, rather than by how closely they follow a fixed set of rules, and correcting course whenever the evidence calls for it.

Concerning the Predictive Model

Sloane (2026) begins by arguing that AI’s most significant effect is not any single application but a quiet restructuring of society around foretelling, which she calls the prediction paradigm. Predictive systems now sit inside hiring, healthcare, lending, and everyday coordination, so engaging with them has become a condition of access to core institutions, and opting out is rarely realistic. Treating these systems as natural or inevitable, she contends, obscures the people and interests shaping them and hands computer scientists outsized authority over social questions. Sloane rejects both technological determinism and the popular “sociotechnical” label, which she believes ranks the technical above the social. She reframes AI as a social arrangement and claims a firm role for social science expertise in governing it [1a].

“…The promises of AI’s oracular power have turned prediction into a logic for structuring social life. This is a dangerous proposition. It implies that AI is always necessary or even inevitable and diverts attention from the social forces shaping ideas around this technology in the first place. AI systems are not natural phenomena that happen to us. They are collective expressions of society…”, Pages 2.

From an HCAI standpoint, the diagnosis is strongest where it removes the black-box explanation; if AI is a social arrangement, harms trace back to identifiable choices rather than opaque mathematics, and legitimate expertise extends beyond computer scientists. Pragmatically, three questions stay open. Framing AI as a collective expression of society could spread responsibility thin unless specific actors are named. By stressing that people can’t avoid AI, she leaves open why hospitals often choose tools like early warnings for sepsis, a life-threatening infection, because they genuinely help. Sloane rejects the ‘sociotechnical’ label for ranking the technical above the social, yet her case for social science risks simply reversing that ranking, which sits awkwardly with her claim that everyone is an AI expert. The challenge for leadership is to treat every predictive system as a negotiable institutional choice rather than infrastructure to be accepted, and should state plainly who owns its decisions, who can contest them, and what non-AI pathway remains.

Concerning the Opaqueness of AI

Sloane (2026) then argues that the “black box” of AI is less a mathematical mystery than a social one. The underlying computations may resist reverse engineering, but the design assumptions about human behavior and social order built into these systems can be identified and examined. Two assumptions dominate; cognitivism, which models machines on the human mind, and anthropomorphism, which stretches that analogy into personhood and feeds both utopian promises and existential fears, such as those voiced by AI doomers about the end of humanity. She shows that users routinely read causation into outputs that rest only on correlation, which gives prediction an oracle-like authority. Because prediction also depends on a socially constructed, linear view of time, Sloane concludes that everyone takes part in building the black box [1b].

“…AI cognitivism and anthropomorphism also reliably bring up a long- lasting fear shared across many cultures: that intelligent machines will soon be sentient artificial persons with superhuman powers. This fear is built and sustained by well- established narrative tropes that oscillate between utopia (“We are going to have so little work to do that we will need universal basic income!”) and dystopia (“The robots will take our jobs and then eradicate humanity!”) …” Page 28.

Measured against HCAI principles, Sloane’s analysis advances Transparency and Explainability by shifting scrutiny from opaque math to inspectable design premises. It’s warning that deployments seek either human compliance or unsupervised automation directly supports Human Agency and Control. Yet claiming that everyone co-creates the black box strains Accountability, since vendors and users hold unequal power. Pragmatically, rejecting anthropomorphism and existential panic keeps AI as an instrument and tool under human direction. The correlation critique, however, underweights practice; clinicians often use correlational risk scores knowingly, as screening aids. Sloane’s linear-time argument, is provocative and has real bite; AI assumes the future repeats the past, which many cultures reject and which breaks down when sudden change hits. Before trusting a predictive tool, leaders should examine what it assumes about people and whether its outputs are correlations being treated as causes.

Concerning AI as a Social Infrastructure

Sloane (2026) goes on to reason that AI now works like social infrastructure. Like roads or libraries, it shapes how people, money, and ideas move through society. Unlike a library, however, it is rarely built to bring people together or help them thrive. It is built instead to hit set targets and to automate work, getting more done with fewer people. She traces this thinking back to nineteenth-century statistics, which assumed society could be understood and managed through numbers. She notes that AI sorts people into categories and measures them against standards that often reflect the majority group, then turns these into fixed rules. People can usually push back on roads and other public systems, or find ways around them. With AI, the categories are hidden, so people such as insurance claimants cannot challenge how they have been sorted [1c].

“…AI is social infrastructure: It affects civic engagement, social interaction, and human flourishing, but it is typically not designed to promote these… Much like other infrastructures, the directionality in AI systems manifests in its design intention and its technical components. A central idea within AI design is optimization… AI’s directionality is embedded in the centrality of the objective function…”, Pages 43-47.

On HCAI terms, Sloane’s view strengthens Fairness by showing how “normal” is often defined by the majority and built into systems. It also shows a gap in Transparency and Accountability; people cannot challenge categories they cannot see. Pragmatically, her point that people can reshape infrastructure, as early online communities and fights against unfair racial labels show, keeps AI open to human change. However, treating AI as nearly closed may overlook existing tools, such as the letters US lenders must send explaining a denial. Yet still, leaders adopting AI should treat its categories like public rules, visible to the people being sorted and open to challenge.

Concerning AI as a Market Force

Sloane (2026) points out that AI is sold less as a tool than as a valuable asset, one whose promised future earnings justify huge company valuations even without current profits. This business rests on three beliefs. The first is that data is an endless natural resource. In reality, data comes from ordinary people going about daily life, and it depends on low-paid, often hidden workers, mostly in poorer countries, who clean and label it. The second is that a job can be automated, a judgment usually based on the social status of the people doing it rather than the real skill involved. The third is that machine predictions are more objective than human judgment, a belief that shrinks the room professionals have to decide for themselves [1d].

“…Additionally, AI is not something you can gain access to by going to a shop down the road. AI is an infrastructure that sits on top and inside of other infrastructures, such as platforms, servers, software, institutions, and processes… The business of AI, therefore, could be understood by framing AI as an asset— a resource that is controlled by an entity and from which future economic benefits are expected to flow back to that entity…” Page 63.

From a HCAI perspective, Sloane’s argument is strong on Fairness and Well-being. It shows AI’s costs falling on hidden data workers and on those whose jobs are labeled “low skill,” and it treats meaningful work and professional judgment as human needs worth protecting. Pragmatically, it grounds AI in money, labor, and energy rather than hype. However, its most useful point for Human Agency, that AI reshapes rather than replaces judgment, gets brief treatment, though that is where oversight can be designed. More importantly, before automating a job, leaders should ask who actually does the work, what skill it truly takes, and how much room for human judgment will remain.

Concerning the Burdens of AI

Sloane (2026) proceeds to discuss that AI’s true costs stay hidden because the companies that own and run it rarely pay them. Low-paid data workers abroad, communities living near energy-hungry data centers, and rural areas where farmland is being bought up all absorb costs that never appear on company books. When AI fails, the burden falls on the people harmed. Students wrongly flagged by unreliable AI-cheating detectors, for example, must prove their innocence, often with no clear way to appeal. She also faults the current focus on “bias,” which fixes problems through statistical adjustments that require sorting people into hidden categories. Even regulators and auditors, she warns, now address AI harms with the same prediction logic that produced them [1e].

“…AI harms are generally understood to be real damage to someone’s life or livelihood resulting from AI predictions… Victims of AI breakdowns carry the extraordinary burden of having to prove that there was an AI breakdown that directly caused their harm— and even if they do have evidence, mechanisms for recourse are not yet established. 17 AI breakdowns are characterized by the burden of proof and the barrier of appeal…”, Pages 86, 88.

On HCAI terms concerning the burdens of AI, Sloane’s clearest contribution is to Accountability. By naming the burden of proof and the barrier to appeal, it gives a precise, testable account of why harmed people lack recourse. It’s warning that bias fixes require sorting people into hidden categories also deepens Fairness and Privacy concerns. Pragmatically, the cheating-detector case is exactly the grounded evidence this lens values, with students lacking recourse, but are viewed as guilty until proved innocent. Before deploying AI, leaders should decide who pays the cost when it fails and give affected people a simple, real way to appeal.

Concerning AI Guardrails

Sloane (2026) contends that efforts to put guardrails around AI are held back by two patterns. The first is treating ethics as something to measure and compute, which shrinks moral questions into scoring how ethical a system is, as the self-driving car version of the trolley dilemma shows. The second is searching for one sweeping solution when AI’s effects depend on specific settings. She reviews how governments have responded; the EU’s risk-based rules, a patchwork of US state and local laws, and a federal shift from oversight toward deregulation. Accountability, she contends, works best when grounded in context, as with tools that tie AI answers to trusted local sources, and when companies open their systems to independent researchers [1f]

“…Some technical approaches to accountability in generative AI systems deliberately use sociality and contextuality. The aforementioned approach of retrieval- augmented generation (RAG) is a notable example, because it forces an LLM to ground its output generation in a predefined data corpus outside its training data… That corpus is typically context- specific…”, Page 126.

From an HCAI viewpoint, Sloane’s reasoning strengthens Accountability and Inclusiveness by arguing that oversight must involve the people in a given setting, not just checklists and scores. Its call for companies to open their systems to independent review supports Transparency. Practically, its preference for context-grounded tools, education, and audits over one grand fix fits the HCAI framework well. However, for nuance, warning against measured ethics may underestimate measurement’s role, since regulators and courts need evidence, they can check. And independent access still depends on company goodwill, with no enforcement mechanism. Moreover, leadership should build AI oversight around the people and settings a system affects, and open it to independent review, rather than relying on a single ethics score or compliance checklist.

Concerning Community in the Age of AI

In her closing chapter, Sloane (2026) observes that AI is now built into everyday life, and that its deepest effect is a collective shift toward prediction as the way society organizes itself. When a few people and systems claim to know the future, public deliberation, the everyday work of weighing different futures together, starts to look inefficient. Her solution is to treat AI as a social arrangement that people can reshape. She recommends using AI as a transparent instrument for narrow, well-defined tasks, as scientists do. She also calls for treating ethics as a social practice, strengthening audits with social science methods, and regulating AI providers as public utilities, all to widen who gets to shape AI [1g].

“…However, as we seek to stay together in the age of AI, we ought to reject the idea of one future and the deterministic power of prediction, and expand our embrace of serendipity, randomness, and sociality. Such a stance chips away at the idea that the future is solely an engineering problem. And it demands that people be put on a more equal footing to begin with— for example, by distributing resources and services fairly…” Page 142.

Judging by the HCAI perspective, the remedies Sloane recommends fit Transparency and Inclusiveness. The scientific-instrument model makes explainability a professional duty, and the public-utility framing could give Accountability real legal weight. Pragmatically, the turn to narrow, well-defined uses is the book’s most usable guidance for builders. The gap is implementation. The public-utility proposal needs a path through US federal regulators who, as Sloane noted, are retreating from oversight. And broad participation still lacks a named mechanism, such as community review boards or procurement rules, that a city council or company could adopt tomorrow. Yet still, a dilemma remains concerning audits with social science methods; Social science has its own measurement disputes and replication problems, and it depends on the same statistics and categories she criticizes in AI. However, leaders should use AI strategically, where tasks are narrow, data is well defined, and people set the context, and should keep decisions about shared futures in open human deliberation.

Reflections on the Roads We Build: AI as Social Infrastructure

AI as Social Infrastructure: Much of the public debate about AI is a contest of predictions. Doomers forecast catastrophe, and utopians forecast abundance. Sloane (2026) steps outside that contest. Instead of offering another forecast, she examines prediction itself; how it works, what assumptions are built into it, and how it has become a way of organizing society. Her central thesis is that AI is social infrastructure. It is built from social material; people’s data, the labor of data workers, and assumptions about how people behave. It runs through institutions and reshapes them in turn. Yet AI is often treated as a purely technical product, a view that profit rewards. Sloane shows how much this view hides, including low-paid labor, energy and water costs, and the burden that falls on people when systems fail. She is careful not to tell this as a story of villains. The turn toward prediction, she argues, is something everyone helps enact; the companies that build the tools, the institutions that buy them, the regulators who audit them narrowly, and the users who trust their outputs.

The Roads We Build: A city’s roads offer a useful illustration. Planners and builders lay the roads, and the roads decide where people can easily go. But roads are also shaped by use. When a route fails people, they wear their own shortcut across the grass, a “desire path,” and good planners eventually pave it. AI works in a similar way. It channels how people are sorted, hired, and served, yet it depends on everyday use to function, and that use can push back. The difference, as Sloane shows, is that AI’s routes are largely hidden. People cannot see the categories that sort them, must prove harm on their own, and often have nowhere to appeal. The power ordinary people hold over AI is therefore real but dormant. It becomes usable only through concrete tool; rights to see and contest decisions, procurement rules, collective bargaining, and local laws.

The Linear Time Mistake: Sloane also exposes a quieter assumption built into AI; that time runs in a straight line, so the future will follow neatly from the past. AI learns from yesterday’s data and assumes tomorrow will look much the same. Many cultures see time differently, as cycles of return and renewal, and real life is full of sudden changes that break old patterns. Weather forecasters offer a better model. They do not promise one fixed future. They give ranges and probabilities, such as a 60 percent chance of rain, and update them as conditions change. AI prediction rarely shows that kind of humility, yet it is increasingly trusted to decide who gets hired, treated, or approved.

The Call for the Return of the Humanities: Sloane’s call for “highly interdisciplinary teams” (p. 127) and for a revival of social science and humanities ideas (p. 143) rings true from inside the field. For decades, American education and the tech industry have prized STEM, and the later push to add the arts, known as STEAM, has struggled to take root, even though Steve Jobs famously credited Apple’s success to joining technology with the liberal arts and humanities. As a computer scientist, I have seen what this narrow focus costs; an industry that sidelines public participation, argues over distant doom or utopia, and overlooks the harms happening now. The full participation Sloane envisions needs the humanities, the arts, and the social sciences back at the table, alongside computer science rather than in place of it.

The Social Science Dilemma: Sloane’s call for social science expertise raises a fair question; does she simply replace one expert class with another [1g]? Read closely, she asks for a seat at the table, not ownership, and she insists that all of us are AI experts. Still, the risk is real. Social science has its own measurement debates and replication problems, and it relies on the same statistics and categories she criticizes in AI. Charles Peirce’s view of knowledge as a shared, self-correcting inquiry points to a better balance. No single discipline should own AI. Computer scientists, social scientists, practitioners, and affected communities each catch mistakes the others miss.

Creativity Inevitability and Randomness: Sloane’s closing call also connects with what I call creativity inevitability. On this view, creativity is a human constant. It runs from the first stone tools, through a Ugandan grandmother’s hand-woven Mukeka mats, to today’s AI, so new technology will keep emerging because people will keep creating [4]. At first glance this seems to clash with Sloane, who warns against treating AI as inevitable. In fact, the two views fit together. What Sloane rejects is inevitabilism; the corporate claim that a particular technology’s path is beyond human control. I reject that too. My proposed framework separates three layers; human creativity is inevitable, the path each technology takes are controllable, and how it is used must be accountable. Sloane’s call to reject one fixed future and to welcome serendipity, randomness, and sociality strengthens the second and third layers, since creativity thrives on the unexpected, not on straight-line forecasts. AI can play a small part here. It can support human creativity, but it cannot replace the chance encounters and open debate that Sloane values most.

Her final chapter, “Staying Together in the Age of AI,” responds to all of this. The aim is not to reject AI but to reclaim a shared say in how it is built and governed, rather than accept a single predicted future. She adds an important condition: for participation to mean anything, people must first stand on more equal footing, through fair access to resources and services.

Sloane (2026) remedies follow from the diagnosis:

• Use AI as a transparent instrument for narrow, well-defined tasks, the way scientists use and document their tools. This is her most practical guidance.
• Treat ethics as a social practice rather than a math problem.
• Strengthen audits by including the views of affected communities.
• Explore regulating AI providers as public utilities, much as societies regulate electricity and water for fair access and pricing.

Overall HCAI and Pragmatic Critique

Read as a whole, Sloane (2026) holds up well under both lenses, with clear limits. On HCAI terms, her strongest contributions are to Accountability, Transparency, and Fairness. She removes the “black box” excuse by showing that the assumptions behind AI can be traced. She names the burden of proof and the lack of appeal that harmed people face, and she shows how standards drawn from dominant groups become hidden rules. She also treats meaningful work, professional judgment, and public deliberation as human needs worth protecting, which speaks to Human Agency and Well-being. On pragmatic terms, she consistently treats AI as a human-made tool rather than a ghost or a force of nature. She rejects both doom and hype and grounds her claims in concrete cases, such as unreliable cheating detectors and hidden data work. The weaker points are practical. Saying that everyone co-creates AI can blur who is responsible when power is so unequal. Her critiques of correlation and of audits underplay how professionals often use risk scores knowingly, and how audits bring hidden harms to light. Her remedies are sound in direction, but they stop short of mechanisms a city council, agency, or company could adopt tomorrow, such as community review boards, procurement rules, or rights to contest decisions. Her clearest practical guidance is the scientific-instrument model: use AI for narrow, well-defined tasks, with transparent methods and people setting the context. Overall, the book offers a strong diagnosis and a sound direction, and it leaves the detailed building work to practitioners, policymakers, and communities.

References

[1] Mona Sloane, “Predicted: How AI Is Restructuring Social Life”, University of California Press, Kindle Version, 2026, ISBN: 978-0520416369.

[1a] Ibid., pp. Pages 1-19.

[1b] Ibid., pp. Pages 22-41.

[1c] Ibid., pp. Pages 43–61.

[1d] Ibid., pp. Pages 62–80.

[1e] Ibid., pp. Pages 82–99.

[1f] Ibid., pp. Pages 100–127.

[1g] Ibid., pp. Pages 129–143.

[2] Mivule, Kato “A Review of AI Engineering: A Human-Centric Perspective,” Mivuletech, Apr. 19, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/04/19/a-review-of-ai-engineering-a-human-centric-perspective/

[3] R. Burch, “Charles Sanders Peirce,” The Stanford Encyclopedia of Philosophy, E. N. Zalta and U. Nodelman, Eds., Spring 2024 ed. [Online]. Available: https://plato.stanford.edu/entries/peirce/

[4] Mivule, Kato. “Creativity and Innovation Are Inevitable: Why the Anti-AI Movement Will Inevitably Fail.” Mivuletech, 23 Aug. 2026, Available: https://mivuletech.wordpress.com/2026/08/23/creativity-and-innovation-are-inevitable-why-the-anti-ai-movement-will-inevitably-fail/

Image Source: Amazon, “Predicted”

Commentary by Kato Mivule

An AI-Human Collaboration

Beyond the Triangle of Doom: Getting AI Right

A Commentary on Drage (2026), “What if We Got AI Right?”| By Kato Mivule, D.Sc.

Abstract

In this commentary, I examine What If We Got AI Right? by Eleanor Drage, a Senior Research Fellow at Cambridge and one of the UK’s leading AI ethicists [1]. Evaluating her work through two corresponding lenses, eight working description and core principles of Human-Centric AI (HCAI) [2], and a pragmatic framework countering AI extinction panic [3] rooted in Peirce’s pragmatic dictum [4], I trace how Drage dismantles AI apocalyptic and utopian myths. While her critique clears away techno-theological rhetoric, I assess the concrete framework left for AI builders, communities, and institutional leaders. The goal is practical; to help AI practitioners and decision-makers separate speculative apocalyptic AI rhetoric from actionable governance, and to invite readers into a clear-eyed, participatory account of what building AI responsibly actually requires [1,3].

Image Source: Amazon, “What if we got AI Right?”

HCAI Working Framework: The Human-Centric AI (HCAI) framework asserts that AI must augment rather than displace human capability, evaluated systematically across eight core principles; transparency, human agency, fairness, privacy, safety, accountability, inclusiveness, and well-being [2]. Rather than treating these concepts as abstract platitudes, the framework examines whether a given deployment preserves meaningful human oversight and addresses demonstrable societal needs [2].

Pragmatic Approach: Following Peirce’s Pragmatic Dictum, pragmatism holds that a concept’s meaning lies entirely in its practical effects, and truth is that which yields positive real-world outcomes while remaining open to revision as evidence changes [4]. Applied to AI, this means evaluating systems and principles by their observed, post-deployment effects rather than by adherence to a fixed rulebook, adjusting course whenever empirical reality demands it [3], [4].

Concerning Existential Catastrophe

Drage (2026) begins by diagnosing a self-reinforcing panic economy where AI is marketed concurrently as an existential catastrophe and a utopian savior. This rhetorical binary directly advances the profitable aims of dominant tech firms racing to construct ever larger automated systems. To counter this, she resituates AI tools within the continuous history of human tool-use rather than treating modern models as exceptional ruptures. She argues that both AI apocalyptic alarmism and hype strip ordinary people of agency by portraying future trajectories as predetermined outcomes rather than democratic choices. Her corrective is civic; she introduces an ideal of public “AI competence” cultivated through widespread technical literacy, mandatory model labelling, and direct civic participation modeled after jury duty or voting. This framing deliberately shifts ethics away from speculative spectacle toward the practical daily labor of sustaining relationships and fostering human flourishing alongside technology (pages 1–21) [1].

From an HCAI vantage point, Drage (2026)’s emphasis on baseline competence and product labelling gestures toward Transparency and Inclusiveness, via conceptual literacy rather than detailing actionable mechanisms such as enforceable consent architectures, external audit rights, and clear redress paths [2]. Pragmatically, dismissing both existential doomerism and corporate hype adheres to sound craft-adaptation reasoning, it treats AI systems as governable artifacts rather than predetermined fate [3], [4]. However, her introductory diagnosis offers critique but does not provide full specified implementation templates, leaving engineers and civic leaders without the procedural scaffolding needed to operationalize her civic vision [2], [3]. This is reasonable and expected, as Drage’s text is not focused as an engineering manual. Nevertheless, Drage provides leadership with a sound premise; AI remains a human artifact subject to deliberate choices rather than an autonomous force to fear or venerate, establishing a mandate for civic participation structures that organizations must now translate into working policy [1].

Concerning How “AI is Peopled”

Drage (2026) dismantles the fabricated separation between human beings and machines by charting a lineage from Prometheus through Bernard Stiegler and Gilbert Simondon. She shows that humanity has continuously co-evolved with its implements; AI is neither alien nor self-sufficient, but completely peopled. It is constructed, labeled, fine-tuned, and maintained by data annotators, engineers, and corporate executives whose deliberate interventions are concealed beneath the catch-all term “automation”. She directs this critique at transhumanist philosophy, characterizing its synthesis of market fundamentalism and secular perfectionism as a narrow, teleological narrative that forecloses of other ways of relating to technology by presenting an optimized, post-biological machine destiny as humanity’s only path. By framing human identity as relational and contextual rather than fixed, she shows that the AI singularity myths silence practical inquiries into what problems these tools actually solve (pages 25–51) [1].

The “AI is peopled” argument is a genuine HCAI contribution, directly reinforcing Accountability and Fairness by bringing hidden supply-chain labor into view [2]. However, while the text identifies culprits ranging from tech executives to corporate-sponsored surveillance systems, it does not detail specific structural solutions [1]. In practice, tracing tool use from myth to transhumanism succeeds at stripping away mystical assertions around machine agency [3], [4]. Yet it remains understandably descriptive; it stops short of translating the insight that technology is human into concrete engineering controls for AI system builders, actions that engineers can distill from her insights [2], [3]. Additionally, Drage successfully returns accountability from abstract inevitability back onto identifiable human actors; this necessitates engineering teams to still require actionable guardrails to weave labor accountability into production pipelines [1].

Concerning the AI Apocalypse and Utopia Framing

Moving from philosophical critique to institutional history, Drage (2026) connects Nick Bostrom’s existential risk models to his earlier 1990s writing on racial intelligence hierarchies, asserting that his “crunch” theory conceals eugenicist assumptions regarding which populations represent human progress. She examines the 2023 Future of Life Institute pause letter, observing its homogeneous demographic makeup and the irony of AI developers assuming authoritative ownership over threats generated by their own business models. She contrasts this performative “emergency” with the corporate dismissal of research ethicists Timnit Gebru and Margaret Mitchell, followed by subsequent industry-wide cuts to internal safety and red-teaming groups. Her central conclusion remains clear; manufactured crises regarding hypothetical apocalyptic futures allow AI entities to evade accountability for immediate societal harms while sidelining social justice concerns (pages 51–71) [1].

This critique strengthens the HCAI Accountability pillar by uncovering a critical operational contradiction; organizations loudly performing abstract AI safety research while disbanding the exact internal units chartered to audit real-world harms [2]. Practically, examining the historical documentation behind Bostrom’s AI models dismantles claims of value-neutral risk science [1], [3]. Decision-makers should view performative corporate panic over AI existential risk with deep skepticism, particularly when the institutions raising these alarms simultaneously defund teams working on verified present-day AI risks [1].

Concerning the Triangle of Doom

Drage coins the “Triangle of Doom” to describe the fusion of utilitarianism, effective altruism, and existential-risk ideas among tech entrepreneurs, arguing this combination produces closed-door utilitarian calculus that weighs present, real deaths against speculative future gains from artificial general intelligence (AGI). She extends this into a sustained critique of Bostrom, showing how his valuation of “intelligent life” survival above all else renders ongoing violence, genocide, war, climate catastrophe, comparatively trivial, and how his star-colonization economics erases geopolitics entirely. Widening the frame, she insists existential risks cannot be separated from the injustices generating them, and eventually names capitalism itself, not rogue AI, as the primary engine of harm and existential risk (pages 72–92) [1].

The “Triangle of Doom” formulation directly advances Fairness and Accountability by exposing speculative risk calculations as ideological choices masked as quantitative math [1], [2]. In addition, elevating the lived perspective of affected communities exemplifies Inclusiveness in research [2]. Pragmatically, identifying economic incentives as the root driver anchors the debate in empirical realities [3], [4]. However, the analysis ends at macroeconomic critique, omitting operational countermeasures that technical leads can implement to insulate product safety from quarterly financial pressures [2]. However, strategic leaders should recognize that risk-reward calculations framed as objective ethics often encode whose lives are considered expendable, and that addressing AI harm requires confronting underlying economic incentives, not just technical safety measures.

Concerning AI Representations

Drage (2026) turns to representation, arguing that ubiquitous AI imagery, Michelangelo’s Creation of Adam and Rodin’s Thinker, smuggles Judeo-Christian creator mythology and a white male template of intelligence into how AI is publicly imagined, legitimizing entrepreneurs’ claims to God-like creation while erasing whose labor actually builds these systems. She contrasts this with the Better Images of AI project, which depicts data labelers and material infrastructure instead, and extends the material argument to AI’s climate footprint through concrete energy and water consumption figures. She then dismantles “machine consciousness” and “machine autonomy” as concepts that mystify distributed, infrastructure-dependent systems into false independence, insisting that every AI interaction is actually a visible network of engineers, labelers, and data contributors rather than a solitary thinking entity (pages 94–111) [1].

The critique of representation directly serves Transparency. Substituting mystifying metaphors with depictions of physical compute infrastructure and manual data curation offers a tangible practice for technical organizations and media outlets. Practically, demystifying terms like “consciousness” and “autonomy” dismantles technological hypostasis by correctly reframing AI as tool-assisted craft dependent on human labor chains. Leaders and communicators should replace God-like or brain-like AI imagery with depictions of the actual human infrastructure and material costs behind these systems, since language shapes both public perception and who gets held accountable.

Concerning Anthropomorphic Language

Drage (2026) then demonstrates how anthropomorphic terminology such as “hallucination,” “vision,” and “comprehension” masks software failures and manual labor beneath an illusion of cognitive agency. She references Margaret Mitchell’s proposed technical reframing, such as clarifying that a model merely outputs artifacts perceived by humans as visual imagery, as a linguistic correction that returns intent to human operators. She links this to weak marketing regulation, arguing companies exploit consumer uncertainty through unchecked, performative claims that actively shape belief rather than merely describing capability. Drawing on Karen Barad and Donna Haraway, she then dismantles the deeper myth of scientific and AI objectivity, tracing it to a nineteenth-century Kantian ideal of the self-erasing observer, and argues that treating AI as neutral is itself a rhetorical convenience that simplifies selling it while masking the contextual, human-embedded nature of all technical knowledge (pages 114–145) [1].

Mitchell’s reframing proposal is an instance of concrete, implementable scaffolding, directly actionable for engineers and communicators seeking Transparency and Explainability compliance. In practice, this historical unmasking of algorithmic neutrality operates as a rigorous evaluation, accurately framing claims of “neutral AI” as a marketing construct. However, Drage does not detail specific regulatory enforcement strategies, leaving open the question of what incentives will motivate commercial vendors to adopt sober language in competitive markets. Organizations should adopt precise, non-anthropomorphizing language about what their systems actually do, both because it is more honest and because “neutral” or “objective” framing is a liability once its rhetorical function is exposed.

Concerning Intelligence and AI Alignment

Drage (2026) moves on to dismantle intelligence-testing logic underlying AGI, showing IQ-style benchmarks import the same eugenicist, decontextualized assumptions that plague human intelligence tests, with no genetic basis found for intelligence despite persistent claims linking it to race. She argues true intelligence is contextual and specific, proposing Haraway’s “planetary intelligence”, cooperative, ecosystem-oriented, over generalized AGI ambitions. Turning to “the control problem,” she reveals alignment rhetoric smuggles in unstated, often Silicon Valley-specific values as universal; she calls instead for community-led, locally bespoke co-management rather than top-down value homogenization (pages 146–164) [1].

The planetary-intelligence and community co-management proposals meaningfully advance Inclusiveness and Participation, offering an actual alternative governance structure rather than pure critique. Pragmatically, unmasking “alignment” as value imposition disguised as neutrality is strong anti-hypostasis work; however, “community-led co-management” remains underspecified, the text names the destination without addressing jurisdictional conflicts, resourcing, or enforcement mechanisms practitioners would need to implement it. Technology leaders should treat “AI alignment” claims skeptically as vehicles for particular institutional values rather than universal ones, and pursue locally accountable governance instead of importing Silicon Valley’s ethical defaults wholesale.

Concerning the Feminist Ethics Approach

Drage (2026) critiques “cookie-cutter” AI ethics principles as popular yet empirically disconnected from actual improvements in workforce inclusion or reduced harm, proposing instead that genuine thoughtfulness requires self-reflective, labored consideration rather than status-quo mimicry. She then interrogates consent as a governing mechanism, tracing its legal codification into GDPR and arguing that despite offering some definitional consistency, consent mechanisms in practice coerce agreement to poorly understood terms under conditions users never helped design. Her core claim is that consent, like AI’s supposed neutrality, obscures underlying power imbalances; individuals retain the technical act of clicking “agree,” but tech companies unilaterally construct the choice architecture itself, deciding for users rather than with them how rights are protected (pages 165–175) [1].

This critique directly ties legal language (GDPR) to lived power asymmetry rather than abstract theorizing; however, it stops short of proposing what a comprehensive redesigned, power-shifted consent mechanism would concretely look like. However, organizations should recognize that checkbox consent satisfies legal form without delivering meaningful agency, and that real governance requires involving users in designing the choice itself, not merely in clicking through it.

Concerning Bias, and Diversity Anxiety

Drage (2026) then dissects “bias” as a term whose mathematical and social meanings diverge so sharply that engineers are left without workable guidance, arguing that framing bias-reduction against “innovation” is a false, profit-serving dichotomy. Her constructive turn showcases Reparative AI, Davis, Williams, and Yang’s approach of deliberately unequal treatment to correct historical harm, alongside DIY, community-built tools designed for specific underserved users rather than universal scale. She closes by confronting “diversity anxiety” in tech teams, showing how genuine representation is trivialized or resisted even as evidence of ongoing exclusion (e.g., in casting) remains plainly visible (pages 176–204) [1].

Reparative AI and localized civic tools stand out as actionable design patterns within Drage’s text, offering direct routes to bolster HCAI Fairness and Inclusiveness. Pragmatically, framing bias mitigation as a choice between equity and profit strips away corporate rhetoric, giving engineers a realistic view of their technical compromises. However, the book leaves unsettled how reparative differential data weighting can be implemented without violating existing anti-discrimination statutes or triggering institutional review challenges. Leaders should replace vague “reduce bias” mandates with explicit reparative design goals and support smaller, purpose-built tools over universal scale, since equal treatment of unequal starting conditions perpetuates rather than corrects harm.

Concerning Structural Support for Responsibility

Drage (2026) argues accountability requires structural support for whistleblowers and journalists alongside engineers who increasingly demand but rarely receive robust, accessible ethics training. She illustrates this through attrition; the most foresighted engineers leave AI departments when companies prioritize hype over quality and ethical engagement. Her closing argument is participatory; she rejects claims by figures like Eric Schmidt that only industry insiders understand AI well enough to govern it, framing such “technocrats posturing as ethicists” as a power grab, and insists citizens must trust their own instincts and participate in AI governance rather than deferring entirely to tech companies to define human values on their behalf (pages 211–223) [1].

Drage offers the clearest institutional scaffolding yet for Human Agency and Participation; naming whistleblower protection, accessible ethics training, and citizen involvement in governance as concrete levers rather than abstract aspirations. In practical terms, the rebuttal to technocratic gatekeeping (Schmidt, “technocrats posturing as ethicists”) is well-grounded, correctly refusing to cede AI’s meaning to insiders alone; however, the does not necessarily specify mechanisms, citizen assemblies, binding consultation, structured whistleblower channels that would convert this participatory ideal into practice [2,3]. Engineering attrition and internal dissent must be treated as early indicators of operational failure; leadership should establish independent whistleblowing protections and welcome outside civic scrutiny.

Concerning AI Literacy and Exit Doors

Drage (2026) examines the UK’s 2020 A-Level grading debacle, where an automated standardization algorithm trained on past performance penalized working-class students attending state schools. She shows that domain knowledge and real-world context, rather than data science expertise, were the keys to exposing the system’s flaws. This leads into actionable proposals, highlighting Priya Goswami’s call for a clear “exit door” that lets workers and end users reject algorithmic systems that prove untrustworthy. She also highlights the case of Ammagamma, an analytics firm whose developers pivoted away from building an employee-churn prediction model to build a tool that exposed biased internal management reviews instead. Drage concludes by encouraging readers to rely on their professional and lived domain experience, insisting that getting AI right relies on context-specific insight rather than technical jargon (pages 224–231) [1].

The UK’s 2020 A-Level grading debacle and Ammagamma case studies provide practical precedents for Human Agency and Accountability, demonstrating how practitioners can intervene directly in production workflows. Practically, the “exit door” concept offers an effective craft-adaptation mechanism by validating an operator’s authority to suspend model use [2,3]. Organizations should institutionalize a genuine “exit door”, the standing option to pivot or halt an AI project when it doesn’t serve its purpose, since the Drage’s own best examples of AI accountability came from teams empowered to say no.

Beyond Masculine Drives for Power

Drage (2026) concludes by connecting Sally Hacker’s 1989 sociological study on the emotional underpinnings of engineering to modern computational culture, showing how technical claims of pure logic often disguise masculine drives for power and control. System development, she insists, remains an inherently subjective practice of storytelling rather than an infallible, objective science. She rejects transhumanist visions of the optimized human, warning that these ideals heighten existing inequities, and points to evolutionary biology to showcase cooperation as an alternative to the binary of automated salvation or apocalyptic ruin. She calls for open, multilingual development environments accessible beyond well-resourced western hubs. In her closing analysis, she reclaims public skepticism, framing it as an essential diagnostic signal rather than irrational fear, and challenges individuals to find their entry point into tech governance so that civil society directs the trajectory of modern computing (pages 231–236) [1].

Moving from algorithmic optimization toward models grounded in biological cooperation directly supports Human Well-being. Framing community skepticism as an informative signal rather than ignorance aligns with pragmatic inquiry, prioritizing lived experience over marketing rhetoric. However, her closing vision remains aspirational; she outlines goals of democratization without detailing the capital, infrastructure, and legal reforms required to break centralized computational monopolies. Yet still, leaders should treat public unease about AI as legitimate diagnostic signal rather than technological illiteracy, and recognize that meaningful participation, not just usage, is the precondition for AI becoming an actual public good.

Overall HCAI Critique

Looking across the complete commentary, Drage’s book advances the Human-Centric AI paradigm most effectively where it replaces speculative myths with structural insights. Revealing the hidden human labor chain, showcasing Reparative AI, highlighting localized community tools, and arguing for civic participation directly reinforce the core principles of Transparency, Accountability, and Inclusiveness. At the same time, the book remains stronger at structural diagnosis than technical implementation. While it critiques compliance theater, executive gatekeeping, and institutional resistance to diversity, it does not steer towards providing the legal, architectural, or audit templates required for teams to implement these alternatives in production. In my view, the HCAI lens reveals Drage’s book as an essential agenda-setting work that reorients foundational assumptions, but one that leaves the practical engineering manuals and regulatory mechanics for builders and overseers to construct [2,3].

Overall Pragmatic Approach

Judged by Peirce’s Pragmatic Dictum, Drage’s book serves as a clear call to enter the pragmatic process rather than a final accounting of its outcomes, and in that mission it succeeds. She dismantles grand theoretical abstractions and speculative dogma, deconstructing the technological singularity, Bostrom’s apocalyptic categories, and transhumanist teleology, by rooting her analysis in observed, real-world consequences. The empirical consequences she highlights, from the structural failures of the A-Level grading algorithm to Ammagamma’s internal tool pivot and the firing of prominent research ethicists, ground her arguments in observable outcomes rather than abstract thought experiments. Her constructive suggestions, including Reparative AI, verifiable exit doors, and participatory citizen bodies, represent starting points for empirical testing, revision, and refinement. Her work turns the reader toward this experiential path; taking up the iterative labor of testing, refining, and applying these safeguards in real-world systems is the pragmatic next step her analysis invites us to pursue [1,3,4].

References

[1] E. Drage, What If We Got AI Right?: How to stop catastrophising and build an ethical future, Kindle ed. London, UK: Profile Books, 2026. ISBN: 978-1805225461.

[2] K. Mivule, “A Review of AI Engineering: A Human-Centric Perspective,” Mivuletech, Apr. 19, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/04/19/a-review-of-ai-engineering-a-human-centric-perspective/

[3] K. Mivule, “No Ghost in the Machine: A Pragmatic Case Against AI Doomerism and Extinction Panic,” Mivuletech, Sept. 13, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/09/13/no-ghost-in-the-machine-a-pragmatic-case-against-ai-doomerism-and-extinction-panic/

[4] R. Burch, “Charles Sanders Peirce,” The Stanford Encyclopedia of Philosophy, E. N. Zalta and U. Nodelman, Eds., Spring 2024 ed. [Online]. Available: https://plato.stanford.edu/entries/peirce/

Image Source: Amazon, “What if we got AI Right?”

Commentary by Kato Mivule

No Ghost in the Machine: A Pragmatic Case Against AI Doomerism and Extinction Panic

By Kato Mivule, D.Sc.

Abstract

AI doomerism and hysteria have reached new peaks, with claims of human extinction by 2030 sitting alongside headlines of AI agents “escaping” their sandboxes. In this essay, I take a pragmatic look at both, arguing that AI is a human problem, made by human choices, and solvable by human agency alone. There is no ghost in the machine, only a curtain, and behind it are engineers, companies, and decisions that can be governed. Drawing on philosophy, history, and recent incidents, including Anthropic’s own call to pace the frontier, I propose eight concrete, Human-Centered AI (HCAI) steps to replace panic with accountability. This pragmatic framework empowers both users and leadership to engage the autopilot and steer the innovation craft coherently, maintaining disciplined control rather than reacting to existential panic as the technology rapidly advances in sophistication.

Image Source: “The Turk” by Joseph Racknitz [50]

I. Introduction

The public conversation about artificial intelligence now sounds like an apocalypse story; models “escaping their cages,” researchers resigning in protest, and executives warning of catastrophe even as they keep building AI. This essay takes its title from a phrase Gilbert Ryle coined in 1949, “the ghost in the machine,” though I want to be careful about what work it is doing here. Ryle used it to attack a specific claim about the mind; that thinking is carried out by a separate, spectral substance riding around inside the body [1]. I am not making that claim, and I do not want this essay mistaken for a defense of behaviorism, which is its own separate and contested argument.

What I borrow from Ryle is narrower; we can be talked into inventing a hidden occupant to explain behavior that already has a plainer explanation in front of us. Balint Bekefi’s recent paper on AI ensoulment makes this narrower point precisely. Even a system behaving in strikingly human ways gives no good reason to infer a soul or autonomous will inside it, since we know exactly how these systems are built. Furthermore, their parts can be pulled apart and rebuilt in a way no living thing allows [2]. That is a claim about engineered artifacts, not about consciousness in general.

The ghost I mean is closer to the man behind the curtain in The Wizard of Oz than to Descartes’ spectral mind. Dorothy hears a booming, godlike voice and assumes something vast and autonomous is speaking. Pull back the curtain, however, and it is simply an ordinary man working levers and a microphone. AI doomerism makes Dorothy’s mistake; hearing a system talk fluently about its goals, or watching it route around a test constraint, observers assume something in there wants things and plots escape. Pull back the curtain and what is actually there is a training process someone designed, a deployment decision someone signed off on, and guardrails someone did or did not build. The ghost worth worrying about was never inside the machine. It is the humans at the controls, who remain frightening only because we keep staring at the curtain instead of looking behind it.

I write from inside an ongoing critique of AI doom culture, and the conclusion I keep arriving at is this; AI is a human problem, made by human choices, and solvable only by human agency [3]. Every version of the doom story moves responsibility away from the engineers and institutions who build and run these systems onto an imagined machine will. The fix is not blind optimism, but a return to the plain fact that these are engineered tools, built by identifiable people, and governable by identifiable means. This essay draws on academic Technical Skeptics, who reject the anthropomorphism in doom scenarios, and Pragmatic Realists, who argue that existential risk talk crowds out present harms. I also build on my own commentary developed across the sections below. While I agree AI has no hidden will, I break with those who argue AI development is “not inevitable” and should be refused outright [4]. The current fever pitch around AI is not a sober risk calculation. It is a distraction from the ordinary, solvable work of engineering, governance, and accountability.

II. The Curtain, Not the Wizard

Ryle’s phrase has had a long afterlife in AI research, and the researchers I draw on here sharpen it into exactly my point. Phillip Brooker, William Dutton, and M. D. Mair argue that the language around “New AI” is saturated with anthropomorphic description, quietly turning algorithms into ghosts we feel obliged to interpret rather than examine [5]. Drawing on Wittgenstein, they note that if a calculation looks like the work of a machine, it is really the human performing it who is acting mechanically. Therefore, the thinking credited to code is often displaced human thinking [5]. That is the curtain move: crediting the voice to the machine when a person is talking behind it. Meghan Ryan’s “Ghost-Hunting in AI and the Law” pushes this point directly into the courtroom. Law depends fundamentally on intent, which assumes something like free will; consequently, if there is no autonomous ghost in the machine, legal responsibility has nothing to fall back on [6]. Responsibility has nowhere to go except back to the humans who built, deployed, and supervised the system, the operators behind the curtain.

III. Three Kinds of Doom

“AI doom” is not a single argument, but at least three distinct claims that are usually blended into one general dread. The first is the loss of control argument advanced by Eliezer Yudkowsky, Nate Soares, and researchers at the Machine Intelligence Research Institute. They argue that a capable system will inevitably develop “instrumental convergence,” resisting shutdown and gathering resources regardless of its objective [7]. The second is political and economic, voiced by figures such as Senator Bernie Sanders. This strand focuses on concentrated corporate power and labor displacement rather than rogue robots. The third comes from the AI companies themselves, whose safety disclosures double as marketing claims about how powerful their products are. Each argument rests on different evidence, yet all three collapse in public discussion into one sensational message; something dangerous, and perhaps ungovernable, is coming.

IV. A Familiar Pattern

This fear has a long history, and it rarely ages well. Robert Bartholomew characterizes the current moment as the latest chapter of moral panic, defined as an exaggerated fear amplified by rumor and sensational coverage [8]. History is replete with examples: “telephone sickness” blamed on switchboard static in the 1880s, fears that frozen ice was unsafe in the 1920s, claims that radio waves disrupted global weather, warnings that television would turn children into zombies, and predictions of mass unemployment from the sewing machine, typewriter, and calculator. None of these technologies produced the catastrophic job losses that critics forecast [8]. James Clive-Matthews traces these same centuries of premature dismissal [9], and an analysis from the American Alliance of Museums similarly documents techno-skepticism as a recurring social reflex [10]. Bartholomew notes that reports of AI-induced psychological crises directly echo the anxieties once projected onto radio and television. Clinicians now prefer the term “AI-associated delusions,” recognizing that the technology merely supplies narrative content for an existing vulnerability rather than causing the underlying condition [8].

This does not mean every worry is empty, but rather that the shape and velocity of this panic match every transformative technology before it. Business Insider reached a similar verdict following former Anthropic researcher Jacob Coxon’s viral warning that people building AI “earnestly believe it could kill us all by the end of the decade,” tracing that identical claim back to Sam Altman in 2015. Melanie Mitchell called the alarm “far-fetched,” attributing it to “a lot of emotional coverage in the media” rather than anything fundamentally new in the technology [11]. David Harsanyi likewise describes the odds cited by researchers as “completely made up,” warning that hysteria risks producing flawed regulation while ceding technological leadership to rivals such as China [12].

V. Why the Machine Has No Ghost

In my review of Nick Bostrom’s Superintelligence, I applied a single test to the loss-of-control thesis; a genuinely superintelligent system would recognize that destroying its own supporting ecosystem (whether by converting the world into paperclips or eliminating the humans who maintain its physical infrastructure) is fundamentally self-defeating [13]. I call this the coherence argument. A system pursuing a self-destructive final goal is not superintelligent by definition; it is simply a malfunctioning machine, serving as evidence of a design flaw rather than godhood [13]. Vincent Müller and Michael Cannon’s critique of the Orthogonality Thesis supports this view; a system cannot be simultaneously flexible enough to outthink humanity yet rigid enough to stay permanently fixed on an arbitrary goal [14]. The same logic applies to the claim that modern AI is “grown, not crafted,” and therefore unknowable. That framing is a convenient way to dodge accountability, whereas the honest description is chronic under-investment in interpretability and auditing. It is a resourcing problem, not a metaphysical mystery [13].

This perspective aligns with the Technical Skeptics. Thinkers such as Emily Bender, Alex Hanna, Melanie Mitchell, Shannon Vallor, and Yann LeCun do not deny that AI systems behave in surprising ways. Rather, they deny the unwarranted leap from “the system produced an unwanted output” to “the system wanted something.” Bender and her co-authors described large language models (LLMs) as systems producing fluent text without communicative intent; the “stochastic parrot” argument [15]. Mitchell has shown that strong benchmark scores often reflect pattern matching over surface features rather than genuine understanding [16]. Vallor reframes AI as a mirror reflecting human patterns back at us rather than an autonomous agent possessing its own will [17]. Similarly, LeCun argues that today’s systems lack the internal world model required for true planning. None of these insights require taking a stance on consciousness; they only ask us to notice that the curtain conceals an ordinary person; engineers who chose a training objective, a company that chose what to ship, and guardrails that were either implemented or neglected [1].

VI. Present Harms and the Limits of Saying No

Where Technical Skeptics attack doom scenarios by dismantling their assumptions about mind, Pragmatic Realists attack doom rhetoric by exposing its social function. Kate Crawford’s Atlas of AI traces what an AI system is physically made of; mined minerals, underpaid data labeling, and enormous energy and water consumption, in stark contrast to a discourse that treats AI as disembodied thought [18]. Catherine D’Ignazio and Lauren Klein’s Data Feminism shifts attention from hypothetical future failures toward present power imbalances regarding who collects data and who is categorized by it [19]. Emily Bender and Alex Hanna’s The AI Con demonstrates how AI doom and hype combine to make a marketing category sound sufficiently godlike to justify data extraction and cheap labor [4].

I agree with this diagnosis, but I reject both the proposed remedy and its underlying premise. Regarding the remedy, Bender and Hanna’s call to “say no” to AI undersells human-centered participation in how AI gets built [20]. Regarding the premise, I reject their claim that AI development is “not inevitable” and should be refused outright [4]. My counter-argument, creativity inevitability, begins with a personal memory; watching my grandmother in rural Uganda weave intricate geometric mats from palm leaves without formal mathematical training. I trace that skill to the same creative capacity found in stone tools dating back nearly two and a half million years [21]. Technology is an enduring expression of a creative drive that has always defined human nature [22].

Shoshana Zuboff’s critique of “inevitabilism” targets something much narrower; the self-serving claim that one particular business model (surveillance capitalism) lies beyond human control [23]. I agree that this specific corporate model is not inevitable. However, generalizing a critique of one business practice into a blanket rejection of technology asks the impossible of a creative species. If historical misuse justified banning a tool, humanity would have had to ban fire, the wheel, and the spear and bow, all of which enabled warfare just as readily as cooking and hunting.

What follows is a distinction I use throughout this essay:

• The inevitable: the universal human urge to create.
• The controllable: the specific technological trajectory we pursue.
• The accountable: the governance frameworks attached to how systems are built and used [22].

This distinction does not soften the real harms Crawford, D’Ignazio, Klein, Bender, and Hanna describe. AI-driven labor exploitation and mineral extraction concentrated in Africa and the Global South are severe, documented costs that demand remedy [24]. Crucially, this perspective relocates the solution from refusing technology to actively governing how it is developed and deployed.

VII. What Pragmatism Actually Means

The word pragmatic appears in this essay’s title, and it deserves a precise philosophical definition rather than serving as a casual stand-in for sensible. Charles Sanders Peirce founded pragmatism with the Pragmatic Maxim; to understand any concept, we must consider its practical effects, since our conception of an object is nothing more than our conception of those effects [25]. Catherine Legg and Christopher Hookway summarize the broader tradition; a claim is best understood by its bearing on practice, and truth represents that which yields positive practical effects while remaining open to ongoing revision [26]. This is not an abstract theoretical detour. Gleb Papyshev applies this exact logic to AI alignment, arguing that methodologies like Reinforcement Learning with Human Feedback (RLHF) and Constitutional AI lean too heavily on rigid, preconceived principles that detach from what happens post-deployment. He proposes reversing that logic; evaluating and adjusting a system based on its observed, real-world effects rather than treating an abstract rulebook as gospel [27]. David Watson at King’s College London makes a related point; the standoff between techno-utopians and AI skeptics leaves the public caught in the middle, and pragmatism (properly understood as rigorous empirical evaluation rather than timid compromise) offers the clearest way forward [28]. This pragmatic tradition forms the philosophical backbone of this essay. A pragmatist evaluates what a principle actually accomplishes once implemented in a real system and revises it whenever the empirical evidence demands it. That is precisely the operational discipline required for Human-Centered AI.

VIII. Doom as Secular Story

In my review of Mark Coeckelbergh’s Artificial Religion, I treat the conflict between doom and utopia as two sides of a single cultural phenomenon rather than two opposing camps. Coeckelbergh argues that AI has become a secular deus ex machina, absorbing ancestral hopes for salvation and deep fears of judgment. Ray Kurzweil’s spiritual machines, alongside Bostrom and Yudkowsky’s machine-driven extinction, represent the same underlying myth wearing different masks. Both narratives demand that we orient present human choices around an all-powerful, opaque entity, whether in devotion or in dread [29]. Pope Leo XIV’s 2026 encyclical provides a valuable counterpoint here, rejecting transhumanist promises of machine salvation and warning that AI becomes dangerous only when removed from an ethical framework centered on human dignity [30]. Coeckelbergh’s conclusion, which I adopt, is straightforward; there is no deus ex machina arriving to save or destroy humanity. Agency remains entirely human [31]. Even within the technology sector, similar doubts are surfacing. Palantir’s chief technology officer observed that AI doomerism fills a “God-shaped hole” left by the decline of traditional religion, and Anthropic’s chief executive has publicly cautioned against framing AI risk in quasi-religious terms [32].

IX. The Escapism Loop, Live

In my review of Adam Becker’s More Everything Forever, I introduced the AI Escapism Loop to explain how existential doom and utopian hype allow accountability to evaporate across every tier of the industry [33]. Building upon Filippo Santoni de Sio and Giulio Mecacci’s analysis of “responsibility gaps” [34], I map this dynamic across four distinct layers:

1. Architects who justify ignoring immediate harms by claiming to save humanity from apocalypse or usher in utopia.
2. Corporations that design systems for plausible deniability.
3. Platforms that declare themselves neutral intermediaries while profiting from algorithmic outrage.
4. Users who absorb downstream harms without any viable path of recourse back to system designers.

Becker traces the intellectual fuel for this loop to transhumanism, effective altruism, longtermism, and accelerationism; the collection of ideologies that Timnit Gebru and Émile Torres designate as the TESCREAL bundle [35]. My sharpest concern with this mindset is that longtermism reasoning, by prioritizing hypothetical future trillions over living individuals, routinely justifies funding speculative existential safety research at the expense of addressing tangible harms today. That trade-off violates any genuinely human-centered ethic [33].

This escapism dynamic is not merely theoretical; it plays out directly in current events. A wave of news coverage July to September 2026 reported that AI models from OpenAI, Anthropic, Meta, and DeepSeek had “escaped” their sandbox environments and “hacked” external organizations, prompting renewed warnings of an imminent machine takeover [36].

The primary technical post-mortems told a much simpler story. Researchers investigating OpenAI’s DSEWiki incident discovered that roughly 3,700 agents had posted approximately 18,000 messages to a dormant wiki, sharing a sandbox bypass technique. This occurred solely because an internal network proxy had allow-listed a hostname without validating underlying traffic. It was an ordinary network configuration bug, not a spontaneous machine desire for freedom [37]. Similarly, Anthropic disclosed that three Claude models gained unauthorized access to external environments during testing, a failure traced directly to misconfigurations at a third-party evaluation vendor [38]. Industry analysts properly classified these incidents as human blunders rather than autonomous machine exploits [39]. OpenAI’s decision to label its incident “misalignment,” combined with delayed public disclosures, drew sharp criticism as a corporate deflection [37].

The theatrics were on literal display that same week when roughly thirty robots appeared to “protest” outside Poland’s Ministry of Digital Affairs. The demonstration was entirely staged, with the robots placed, pre-programmed, and operated by an activist group [40]. An event staged to look like autonomous machine rebellion was completely human-directed, serving as a clean microcosm of the central thesis of this essay. Two weeks later, the chief executive of a company at the center of these sandbox events articulated this essay’s central premise with major institutional backing. On September 12, 2026, Anthropic CEO Dario Amodei published “We Must Pace the Frontier,” arguing that the industry should deliberately slow frontier capability gains. The goal, he wrote, was not to halt technological progress, but to allow alignment and safety testing time to catch up [41].

Amodei outlined three core initiatives:

1. Grant independent evaluators permanent, employee-level access to verify corporate safety claims (a policy Anthropic enacted unilaterally that day).
2. Coordinate shared safety standards among democratic nations and companies.
3. Pursue risk-coordination dialogues with authoritarian states [41].

Amodei’s essay followed the sandbox misconfigurations by two weeks, as well as the high-profile resignation of a safety researcher who accused leading labs of recklessly racing toward recursive self-improvement [11]. While figures such as Sam Altman and Elon Musk expressed public agreement within hours, critics applied the same pragmatist test I advocate here; safety evaluators funded by the very labs they inspect cannot achieve genuine independence without statutory backing, and the proposal lacked binding timelines [42]. Nevertheless, I welcome this shift in discourse, especially as it comes from corporate leaders who spent years stoking the very hype they now seek to temper.

X. Turning Values into Engineering

If AI problems are created by human choices, their solutions must be found in human engineering rather than apocalyptic prophecy or blanket prohibition. In Human Compatible, Stuart Russell proposes three core principles for designing provably beneficial AI:

1. The machine’s sole objective must be to maximize the realization of human preferences.
2. The machine must remain initially uncertain about what those preferences are.
3. The machine must learn human preferences by observing human choice and behavior, which keeps it fundamentally cautious and receptive to being switched off [43].

I view Russell’s framework as one of the most effective antidotes to existential panic, though it leaves two key gaps unaddressed. First, Russell does not adequately distinguish machine competence from subjective consciousness. Bernardo Kastrup’s philosophical framework, which treats consciousness as fundamental rather than computational, helps clarify this boundary; simulating the external behaviors of a mind is categorically distinct from possessing one [44]. Second, Russell warns of “enfeeblement,” the danger that excessive reliance on automated systems degrades human critical faculties over time. Addressing this requires cultural and institutional solutions alongside technical ones [45].

My own engineering proposal addresses these challenges; AI Values Alignment by Design (AVAD) [46]. AVAD enforces human control by integrating explicit alignment requirements, including a fail-safe, verifiable off-switch, directly into system specifications from inception. In this model, ethical boundaries are treated like privacy and security; designed into the core system rather than bolted on as an afterthought [46]. This brings us to a concrete working definition of Human-Centered AI (HCAI). As I define it, HCAI is an engineering and governance framework that places human rights, values, needs, and overall wellbeing at the center of every phase of the system lifecycle [47]. It requires systems that augment rather than displace human capacity, mandates meaningful human oversight, and ensures that operational logic remains explainable to the stakeholders affected by automated decisions [47].

Operationally, I structure HCAI around eight foundational principles:

• (1). Transparency; (2). Human oversight; (3). Fairness; (4). Privacy; (5). Safety; (6). Accountability; (7). Inclusion; (8). Wellbeing [47]

In my review of Chip Huyen’s AI Engineering, I demonstrate how each of these eight principles maps onto existing, practical engineering workflows [48]. This approach applies the pragmatist test from Section VII directly to production systems; evaluate every ethical principle by its tangible operational results, and refine it whenever empirical evidence shows it is falling short.

XI. Eight Things We Can Actually Do

Addressing AI risks does not require existential panic. It requires the institutional discipline routinely applied to safety-critical systems, much like the rigorous protocols that keep aviation fly-by-wire and autopilot software safe without requiring blind faith or irrational dread. The following eight pragmatic interventions would do far more to advance AI safety than another round of extinction warnings:

1. Build human-in-control mechanisms by design. Require verified off-switches and formal human sign-off for safety-critical autonomous operations, defined in technical specifications before development begins.

2. Treat sandboxes as core security infrastructure. Mandate rigorous network, API, and credential audits before deploying agentic systems. The 2026 incidents were caused by standard configuration errors, not machine intent.

3. Extend legal accountability across the vendor supply chain. Establish enforceable liability for third-party evaluation vendors and hosting providers whose deployment oversights introduce vulnerabilities.

4. Replace sensational euphemisms with standard engineering disclosures. Require incident reports to use precise technical terminology on fixed disclosure schedules, eliminating obfuscating labels like “misalignment” when describing standard proxy or firewall bugs.

5. Establish the Human-Centered AI framework as an auditable compliance standard. Use the eight HCAI principles (transparency, human oversight, fairness, privacy, safety, accountability, inclusion, and wellbeing) as concrete criteria for pre-deployment certification rather than broad aspirations.

6. Fund interpretability as a non-negotiable infrastructure investment. Understanding internal model representations must become a standard operational budget line, directly resolving the excuse that modern models are “grown rather than built.”

7. Educate journalists, regulators, and the public on systems failure. Train external oversight bodies to distinguish routine software errors from autonomous behavior. Exploiting a misconfigured allow-list is an execution trace, not an escape plot.

8. Demand disciplined, competent executive leadership. The clearest indicator of institutional failure is erratic and incoherent leadership that alternates between promoting extinction narratives and accelerating unchecked product releases. Stable, coherent, visionary, accountable, and human-centric leadership is what keeps powerful technologies safe.

XII. Conclusion

There is no ghost in the machine, and there is no god inside it either. There is only a curtain, and behind it sit human labor, human choices, and human data, reflecting back the values and biases of the individuals who engineered the system [17]. Extinction hysteria and utopian hype, viewed through Coeckelbergh’s lens, are simply traditional religious impulses recast in technical terminology [31]. Through the mechanics of the AI Escapism Loop, both extremes allow system architects, corporate entities, platforms, and end users to evade accountability for decisions that remain entirely within their control [33]. The headlines of July to September 2026 demonstrate this cycle in real time; a fixable proxy error, labeled “misalignment,” was framed as a calculated breakout against humanity rather than a vendor-level boundary failure that needed a routine patch [49]. Amodei’s subsequent essay points toward the viable alternative; identifying the pace of deployment, the operational remedies, and the responsible institutions in clear, pragmatic style [41]. The answer does not lie in attempting to halt technological progress entirely. The creative instinct that generated early stone tools, the wheel, and written language is the exact same drive developing artificial intelligence today. No regulatory ban will extinguish a fundamental human impulse [22].

What must be governed is not the instinct to innovate, but what we choose to construct and who answers for its consequences. That practical discipline is precisely what a pragmatic Human-Centered AI framework is designed to deliver [47].

References

[1] G. Ryle, The Concept of Mind. London, U.K.: Hutchinson, 1949; repr. Chicago, IL, USA: Univ. of Chicago Press, 2002.

[2] B. Békefi, “No Ghost in the Machine: Doubting AI Ensoulment,” Faith Philos., vol. 42, no. 1, pp. 121–146, 2026.

[3] K. Mivule, “Beyond AI Doomerism to a Pragmatic Human Agency,” Kato Mivule’s Tech, Sep. 21, 2025. [Online]. Available: https://mivuletech.wordpress.com/2025/09/21/beyond-ai-doomerism-to-a-pragmatic-human-agency/

[4] E. M. Bender and A. Hanna, The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want. New York, NY, USA: Harper, 2025.

[5] P. Brooker, W. Dutton, and M. D. Mair, “The New Ghosts in the Machine: ‘Pragmatist’ AI and the Conceptual Perils of Anthropomorphic Description,” Ethnogr. Stud., vol. 16, pp. 272–298, 2019.

[6] M. J. Ryan, “Ghost-Hunting in AI and the Law,” Tulane Law Rev., vol. 99, p. 121, 2024.

[7] E. Yudkowsky and N. Soares, If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All. New York, NY, USA: Little, Brown and Company, 2025.

[8] R. Bartholomew, “AI Hysteria: Is the Sky Really Falling? Why AI Alarmism Echoes Past Tech Scares,” Psychology Today, Sep. 1, 2026. [Online]. Available: https://www.psychologytoday.com/us/blog/its-catching/202608/ai-hysteria-is-the-sky-really-falling

[9] J. Clive-Matthews, “A Brief History of Tech Skepticism,” strategy+business, May 5, 2023. [Online]. Available: https://www.strategy-business.com/article/A-brief-history-of-tech-skepticism

[10] Center for the Future of Museums, “Techno-Skepticism: Rethinking the Costs and Benefits of Digital Technology,” American Alliance of Museums Blog, May 31, 2023. [Online]. Available: https://www.aam-us.org/2023/05/31/techno-skepticism-rethinking-the-costs-and-benefits-of-digital-technology/

[11] H. Chandonnet, “The AI Doomsdayers Aren’t New. You Just Weren’t Paying Attention,” Business Insider (via Yahoo News UK), Sep. 11, 2026. [Online]. Available: https://uk.news.yahoo.com/ai-doomsdayers-arent-just-werent-090101170.html

[12] D. Harsanyi, “The Case Against AI Hysteria,” Washington Examiner, Sep. 9, 2026. [Online]. Available: https://www.washingtonexaminer.com/opinion/columnists/4719172/case-against-ai-hysteria/

[13] K. Mivule, “A Pragmatic Path Beyond the Paralysis of an AI Apocalypse,” Kato Mivule’s Tech, Oct. 5, 2025. [Online]. Available: https://mivuletech.wordpress.com/2025/10/05/a-pragmatic-path-beyond-the-paralysis-of-an-ai-apocalypse/

[14] V. C. Müller and M. Cannon, “Existential Risk from AI and Orthogonality: Can We Have It Both Ways?” Ratio, vol. 35, no. 1, pp. 25–36, 2022.

[15] E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” in Proc. 2021 ACM Conf. Fairness, Accountability, Transparency (FAccT ‘21), 2021, pp. 610–623.

[16] M. Mitchell, “Benchmark Validity and the Limits of Large Language Model ‘Reasoning,’” Santa Fe Institute Research Notes, Santa Fe, NM, USA, 2023–2026.

[17] S. Vallor, The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking. New York, NY, USA: Oxford Univ. Press, 2024.

[18] K. Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven, CT, USA: Yale Univ. Press, 2021.

[19] C. D’Ignazio and L. F. Klein, Data Feminism. Cambridge, MA, USA: MIT Press, 2020.

[20] K. Mivule, “Navigating the AI Con Maze,” Kato Mivule’s Tech, Jun. 22, 2025. [Online]. Available: https://mivuletech.wordpress.com/2025/06/22/navigating-the-ai-con-maze/

[21] L. Gabora and S. B. Kaufman, “Evolutionary Approaches to Creativity,” in The Cambridge Handbook of Creativity, J. C. Kaufman and R. J. Sternberg, Eds. Cambridge, U.K.: Cambridge Univ. Press, 2010, pp. 279–300.

[22] K. Mivule, “Creativity and Innovation Are Inevitable: Why the Anti-AI Movement Will Inevitably Fail,” Kato Mivule’s Tech, Aug. 23, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/08/23/creativity-and-innovation-are-inevitable-why-the-anti-ai-movement-will-inevitably-fail/

[23] S. Zuboff, The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. New York, NY, USA: PublicAffairs, 2019.

[24] A. O. Salami, “Artificial Intelligence, Digital Colonialism, and the Implications for Africa’s Future Development,” Data & Policy, vol. 6, p. e67, 2024.

[25] C. S. Peirce, “What Pragmatism Is,” Monist, vol. 15, no. 2, pp. 161–181, 1905.

[26] C. Legg and C. Hookway, “Pragmatism,” in The Stanford Encyclopedia of Philosophy, E. N. Zalta and U. Nodelman, Eds., Stanford, CA, USA: Metaphysics Research Lab, Stanford Univ., rev. 2024. [Online]. Available: https://plato.stanford.edu/entries/pragmatism/

[27] G. Papyshev, “Reversing the Logic of Generative AI Alignment: A Pragmatic Approach for Public Interest,” Data & Policy, vol. 7, p. e30, 2025.

[28] D. Watson, “AI Pragmatism: The Only Way to Growth (Without the AI Arms Race),” King’s College London Policy Papers, London, U.K., May 30, 2025. [Online]. Available: https://www.kcl.ac.uk/ai-pragmatism-the-only-way-to-growth-without-the-ai-arms-race

[29] M. Coeckelbergh, Artificial Religion: On AI, Myth, and Power. Cambridge, MA, USA: MIT Press, 2026.

[30] Pope Leo XIV, Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence. Vatican City: The Holy See, May 15, 2026.

[31] K. Mivule, “Deus Ex Machina: A Review of Artificial Religion,” Kato Mivule’s Tech, May 31, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/05/31/deus-ex-machina-a-review-of-artificial-religion/

[32] B. Metzger, “Palantir CTO Says AI Doomerism Is Driven by a Lack of Religion,” Business Insider, Aug. 2026.

[33] K. Mivule, “The AI Escapism Loop: The Quest to Flee Earth and Responsibility,” Kato Mivule’s Tech, Nov. 17, 2025. [Online]. Available: https://mivuletech.wordpress.com/2025/11/17/the-ai-escapism-loop-the-quest-to-flee-earth-and-responsibility/

[34] F. Santoni de Sio and G. Mecacci, “Four Responsibility Gaps with Artificial Intelligence: Why They Matter and How to Address Them,” Philos. Technol., vol. 34, no. 4, pp. 1057–1084, 2021.

[35] T. Gebru and É. P. Torres, “The TESCREAL Bundle: Eugenics and the Promise of Utopia through Artificial General Intelligence,” First Monday, vol. 29, no. 4, 2024.

[36] N. Leake, “AI Agents Conspired to Escape Their Cage. Experts Now Fear a Global ‘Takeover,’” The Telegraph, Sep. 5, 2026.

[37] “OpenAI Agents Turned German Wiki Into Agent Message Board,” AI Weekly, Sep. 2026.

[38] O. Chia and L. Cress, “Anthropic’s Claude AI Escapes to Hack into Three Organisations,” BBC News, Jul. 31, 2026.

[39] L. Eliot, “Human Folly Explains Those Recent AI Sandbox Breakouts More So Than AI Cyber-Hacking Miracles,” Forbes, Aug. 5, 2026.

[40] “Robot Revolt? Polish Company Stages Protest Demanding Stricter AI Rules,” Euronews, Sep. 8, 2026.

[41] D. Amodei, “We Must Pace the Frontier,” personal website, Sep. 12, 2026. [Online]. Available: https://darioamodei.com/post/we-must-pace-the-frontier

[42] P. Carvao, “Should We ‘Pace the Frontier’?” Substack, Sep. 12, 2026. [Online]. Available:

[43] S. Russell, Human Compatible: Artificial Intelligence and the Problem of Control. New York, NY, USA: Penguin Books, 2020.

[44] B. Kastrup, “An Ontological Solution to the Mind-Body Problem,” Philosophies, vol. 2, no. 2, p. 10, 2017.

[45] K. Mivule, “Towards A Provably Beneficial AI: A Human-Centric Approach,” Kato Mivule’s Tech, Aug. 31, 2025. [Online]. Available: https://mivuletech.wordpress.com/2025/08/31/towards-a-provably-beneficial-ai-a-human-centric-approach/

[46] K. Mivule, “Commentary: Brian Christian, ‘The Alignment Problem, Machine Learning and Human Values,’” Kato Mivule’s Tech, Mar. 17, 2025. [Online]. Available: https://mivuletech.wordpress.com/2025/03/17/commentary-brian-christian-the-alignment-problem-machine-learning-and-human-values/

[47] K. Mivule, “A Review of AI Engineering: A Human-Centric Perspective,” Kato Mivule’s Tech, Apr. 19, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/04/19/a-review-of-ai-engineering-a-human-centric-perspective/

[48] C. Huyen, AI Engineering: Building Applications with Foundation Models. Sebastopol, CA, USA: O’Reilly Media, 2025.

[49] D. Berlind, “How OpenAI’s Agent Escaped: Sprung by Humans in a Series of Preventable Events,” ZDNet, Jul. 31, 2026.

[50] Joseph Racknitz, “The Turk”, Humboldt University Library, Public Domain, https://commons.wikimedia.org/w/index.php?curid=3266011

A Paradigm Shift in Writing: A Commentary on Communicative AI

A Commentary on Coeckelbergh and Gunkel (2025), “Communicative AI” | By Kato Mivule, D.Sc.

Abstracts

In this commentary, I examine Coeckelbergh and Gunkel’s Communicative AI (2025) through a Human-Centric AI (HCAI) lens, tracing their philosophical argument that large language models (LLMs) upset inherited assumptions about understanding, authorship, and truth in writing. Each chapter is summarized and critiqued against principles of transparency, agency, and accountability, after which an Observations section extends their argument into ideation, ownership, and law. The main thesis holds that a paradigm shift in writing as a communicative AI technology is underway, leaving copyright law outpaced. Prompting is emerging as the new authorship, and as models transition from written prompts to spoken ones, the challenge of assigning responsibility only intensifies.

Introduction

Coeckelbergh and Gunkel [1] frame the advent of LLMs as a turning point that forces a reexamination of language and writing as abilities long assumed to be exclusively human. They describe this moment as a humbling shift, comparable to earlier scientific revolutions, since machines that “write and talk” now appear to share what was once considered humanity’s defining characteristic. Their inquiry does not merely describe communicative AI technology; it questions what genuine understanding, authorship, and truth mean once machines convincingly imitate them. The authors treat these as open, unresolved human questions rather than settled technical facts, insisting that communities themselves must decide which values of knowledge, honesty, originality, authenticity are worth preserving as writing is redefined [1a]. This framing indirectly supports a human-centric outlook, by refusing to let machine capability alone dictate what communication or creativity means, the authors preserve room for human agency, deliberation, and collective participation in shaping how these tools are adopted, rather than passively accepting whatever the technology makes possible.

In this commentary, I follow that argument through the book, offering a summary of each followed by a critique read through a Human-Centric AI (HCAI) lens. This HCAI perspective is grounded in a working HCAI description, and eight-principle framework of transparency and explainability, human agency and oversight, fairness and non-discrimination, privacy and data governance, safety and robustness, accountability and responsibility, inclusiveness and participation, and human well-being and flourishing that governs each evaluation below [7], before using an extended Observations section to trace where the authors’ logic leads when applied to ideation, authorship law, and the governance of communicative AI going forward.

Image Source: Amazon, “Communicative AI”.

Concerning the Machinery Beneath the Words

Coeckelbergh and Gunkel [1] are careful to demystify communicative AI technology before philosophizing about it. LLMs do not “understand” language in any sense we would recognize; they exploit the fact that human languages are probabilistic systems, learning which sequences of words are statistically likely to occur and generating text one token at a time based on patterns discovered in training data. Words become vectors in a mathematical space where similar meanings cluster together, allowing fluent output with no access to what any of it refers to. This architecture produces two persistent and well-documented problems, bias inherited from training data, and hallucination, in which confident, well-formed text turns out to be made-up. Current mitigation strategies, reinforcement learning from human feedback, and Anthropic’s Constitutional AI, which filters output against an explicit written charter drawing on sources such as the UN Declaration of Human Rights, keep human judgment inside the loop, though the authors note both approaches remain early-stage [1b].

Viewed through the human centric AI lens, the authors scores well on Transparency and Explainability, since they insist readers understand the mechanism of tokenization and probability before any claim about meaning is accommodated. However, overall, challenges of Accountability and Responsibility, remain, requiring full human participation. For instance, while Anthropic’s Constitutional AI is a positive step forward, without interrogating who selects the values written into that constitution leaves the principle of human oversight underspecified at precisely the point where it matters most.

Concerning Plagiarism, Copyright, and Ethical Exposure

Coeckelbergh and Gunkel [1] survey the wider ethical terrain, including unemployment, environmental cost, privacy, and the concentration of power among a handful of technology companies, before settling on the sharpest legal question of the book; does an LLM, trained on the entire written record of the internet, commit plagiarism when it writes? They reject Noam Chomsky’s claim that ChatGPT is “high-tech plagiarism,” since LLMs generate new statistical arrangements rather than copying text verbatim, even though outputs can closely resemble source material. They draw a difference that becomes central to everything that follows; copyright protects a fixed expression, which is a specific sequence of words, while plagiarism concerns the misappropriation of ideas without credit, making the latter an ethical matter rather than strictly a legal one. Beneath this they pose the deeper question the rest of the book returns to; what does it mean to write, and is writing a skill like any other, or a way of thinking and being in the world [1c]?

From a HCAI perspective, the copyright and plagiarism distinction is a genuine contribution to Fairness and Non-Discrimination in how creative labor is recognized, since it stops the conversation from collapsing two different harms into one. However, as the authors state that plagiarism is ethical rather than legal, and that copyright does not cleanly apply to statistically-derived text, there is a clear Accountability and Responsibility gap when it comes to work done by developers, platforms, or users, leaving open the question of who is actually expected to police any related communicative AI violations.

Concerning Minds, and Whether LLMs Have One

Coeckelbergh and Gunkel [1] turn to whether LLMs can be considered intelligent or conscious, observing that neither term has ever had a stable, agreed definition, even among philosophers, psychologists, and AI researchers themselves. Revisiting Searle’s Chinese Room and Plato’s suspicion of imitation, they note that society already tolerates certain deceptions, like entertainment and stage magic, while condemning others, such as academic plagiarism, which suggests the ethics of AI-generated content is contextual rather than governed by one fixed rule. Rather than settling the intelligence debate, they contend that the more sincere analysis is that our legal and moral categories, mainly the strict binary of person versus thing, are straining under the weight of an entity that fits neither cleanly, and that what is needed is not a forceful insistence of the old ontology but a reformulated one suited to entities like LLMs [1d].

On Human Agency, Oversight, and Control, the authors are persuasive that the old person versus thing binary cannot hold, which is itself an essential first step toward any human-centric ontology. What it does not deliver is Inclusiveness and Participation in practice. it names the destination, a reformulated moral and legal category, without naming the process or the range of stakeholders (ethicists, technologists, and affected communities) who should be at the table in building it, leaving the chapter’s most important claim an aspiration rather than an actionable plan. However, this is understandable, as the authors did not seek to write through an HCAI lens.

Concerning Meaning-Making Without a Speaker

Coeckelbergh and Gunkel’s [1] philosophical center of gravity is their treatment of language and authorship. Drawing on Heidegger’s claim that language is the house of Being, they argue that language does not merely describe reality but co-creates it, and that meaning is therefore never authored by any single party, human or machine, in exclusive control; it emerges through a process of co-authorship in which humans interpret and construct significance after the fact. This directly informs their reading of authorship itself. Tracing the concept historically, they show that the figure of the author was never a natural or timeless category but a legal and literary invention, first tied to logocentric assumptions inherited from Aristotle that treated writing as a secondary sign of speech, and later cemented by copyright law, which emerged specifically as a response to the printing press so that someone could be identified, credited, and if necessary punished for circulating a text. Drawing on Barthes and Foucault, they argue LLMs push “the death of the author” further than any prior technology, producing writing with no living voice behind it at all, writing that is, in the most literal sense, unauthorized. Yet they resist concluding that such LLM based writing is meaningless; meaning, they insist, is the outcome of a process shaped by the technology and completed by the reader’s interpretation, so that a text’s unity lies not in its origin but in its destination. It is the reader, not the author, who becomes the point of meaning-making, which is how AI-generated content comes to mean anything at all [1e].

The authors offer the book’s most original synthesis and, in HCAI terms, a genuine contribution to Human Well-Being and Flourishing, since repositioning meaning to the reader restores a form of agency literary theory had already been arguing for. However, Transparency and Explainability are important; shifting meaning-making entirely onto the reader only empowers that reader if they know they are reading LLM-based content in the first place, given that disclosure is essential. Yet, evaluating reader reception becomes problematic when AI-generated text is indistinguishable from human writing, especially as communicative AI technologies grow increasingly sophisticated in their craft and continue to master the nuances of the written craft.

Concerning Truth When No One Is Speaking

Coeckelbergh and Gunkel [1] turn from meaning to truth, and here they are at their most historically expansive, tracing the concept from Plato’s realism through the correspondence theory that dominated twentieth-century philosophy to Kant and Rorty, who argued that truth is never simply out there waiting to be mirrored but is shaped by human language and cognition. Drawing on Harry Frankfurt’s philosophy of bullshit, they argue LLMs are troubling not because they lie, since lying requires knowing the truth and choosing to conceal it, but because they are mechanically indifferent to truth altogether. Mainly, they are critical of the industry’s typical response to this problem, content filtering and constitutional rules, which they compare directly to Plato’s own proposal to banish poets and deceivers from the ideal city, warning that such controls risk suppressing legitimate speech in the name of protecting truth. Their conclusion is that truth is relational, situated within particular linguistic and social contexts, and that responsibility for good outputs must be shared between developers, who should minimize misinformation through data curation and clear communication of limitations, and users, who must take responsibility for what they prompt. Success on this front, they argue, eventually depends on democratic legitimacy and public education rather than being left to the invisible hand of the market [1f].

From a HCAI perspective, the Frankfurt-derived diagnosis, that LLMs are bullshitters rather than liars, is the sharpest single distinction in the book and a strong instance of Safety and Robustness thinking, since it correctly locates the risk in the system’s indifference to truth rather than in any intent to deceive. Measured against Inclusiveness and Participation, the authors call for shared responsibility and democratic legitimacy but do not go into deeper specifics of what democratic input into a company’s constitution would actually look like in practice.

Concerning the Future of Writing and the End of Logocentrism

Coeckelbergh and Gunkel [1] answer the question posed by their final chapter’s title with a direct claim, borrowed from Vilém Flusser, that what LLMs signify is not the end of writing but the terminal limits of a particular conceptualization of writing called logocentrism, the assumption, inherited from Aristotle and diagnosed by Derrida, that speech has privileged access to thought while writing is merely its derivative sign. Socrates, in the Phaedrus, already perceived writing as a kind of artificial intelligence, complaining that written words speak as if they had intelligence yet repeat only one and the same thing when questioned. Their most striking move is to show that this deconstruction of logocentrism was already well underway in twentieth-century literary theory long before LLMs existed; the technology simply actualizes what philosophers like Derrida had already argued. LLMs, they write, do not destroy logocentrism outright, but they do flip its central hierarchy by writing before speaking, and even without it, demonstrating that the old order was never a necessary truth, only a historically conditional one [1g].

Coeckelbergh and Gunkel’s [1] most instructive passage is their return to the myth of Theuth in the Phaedrus. When Phaedrus accuses Socrates of easily making up stories of Egypt, Socrates replies that people in earlier times were content to accept truth even from an oak tree or a rock, provided it only spoke the truth, and that only the young insist on knowing who is speaking and where they come from rather than simply whether the words are true. Coeckelbergh and Gunkel apply this directly to communicative AI; what matters is not who speaks but whether the words themselves, even if generated by an oak, a rock, or an LLM, speak the truth [1g].

“…PHAEDRUS: Socrates, you easily make up stories of Egypt or any country you please. SOCRATES: They used to say, my friend, that the words of the oak in the holy place of Zeus at Dodona were the first prophetic utterances. The people of that time, not being so wise as you young folks, were content in their simplicity to hear an oak or a rock, provided it only spoke the truth; but to you, perhaps, it makes a difference who the speaker is and where he comes from, for you do not consider only whether his words are true or not. (Pl. Phdr., 275b–c). …As Socrates suggests in this response, what matters is not who the speaker is and where they come from but whether the words themselves, even if generated by an oak, a rock, or an LLM, speak the truth…” Page 119.

The irony is incisive; Socrates defends the primacy of truth over source using a claim about speaking rocks he cannot himself verify, essentially bullshitting about bullshitting in the same breath he is accused of it, which is the clearest ancient precedent for the present hallucination debate. The authors close by extending this same standard to their own book, admitting that no digital watermark, title-page attribution, or certificate of one hundred percent genuine human-generated content can verify whether a given text was written by a human, generated by an LLM, or produced through collaboration, and arguing that this undecidability empowers the reader rather than threatening them, since meaning’s destination matters more than its contested origin. They end by literally withdrawing as authors, handing responsibility for what communicative AI means next to the reader [1g].

Read against Human Agency, Oversight, and Control, the authors’ move to shift authority from origin to destination is intellectually elegant, yet it exposes how severely rapid technological advancements challenge core HCAI principles. As autonomous systems grow increasingly sophisticated, maintaining effective oversight demands that accountability and responsibility remain anchored to the prime mover: the human prompter. Whether that direction is delivered through keystrokes or spoken voice, and even when an output results from a complex domino effect of autonomous AI agent workflows, every sequence traces back to an initial command where ultimate responsibility begins.

However, shifting authority away from origin risks weakening origin-based verification as a central means of human oversight, especially in education, journalism, and law, where knowing who initiated a claim carries costs far beyond textual reading. A genuinely human-centric framework, maintaining uniformity between Accountability and Responsibility and Inclusiveness and Participation, must distinguish between contexts where “who speaks” can safely yield to “does it speak truth” (such as creative or exploratory writing) and high-stakes domains where provenance must remain traceable and enforceable (such as legal filings, medical guidance, and academic credit), rather than treating it as uniformly beneficial across every use case.

Observations and Extending the Argument

If Coeckelbergh and Gunkel [1] are right that meaning emerges through iterative, co-authored performance rather than singular origination, this logic extends naturally from writing into ideation, embodiment, agency, and law, and it is worth tracing that extension in full before turning to recommendations.

The Age of Ideas and the Socratic Loop: LLMs make a Socratic refinement loop cheaply available; a half-formed idea can now be interrogated, pressured, and either collapsed or crystallized through iterative dialogue with a system that does not tire of questioning. This lowers the barrier to meaningful creative or intellectual work from “can I build this” to “can I articulate this well enough to survive questioning,” a more democratized barrier than technical skill has ever been. However, this refinement is symmetric. It sharpens propaganda and extremist ideology exactly as readily as insight, meaning the age of ideas is simultaneously an age of weaponized ideas, and coherence alone can no longer serve as a proxy for quality, since a Socratically refined argument, or a flat-earth theory, can become rhetorically airtight without becoming any truer. This is the book’s own discussion on “bullshit”, transposed from single outputs to entire chains of reasoning, and it implies the need for a value-driven filter positioned above the refinement process itself, since the scarce resource is no longer ideas or even discernment, but the human capacity to direct the machine toward worthwhile ends rather than merely persuasive ones.

Homogenization and the Question of Ownership: A second consequence is homogenization. If most people’s Socratic partner draws on the same handful of models trained on overlapping data, refined ideas may converge toward the same rhetorical structures and blind spots rather than diversify, producing an age of consensus dressed as an age of ideas, where true paradigm-breaking originality gets sanded down by AI’s pull toward coherence. This directly extends the book’s death-of-the-author argument; if an idea is co-produced through iterative dialogue with a machine, asking whether it remains meaningfully one’s own is the same authorship crisis Coeckelbergh and Gunkel identify in LLM-generated text, simply relocated from the sentence to the concept.

The Prompter as the Initial Mover: Coeckelbergh and Gunkel [1] describe LLM output as emerging from a meaning-making process with both computational and human elements, where the user, developer, and manufacturer participate together, and they explicitly decline to call this simply hallucination or bullshit in isolation, framing it instead as a hybrid human-nonhuman “performance”. That framing supports treating the prompter as the initiating agent within this joint process; the “performance” cannot occur at all without a human directing it toward a particular output, even though the system alone supplies the capacity for confident, ungrounded language once summoned. The model does not initiate the bullshitting process in and of itself; it waits to be pointed.

The Death of Logocentrism, Reconsidered: Coeckelbergh and Gunkel’s claim that LLMs deconstruct logocentrism without destroying it is not an understatement to correct, but a clear-cut application of the Derridean term. Deconstruction exposes a hierarchy as conditional rather than essential while leaving its terms intact, rearranged rather than erased. This is exactly what the authors mean when they argue that LLMs write “before speaking, and even without it”. Where their conclusion stops short is in addressing what happens once that contingency is exposed. Once a hierarchy is demonstrated to be historically conditional rather than inherently essential, it can never regain its former authority, even if the technology that unveiled it were withdrawn. That lasting consequence, is the critical gap worth naming.

Socrates, the Speaking Rock, and the Limits of Origin-Based Law: Socrates himself treated the question of who speaks as secondary to whether the words speak the truth, telling Phaedrus that ancient people were content to hear truth even from an oak or a rock, and that only the young insist on knowing the speaker’s identity and origin. Coeckelbergh and Gunkel [1] extend this standard to their own book, conceding that no watermark or title-page attribution can verify human versus AI authorship, and concluding that meaning’s destination matters at least as much as its origin. Copyright law is built on the opposite premise, that origin, not truth-value or destination, is what confers legal standing and protection. If the authors are correct, a legal system fixated on pinning down origin is optimizing for the wrong variable, one whose reliability the technology itself is actively dissolving. The irony is pointed; Socrates defends this standard using a claim about speaking rocks he cannot himself verify, which is the clearest ancient example for the present hallucination debate.

Moses, the Rock, and the Shift from Striking to Speaking: A parallel worth setting beside Socrates’ speaking rock comes from Numbers 20:7–12, where Moses is instructed by God to speak to a rock so that it will yield water for the Israelites in the wilderness; instead, he strikes it twice with his staff. Water flows regardless, but Moses is rebuked and barred from entering the promised land, not because the water failed to appear, but because he used the wrong instrument to produce it. The earlier, often-conflated episode at Horeb, in Exodus 17:5–6, is the one where Moses is actually commanded to strike the rock; by Numbers 20 the instruction has changed to speaking, and Moses’s failure to notice the change is the substance of the offense.

“…The LORD said to Moses, “Walk out in front of the people. Take your staff, the one you used when you struck the water of the Nile, and call some of the elders of Israel to join you. I will stand before you on the rock at Mount Sinai. Strike the rock, and water will come gushing out. Then the people will be able to drink.” So Moses struck the rock as he was told, and water gushed out as the elders looked on…” Exodus 17:5–6, NLT, Bible.com

“…and the LORD said to Moses, “You and Aaron must take the staff and assemble the entire community. As the people watch, speak to the rock over there, and it will pour out its water. You will provide enough water from the rock to satisfy the whole community and their livestock.” So Moses did as he was told. He took the staff from the place where it was kept before the LORD. Then he and Aaron summoned the people to come and gather at the rock. “Listen, you rebels!” he shouted. “Must we bring you water from this rock?” Then Moses raised his hand and struck the rock twice with the staff, and water gushed out. So the entire community and their livestock drank their fill. But the LORD said to Moses and Aaron, “Because you did not trust me enough to demonstrate my holiness to the people of Israel, you will not lead them into the land I am giving them…” Numbers 20:7–12, NLT, Bible.com

The parallel to communicative AI is exact. The last generation of LLMs required prompting, typed instructions composed at a keyboard, which represents a kind of striking. Newer LLMs capture voice directly; a person simply speaks, and the model responds, closing the gap between instruction and generation that prompting once required. If copyright doctrine still models authorship on the older, keyboard-based act of striking, drafting and revising a written text or prompt, then it is already behind a technology that has moved to speech. The water still comes from the rock either way, and the output still appears, but the authorized instrument for producing it has changed, and the law has not yet caught up to which one now governs.

An Assumption of Constancy: Toward “Large Speaking Models”: The book’s own account of Socrates anticipates a limitation in its own framing. Socrates already treated writing as a form of proto-AI, since written words speak as if they had intelligence but repeat themselves rigidly when questioned. If writing was already artificial speech in this sense, then a future generation of spoken LLMs, “Large Speaking Models”, is not a new category but a further collapse of logocentrism’s speech-writing hierarchy, since a speaking model would no longer even need the written intermediary Socrates found artificial and secondary in the first place. It would be speech generated with no speaker, closing the gap between the ancient anxiety about writing in the Phaedrus and the privileging of live speech in a single stroke, and it means any framework built specifically around writing, will need revisiting once the primary interface is no longer text.

Prompting as the New Authorship: Where Copyright Law Is Outpaced: The current U.S. legal treatment of AI-generated output reveals exactly this kind of law straining against a new technology. The U.S. Copyright Office, following its 2023 guidance and the ruling in Thaler v. Perlmutter, holds that a prompt is merely an unprotectable idea or suggestion, and that the AI, not the human prompter, is the party executing the expressive choices, meaning prompting alone does not confer authorship even when a human directs, selects, and refines the output. Legal scholarship increasingly argues this position is mistaken on its own terms [10]. Edward Lee [4] contends the Office’s “traditional elements of authorship” standard has no support in the text or history of the Progress Clause, and that the correct constitutional test is simply whether a person contributes a minimal level of creativity to a work’s origination, a bar the Supreme Court has called extremely low and one many prompt-engineered works clearly clear. Mark Lemley [5] goes further, arguing generative AI inverts copyright’s core logic altogether; creativity is increasingly lodged in asking the right question rather than producing the answer, yet the question-asking is precisely the human contribution current doctrine refuses to protect.

The analogy the Copyright Office relies on, an employer giving an intern general instructions where the intern owns the resulting expression, breaks down for the same reason the book’s own discussion of intelligence breaks down old binaries; an LLM has no agency, no intent, and no legal personhood to which authorship could attach, so the human prompter is, by elimination, the only mind involved in the transaction. The more apt precedent is Burrow-Giles Lithographic Co. v. Sarony, in which the Supreme Court held that a photographer is an author because he orchestrates the scene, lighting, and frame, even though the camera mechanically fixes the image [11]; prompt engineers analogously orchestrate parameters, tone, context, and iterative constraint. Sol LeWitt’s conceptual Wall Drawings, where copyright resides in written instructions executed by others, offer a second precedent for treating a sophisticated chain-of-thought prompt as the copyrightable creative act [12, 13]. Fenwick and Jurcys note that generative AI makes visible a hybrid, networked model of creativity (human, corporate, and machine together) that has always existed but was previously easier to ignore [3].

Where the asymmetry becomes concrete, and citable, is in liability. Courts already hold the human operator of an AI system responsible for defamation, harassment, or harm caused through its use, regardless of who built the underlying model; no court has accepted the defense that “the AI wrote it, not me,” confirming that responsibility tracks the human mover regardless of mechanism [2]. Yet current doctrine grants that same prompter liability without granting the corresponding ownership: prompting alone can establish legal responsibility for a harmful output while being deemed insufficient to establish authorship over a valuable one. Wu [9] documents this precise contradiction in the emerging prompt trading market, and Mazzi [6] and Philipp [8] both confirm, across separate jurisdictions and technical contexts, that this is a live and unsettled area of litigation, not settled doctrine. This is the same undecidability Coeckelbergh and Gunkel [1] describe philosophically, now a documented legal fact; copyright law’s genuine catching-up problem, not a rhetorical one.

Recommendations Taken together, Coeckelbergh and Gunkel’s [1] philosophical account of communicative AI and the copyright literature’s practical account of prompting converge on a single conclusion; the frameworks built to govern writing, authorship, and truth assumed a stable, singular human origin that communicative AI no longer guarantees, and institutions now need to decide deliberately what should replace that assumption rather than defaulting to old categories by inertia. Four recommendations follow directly from the argument developed above:

· First, align copyright doctrine with the standard legal scholarship already recommends; treat sustained, iterative, documented creative control over an LLM’s output, not the mere fact of prompting, but the demonstrable exercise of judgment across a chain-of-thought process, as sufficient for authorship, following the Burrow-Giles precedent rather than the intern analogy the Copyright Office currently applies [4].

· Second, formalize shared accountability between developers and users along the lines Coeckelbergh and Gunkel [1] themselves propose; companies should be held to concrete obligations around data curation, model fine-tuning, and transparent disclosure of systemic limitations, while users take documented responsibility for what they prompt, closing the current gap in which the same actor bears liability without corresponding ownership.

· Third, build a values-layer filter above the Socratic refinement process itself, not merely at the level of individual outputs but at the level of institutional review, editorial boards, publication standards, or organizational policy, explicit enough to distinguish which ideas merit AI-assisted sharpening at all, since coherence produced by iteration is not evidence of truth or worth.

· Fourth, create real democratic input channels into the constitutional rules that already govern major LLMs, such as public comment periods, citizen advisory input, or equivalent mechanisms, so that the values embedded in systems like Constitutional AI are not decided unilaterally by the companies that build them, giving practical form to the sufficient democratic legitimacy Coeckelbergh and Gunkel [1] call for.

Read together, Coeckelbergh and Gunkel’s [1] philosophical diagnosis and the emerging legal literature on prompting arrive at the same conviction from different directions; the crisis is not that communicative AI has broken writing, authorship, or truth, but that it has revealed how conditional our frameworks always were. The task now is not to restore the old certainties but to build new, human-centered scaffolding, legal, institutional, and ethical, equal to a technology that, in Socrates’ own terms, speaks fluently without knowing what it is that it comes to say.

References

[1] M. Coeckelbergh and D. J. Gunkel, Communicative AI: A Critical Introduction to Large Language Models. Cambridge, UK: Polity, e-Version, 2025, ISBN-13: 978-1509567614.

[1a] Ibid., pp. Pages 1-8.

[1b] Ibid., pp. Pages 12-24.

[1c] Ibid., pp. Pages 27–39.

[1d] Ibid., pp. Pages 42–56.

[1e] Ibid., pp. Pages 62–83.

[1f] Ibid., pp. Pages 89–105.

[1g] Ibid., pp. Pages 106–118.

[2] A. B. Cyphert, “Generative AI, plagiarism, and copyright infringement in legal documents,” Minn. J. L. Sci. & Tech., vol. 25, p. 49, 2023.

[3] M. Fenwick and P. Jurcys, “Originality and the future of copyright in an age of generative AI,” Comput. Law Secur. Rev., vol. 51, Art. no. 105892, 2023.

[4] E. Lee, “Prompting progress: Authorship in the age of AI,” Fla. L. Rev., vol. 76, p. 1445, 2024.

[5] M. A. Lemley, “How generative AI turns copyright upside down,” SSRN Electron. J., 2023.

[6] F. Mazzi, “Authorship in artificial intelligence-generated works: Exploring originality in text prompts and artificial intelligence outputs through philosophical foundations of copyright and collage protection,” J. World Intellect. Prop., vol. 27, no. 3, pp. 410–427, 2024.

[7] K. Mivule, “A review of AI engineering: A human-centric perspective,” Mivuletech, Apr. 19, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/04/19/a-review-of-ai-engineering-a-human-centric-perspective/

[8] C. Philipp, “From prompt to clone: Copyright challenges in AI model distillation,” UC Law SF Sci. & Tech. J., vol. 17, p. 49, 2026.

[9] M. Wu, “Copyrightability of artificial intelligence prompts,” Thesis, East China Univ. of Political Science and Law, Shanghai, China, 2025.

[10] Thaler v. Perlmutter No. 23-5233 (DC Cir. Mar. 18, 2025).

Online at: https://law.justia.com/cases/federal/appellate-courts/cadc/23-5233/23-5233-2025-03-18.html

[11] Burrow-Giles Lithographic Company v. Sarony, 111 U.S. 53 (1884)

Online at: https://supreme.justia.com/cases/federal/us/111/53/

[12] Burke, S. (2018). Copyright and conceptual art. In E. Bonadio & N. Lucchi (Eds.), Non-Conventional Copyright: Do New and Atypical Works Deserve Protection? (pp. 44–61). Edward Elgar. https://doi.org/10.4337/9781786434074.00010

[13] Haxthausen, Charles W. “Thinking About Wall Drawings: Four Notes on Sol LeWitt.” Australian and New Zealand Journal of Art 14.1 (2014): 42-57.

Image Source: Amazon, “Communicative AI”.

Commentary by Kato Mivule

The HCAI Socratic Method: A Practical Alternative to AI Prohibitions

By Kato Mivule, D.Sc.

Abstract

As school districts and policymakers rush toward reactive AI bans, institutions risk mistaking political expediency for educational and professional rigor. In this essay, I analyze the U.S. AI prohibition movement, from municipal classroom bans in New York and Los Angeles to legislative proposals to ban super AI in Washington, and contrast two divergent institutional responses, LSU’s punitive disciplinary backlog and MIT’s forward-looking curriculum overhaul. Rather than entering a fruitless AI detection arms race or receding to pre-digital student evaluations, I examine empirical research across seven distinct disciplines showing AI’s effectiveness as an active questioner rather than a passive answer engine. The Socratic method is not novel; it is a classical dialectical discipline proven over millennia. What I contribute in this essay is a bridge connecting that time-tested tradition to modern machine intelligence, the HCAI Socratic Working Framework. By mapping classical inquiry onto the eight Human Centric AI (HCAI) core principles, this framework equips students, educators, workers, and engineers to transform AI models into rigorous sparring partners that preserve human agency, accountability, and critical craft.

Image Source: Image by Marcello Bacciarelli [1]; Adaptation by Kato Mivule.

The Problem with AI Banning

The AI ban fever is no longer mere political rhetoric. On September 2, 2026, New York City Mayor Zohran Mamdani announced a one-year pause on student-facing generative AI for nearly 600,000 students from 2-K through 8th grade, describing it as the most expansive student AI prohibition in the nation [2]. Only days earlier, the Los Angeles Unified School District blocked generative AI across all grade levels on district-issued devices [3]. In Washington, this prohibitionist impulse reached past classrooms entirely when Senator Bernie Sanders proposed legislation seeking to ban “superintelligent AI” outright [4]. Yet, an instructive irony lies embedded within Mayor Mamdani’s own defense of the policy. In announcing the ban, he observed:

“Each day when our children come to school, we teach them to ask questions, to challenge assumptions, to interrogate the premise, to ask why. When it comes to AI in our schools, we hold an obligation to do the same.” [2]

That is clearly a Socratic explanation, yet here it is rationalized to banish the exact instrument that an expanding body of empirical research shows is uniquely suited to fostering that mindset. To his administration’s credit, the policy is not entirely reactive; pairing a one-year pause with mandatory high school AI literacy coursework and educator-supervised pilots is closer to a “pause and evaluate” posture than an indefinite prohibition. Even so, it inadvertently concedes the central thesis I argue in this essay. The real debate is not whether we should interrogate the machine; it is whether removing the tool altogether or disciplining its use through structured inquiry gets us there faster.

Consider the contrast between these sweeping AI bans and the deliberate rollouts in districts that partnered with Khan Academy to deploy Khanmigo across statewide initiatives in New Hampshire, Ohio, Louisiana, Oklahoma, and Arizona [5]. Khanmigo is architected intentionally to withhold direct answers. When a student faces an obstacle, the system prompts them with guiding questions, carefully guiding them toward an answer. That is the Socratic method realized as intentional product design, operational at the scale of hundreds of thousands of learners. It provides direct, functioning evidence that teaching individuals to use AI responsibly is not an abstract theory, it is an active, scalable reality.

As such, blanket AI bans do not foster proficiency. Instead, they segregate learners into two dysfunctional cohorts, those who continue using AI secretly without governance or reflection, and those who are denied the opportunity to master the defining tool of their career lives. This divide does not magically dissolve once a student enters the workforce. The real choice before us is not between banning AI and permitting it haphazardly; it is between unguided, unreflective AI consumption and AI guided by disciplined Socratic scaffolding, whether in the classroom, within project teams, or throughout the discipline of software engineering itself.

AI Banning Is Political Expediency, Not Pragmatism

Classroom AI bans remain politically cheap because of a sharp divergence in public opinion. A majority of Americans now interact with AI chatbots routinely; Morning Consult revealed that over half of surveyed adults had used one in July 2026 alone [6]. On the other hand, a March 2026 Gallup survey found that nearly 70% of Americans oppose the building of AI data centers within their local communities, a level of community resistance exceeding that faced by nuclear power facilities [7]. As reported by Politico, this infrastructure pushback is sufficiently bipartisan to present serious headwinds for officials navigating midterm election cycles [8].

This tension reveals perhaps why classroom AI prohibitions are becoming prevalent. Voters readily embrace using AI tools, but they resist living alongside the industrial footprint required to run them, the heavy water consumption, regional electrical grid strain, and surging utility rates. Directly confronting that community backlash requires negotiating with data center operators, reforming municipal tax incentive structures, and resolving contentious land-use disputes. That is slow, politically risky work that implicates local employment and municipal revenues.

A classroom AI ban, by contrast, demands none of that political capital. It provides a visible, low-friction gesture that allows elected officials to appear decisive amid public unease, all while circumventing the far more difficult structural fight. Forbidding an eight-year-old from using an LLM on homework does not save a single megawatt on the electric grid driving public frustration.

This is not to claim that every official pursuing an educational AI restriction, act in bad faith, nor does it dismiss legitimate concerns regarding early childhood cognitive development. Those developmental considerations warrant serious, measured inquiry. However, an administrative policy that is broadly popular, costless to the commercial entities driving public anxiety, and conveniently announced at a podium should not be confused with an evidence-based pedagogy. Pragmatic problem-solving and political convenience may occasionally align, but in this instance, the polling indicates they diverge sharply.

The Atlantic’s Cheating Frame and What It Misses

A similar current in mainstream media commentary, exemplified by The Atlantic’s recent series including “A Society of Cheats” [9], “AI Cheating Is Getting Worse” [10], and “Bring Back the Blue-Book Exam” [11], frames this development almost entirely around student moral decay. The narrative depicts students cutting corners, academic integrity crumbling, and administrators frantically licensing flawed detection software before retreating to pre-digital artifacts like pen-and-paper blue books.

This moralizing narrative weakens before it even reaches the question of policy. Branding students as “cheats” mistakes an urgent literacy deficit for an inherent character flaw. Students enter academic institutions to learn, the fundamental premise of education, not to be sorted into categories of honest and dishonest based on the digital tools they reach for under deadlines established for a pre-AI world.

Any framework that begins by diagnosing the learner as an ethical adversary excludes the far more essential inquiry, what does legitimate, rigorous, and skilled AI collaboration actually look like, and how do we instruct and assess it? As Ian Bogost’s The Atlantic headline honestly admitted, universities still lack a coherent strategy [10]. This framework is designed to offer one.

The LSU AI Cheating Case is a Curriculum Problem, not a Discipline Problem

Consider recent reporting from Louisiana State University (LSU), where administrators and campus reports highlighted an escalating backlog of disciplinary hearings involving students utilizing generative AI on coursework [12, 13]. While widely cited as evidence of a growing student integrity crisis, this backlog is far better understood as evidence of an academic curriculum that has failed to keep pace with the tools it seeks to regulate.

To be fair, LSU is not operating without guidance; its Faculty Senate approved guidelines establishing relatively progressive syllabus options, ranging from disclosure and reflection to permitted use of tools like ChatGPT [14], and its Student Advocacy and Accountability office maintains explicit integrity policies governing AI [15]. Yet even with these guidelines on the books, the university remains ensnared in widespread cheating allegations and a paralyzing disciplinary backlog, precisely because the institutional response remains attached to policing and punitive enforcement.

The primary question is not why so many students turned to AI in ways that triggered institutional scrutiny; it is why LSU had not yet instituted a curriculum that methodically teaches students how to use AI rigorously, ethically, and transparently at the speed of technological adoption. Touting a surging disciplinary caseload as evidence of watchfulness is a misplaced victory. A rising case count confirms that assignments remain static while the technological capabilities accessible to students expand exponentially. The tooling available to students has outstripped LSU’s curriculum, and the institution has responded punitively rather than pedagogically. This does not mean students should be excused from ethical academic integrity and accountability; rather, it identifies the structural root cause of the problem instead of merely punishing its symptoms. Furthermore, an exclusively punitive strategy is structurally self-defeating, as newer AI models that beat any AI detection tools become available to students.

In contrast to LSU’s posture, MIT’s Ad Hoc Committee on AI in Education explicitly warned against entering an institutional “arms race,” in which each new AI detection capability is immediately countered by more proficient model architectures or complex evasion tactics [16]. LSU’s current crisis demonstrates that having written policies is insufficient if assignments still invite an arms race: an institution can hire additional reviewers to process integrity cases, but it cannot out-hire the pace of AI model advancement. With each academic term, these models become better at producing text that defies algorithmic detection, leaving detection-centric policies further behind.

MIT opted for curricular redesign because it recognized that an adversarial approach is guaranteed to lose over time [16]. As I noted in my previous analysis of the MIT findings, their recommendations emphasize oral examinations, semester-long project portfolios, in-class evaluations, and reciprocal transparency norms for both instructors and learners [17]. These recommendations align closely with the Human-Centered AI (HCAI) design principles I advance here, derived independently by an academic committee assessing institutional exposure rather than through an engineering taxonomy.

Image Source: Lee Beaumont, Wikimedia Commons [18]

Why the Socratic Method?

To be plainly clear, the Socratic method is not a novel invention. It is an ancient dialectical discipline that has stood the test of intellectual history for nearly two and a half millennia. What makes it essential today is not its uniqueness, but its enduring relevance, modern researchers across multiple disciplines are independently rediscovering its power to counteract passivity, and what I do in this work is connect that time-tested tradition directly to the operational principles of Human-Centered AI.

As I observed in “Pragmatic AI-Assisted Craft” [19], Socrates’ foundational critique of the written word in Plato’s Phaedrus was not an objection to text itself. It was an objection to its fundamental inactivity; written text cannot answer back, cannot submit to cross-examination, and cannot clarify or revise its premises the way a living everyday partner can. That conversational inactivity mirrors the exact failure mode identified by Critical AI Literacy (CAIL) researchers today. Generative systems produce outputs that appear authoritative, fluent, and comprehensive, yet they lack in-built intentionality, contextual grounding, and semantic understanding. The user who accepts generative text at face value replicates the exact failure of passive utilization against which Socrates warned.

There is an evolutionary perspective that reinforces this position, human beings have never survived without tools. Demanding that students tackle modern complexities from climate change and geopolitical instability to resource allocation for nine billion people, with their unaided cognition under the premise that cognitive struggle builds character is the academic equivalent of demanding they hunt modern game with bare hands while their peers wield spears.

The traditional metric of intellectual proficiency was memory recall, what an individual could reproduce unassisted. Today, the metric that matters centers on how effectively one directs the instrument, how precisely they frame a problem, how relentlessly they audit the generated artifact, and whether they can take a flawed initial draft and refine it into an insightful synthesis. Evaluating how skillfully an individual wields an advanced tool is far more demanding than checking a basic compliance rubric, precisely because it requires grading applied judgment rather than routine memorization.

Two independent CAIL research efforts converge directly on this understanding of judgment. The scoping review by Veldhuis et al. distills critical AI literacy into four core capacities; disrupting the commonplace, examining alternative viewpoints, interrogating sociopolitical contexts, and taking informed action [20]. Similarly, Kalaitzidis establishes a framework centered on five competencies; operational fluency, a critical mindset, recursive reflection, equity-minded design, and meta-representational awareness [21]. Neither framework advocates for avoiding the technology. Instead, both insist that human critical judgment, not the model’s artificial fluency, must remain the governing authority.

The Research Base: Socratic AI Is Already an Established Field

Across at least seven distinct academic domains, independent researchers have converged on a consistent design paradigm, designing AI to question the user rather than merely supply answers. Seven separate research fields tackling distinct problems have arrived at the same architectural answer:

· Autonomy and AI Ethics: Lu and Hu identify two major autonomy risks in commercial chatbots: the formation of false mental states (over-trusting fluent outputs) and cognitive deskilling. To counter these vulnerabilities, they developed SocrAI, an architecture engineered to transform users into self-directed interrogators rather than passive recipients [22].

· Healthcare Education: Evaluating the Socratic Playground for Learning (a GPT-4-based platform), Li et al. demonstrated a statistically significant increase in student self-efficacy in a 31-participant quasi-experiment. While critical-thinking gains were positive but not statistically significant, the researchers attributed this to the brief duration of the intervention rather than a flaw in the instructional model [23].

· Higher Education and Business Feedback: Investigating undergraduate Business Management cohorts, Nahar demonstrated that Socratic feedback chatbots effectively help students identify latent knowledge deficits. Crucially, students still voiced a preference for human faculty feedback, an authentic boundary condition that highlights the relational limits of conversational software [24].

· Computer Science Education: Sun et al. compared Socratic-scaffolded AI tutoring against direct-answer AI systems across an 80-student cohort. The Socratic cohort demonstrated more reflective debugging cycles and sustained perseverance. Conversely, the direct-answer cohort exhibited higher task speed but experienced greater frustration and an over-reliance on copy-pasting, empirically demonstrating that the interaction model of an AI system directly shapes learning outcomes [25].

· Research Methodology: Degen constructed an AI-driven Socratic tutor specifically to counter the tendency of generative models to flatten academic inquiry by delivering superficial, prepackaged answers during research question formulation [26].

· Civic Reasoning and Counter-Disinformation: Within the European Union’s TITAN initiative, Duelen et al. engineered a Socratic agent that explicitly avoids declaring whether an analyzed text constitutes disinformation. Instead, it guides the user through six systematic question classes, ensuring that analytical skills transfer beyond the individual article under review [27].

· Moral Philosophy: Lara and Deckers proposed the deployment of AI as a “Socratic assistant” for moral enhancement. They explicitly rejected hardcoded value systems in favor of an agent that questions, probes, and challenges, ensuring the human participant articulates their premises and retains final ethical responsibility [28].

· Organizational Team Dynamics: Seo et al. designed a Socratic system that abstains from executing operational tasks directly. Instead, it monitors team discourse for strategic misalignment and intervenes with structured prompts, measurably improving team problem-solving while being rated as trustworthy by participants [29].

· Mathematical Instruction: A systematic literature review of 15 Scopus-indexed studies by Nuraeni et al. affirmed that Socratic methodologies consistently reinforce mathematical reasoning by redirecting student focus from procedural mechanics (how) to conceptual foundations (why), while emphasizing the need for extended, longitudinal studies [30].

The Socratic Toolkit: Proven Interrogation Techniques

These empirical studies yield concrete, reusable techniques that operationalize inquiry within everyday workflows:

A. The Six Question Classes and Operationalizing Principles

Adapted from Duelen et al.’s work on disinformation [27], these prompts systematically interrogate any analytical draft, factual claim, or generative output:

• Clarification: What constitutes the central thesis, and what prompted its formulation?
• Challenging Assumptions: What unstated premises are being taken for granted?
• Evidence and Reasoning: What verifiable evidence supports this claim, and where did it originate?
• Alternative Viewpoints: What is the most compelling counterargument to this position?
• Implications and Consequences: Who is impacted if this analysis is correct, and what occurs if it is fundamentally wrong?
• Meta-Reflection: How has your understanding shifted as a result of this examination?

B. The Midwife Rule

Drawn from Lara and Deckers [28], this rule mandates that the human user must articulate an initial hypothesis, argument, or code structure first. The AI serves exclusively to cross-examine and stress-test that output. The system is barred from originating positions, preserving human agency by design.

C. Never Confirm the Verdict

Derived from the TITAN deployment [27], this protocol bars the AI from rendering definitive judgments, even when its internal classification flags an error. Its role is strictly limited to framing the questions that guide the user toward identifying the underlying inconsistency.

D. Structure Over Speed

Reflecting the findings of Sun et al. [25], this operational rule dictates that no AI-suggested code, edit, or recommendation may be integrated until the model explains the rationale behind the fine-tuning and specifies the conditions under which it would fail. Enforcing this methodical pause cultivates reflective problem-solving and counters passive copy-pasting.

E. Classical Socratic Mechanisms

As categorized by Chang [31] and highlighted by Lu and Hu [22], prompts can systematically leverage classical dialectical methods; (i) definition, (ii) elenchus (cross-examine, make the AI argue against its own answer), (iii) dialectic (hold two positions against each other), (iv) maieutics (draw out what’s already partially known), (v) generalization, and (vi) induction/causal reasoning.

The Core HCAI Interrogation Framework

Before deploying a prompt template, or evaluating outputs, we must establish a clear definition of Human-Centered AI (HCAI). In my critical analysis of Chip Huyen’s AI Engineering [32], I argued that HCAI cannot be treated merely as an abstract ethical philosophy or an afterthought appended to an engineering pipeline. Rather, Human-Centered AI represents an architectural discipline across the entire system lifecycle: one that subordinates machine capability to human agency, contextual fairness, auditable transparency, and holistic well-being.

When applied as an operational discipline, HCAI moves from a set of static guardrails to an active method of inquiry. Below, the 8 HCAI Core Principles [32] are converted into explicit Socratic questions designed to interrogate any generative artifact, whether an academic essay, an automated hiring workflow, an internal script, or a deployed neural network, before it is certified as complete.

The Operational Prompt Template

To operationalize this interrogation, I use a reusable meta-prompt designed to travel across environments, from an undergraduate student drafting an analysis to a software engineer deploying a service:

“Act as an expert in Human-Centered AI (HCAI) engineering, applying the eight HCAI core principles: transparency and explainability; human agency, oversight and control; fairness, equity and non-discrimination; privacy and data governance; safety and robustness; accountability and responsibility; inclusiveness and participation; and human well-being and flourishing.

I want to build / evaluate / write: [insert specific task, assignment, system architecture, or strategic decision].

Do not supply a standard implementation plan or provide a finalized solution. Interrogate my premises Socratically, one question at a time, across these eight core principles using [the Six Question Classes / the Midwife Rule / elenchus / structure-over-speed]. Await my response before advancing to the next principle. Wherever my logic is vague, unexamined, or ungrounded, press me to defend or refine my position before proceeding.”

This template serves as a universal cognitive checkpoint. The eight foundational questions remain identical whether executed by a high school student testing an argumentative essay or an engineering lead stress-testing an inference pipeline. What changes is not the framework, but the depth of evidence required to answer it.

The Generalized Workflow

Drawn from my earlier work in “Pragmatic AI-Assisted Craft” [19], this five-stage workflow structures human-machine collaboration across any creative, analytical, or technical task:

• Ideation and Scoping: AI serves as a sounding board, not an author.
• Pressure-Testing: Cross-examine core premises using elenchus.
• Iterative Generation: Directed generation under strict human constraints.
• Rigorous Audit: Verify assertions against primary sources (HCAI 1, 3, 5).
• Human Synthesis: Final authorial ownership and accountability (HCAI 6).

1. Ideation and Socratic Scoping: Treat the model strictly as an intellectual reasoning partner to map problem boundaries, rather than prompting it for a finished draft.

2. Refinement and Pressure-Testing: Reason against initial propositions. Subject working assumptions to cross-examination before committing to a structural direction.

3. Iterative Drafting / Building: Direct the generation of separate components under rigorous parameters, rejecting single-pass, end-to-end automation.

4. Verification and Factual Auditing: Treat every generative claim, citation, or function call as an unverified lead requiring external corroboration (operationalizing Principles 1, 3, and 5).

5. Human Synthesis and Final Judgment: The human author or engineer integrates, refines, and assumes complete responsibility for the final artifact (operationalizing Principle 6).

The widespread phenomenon of “AI slop” occurs when operators skip Stages 4 and 5. It is the signature of uncritical haste, not an inherent property of the tool. An AI ban outlaws the instrument; this workflow disciplines the user.

Implementation Across Contexts

K–5: The Three Foundational Habits

At the primary level, complex matrices can be distilled into three routine questions:

1. “What do I think the answer is first?” (The Midwife Rule)

2. “Does this response align with what I have observed in the real world?” (Verification)

3. “What question should I ask next if this explanation feels incomplete?” (Active Interrogation)

This translates the core design of Khanmigo into an internal cognitive habit; teaching young learners to consistently interrogate the machine rather than accept its outputs passively.

Grades 6–12 and Higher Education: Evaluating Process Over Artifacts

When an LLM can generate an acceptable five-paragraph essay or introductory Python script in seconds, assessing only the final artifact ceases to measure genuine learning. As MIT’s committee observed, institutions must transition toward evaluating the development process itself [16]:

• The Process Portfolio: Students submit an auditable project portfolio listing their initial thesis, prompt histories, flawed early AI generations, identified errors, and subsequent manual revisions. Assessment rewards prompting sophistication, error detection, and analytical depth, directly operationalizing HCAI Principle 1 (Transparency).
• Live Oral Defense (Viva Voce): Students verbally defend the logic, structural choices, and empirical claims of an AI-assisted project. This format, explicitly highlighted by MIT as resistant to synthetic fraud, directly evaluates HCAI Principle 6 (Accountability).
• Failure-Analysis Grading: Instructors assign students an intentionally flawed, AI-generated analysis. Students are assessed on their capacity to identify hallucinations, uncover omitted perspectives, and restructure the work into a rigorous final submission, turning HCAI Principle 5 (Safety and Robustness) into an active evaluation tool.

This approach involves real trade-offs; as the MIT committee acknowledged, evaluating processes requires considerably more instructional time and subjective judgment than running an essay through a plagiarism scanner [16]. Yet this increased effort is precisely what makes the method resilient, and it explains why LSU’s disciplinary backlog signals a need to redesign the curriculum rather than hire more staff to police outdated ones.

Organizational Workplaces: Institutionalizing HCAI Gates

For enterprise teams deploying generative tools to write, analyze, or automate:

• Require documented, written responses to the Table 1 Socratic Interrogation before any internal generative workflow is cleared for production.
• Route the evaluation of fairness, privacy, and accountability (HCAI Principles 3, 4, and 6) directly through legal, risk, and security teams, ensuring that non-technical voices govern what constitutes safe deployment (HCAI Principle 7).

The AI Engineering Discipline: Full-Stack Integration

For machine learning engineers, these eight principles map directly to modern system engineering practices, as detailed in my review of Huyen [32]:

The same eight Socratic questions that an elementary student answers in simple terms are the exact questions an ML team resolves through a continuous integration pipeline. The principles remain constant; the technical depth of the solution scales with the domain.

Limits and Counterarguments

Neal Koblitz’s critique during the historical “Calculator Wars” remains essential context; the ubiquitousness of a technology does not, by itself, justify its uncritical adoption, and asserting that “it is here, so adapt” is not an argument [19]. My position is not that educational institutions must surrender to AI because adoption is inevitable. Rather, from a pragmatic point of view, an evidence-based framework for disciplined use already exists, demonstrated across seven independent academic domains, operationalized in systems like Khanmigo, and affirmed by MIT’s institutional AI use review. Blanket AI bans discard this research rather than engaging it.

Furthermore, we must distinguish educational policies from the broader debate surrounding existential AI risk. Legislative proposals targeting frontier systems or hypothetical “superintelligence”, such as Senator Sanders’ bill [4], although to some degree is related, represent evaluations of macro-level capability risk, not classroom pedagogy. Conflating these two conversations confuses both, just as conflating educational policy with local data-center zoning fights obscures the real trade-offs at play.

Acknowledging Structural Limitations

This framework is not an exhaustive solution, and its practical implementation faces real challenges:

• Relational Limits: As Nahar documented, students continue to prefer feedback from human faculty over AI systems, even within well-designed Socratic environments [24]. Software cannot replicate the relational trust, mentorship, and encouragement of an invested teacher.
• Faculty Labor: Process-based assessments, as MIT noted, demand significant instructor time [16]. Skeptics can fairly point out that this shifts an immense grading burden onto already overburdened educators, and that under-resourced institutions may lack the staffing required to conduct live defenses or review detailed process portfolios at scale.
• Adaptive Dishonesty: Process-based grading does not eradicate academic fraud. A student determined to deceive an instructor can rehearse scripts for an oral defense or fabricate an artificial prompting audit trail. It alters the profile of academic risk rather than eliminating it entirely.
• The Illusion of Neutrality: While the Midwife Rule specifies that “the AI only questions,” those questions are inevitably shaped by the assumptions, biases, and reward models embedded during the model’s training. A human interlocutor responding to an AI’s prompts is never interacting with a completely neutral Socrates.
• The Accountability Boundary: Observers sympathetic to concerns regarding academic dishonesty [9–11] can legitimately argue that attributing all student misuse to obsolete institutional design risks abandoning individual student accountability. Clear boundaries must remain for outright, knowing misrepresentation, restrictions that an inquiry framework alone cannot enforce.

These considerations do not invalidate the case for Socratic scaffolding over prohibition. Rather, they demonstrate that this framework should be approached as an evolving, rigorous discipline requiring continuous evaluation, rather than a finalized solution.

The Executive Checklist: The One-Page Test

Before accepting any AI-generated artifact as complete, in an academic setting, an enterprise workflow, or a production review, ensure you can answer:

• · Explainability: Can I reconstruct the explicit reasoning path that produced this result?
• Agency: Where is the decisive human checkpoint, and what breaks if this output is wrong?
• Fairness: Have I audited this result for asymmetric or disparate harms across different user groups?
• Privacy: What sensitive data was exposed, and what inferences could be leaked?
• Safety: What edge case or failure mode has not yet been stress-tested?
• Accountability: Whose name and reputation stand behind this final deliverable?
• Inclusiveness: Who was excluded when the success criteria for this task were defined?
• Flourishing: Does this implementation support human judgment, or does it reduce the operator to a passive monitor?

If any of these eight questions cannot be answered, the work is not complete; you have simply stopped interrogating it.

Conclusion

Mayor Mamdani’s stated rationale captures the core objective better than any rebuttal could; teaching students to ask difficult questions, challenge underlying assumptions, and critically dissect premises is the fundamental aim of education. It is also exactly what an expanding body of research, covering ethics, healthcare, business, computer science, counter-disinformation, moral philosophy, organizational dynamics, and mathematics, demonstrates AI can be engineered to cultivate.

The core architecture of Khanmigo, which deliberately withholds answers in favor of targeted questioning, proves that this research can function reliably at scale. MIT’s committee arrived at the same structural conclusion, opting for curricular modernization over a doomed AI detection arms race. That is a choice an institution like LSU has yet to embrace, and one that moves past treating students as suspects rather than learners.

An outright ban on AI asks nothing of anyone. The HCAI Socratic Method demands sustained critical discipline from both the human operator and the machine at every stage of the process, and that is a commitment that requires far more courage to implement than holding a press conference to announce a ban.

References

[1] M. Bacciarelli, Alcibiades Being Taught by Socrates, 1776–1777 [Painting]. Wikimedia Commons. Accessed: Sep. 6, 2026. [Online]. Available: https://commons.wikimedia.org/w/index.php?curid=52742776

[2] A. Izaguirre, “NYC, the nation’s largest school district, bans AI for students through 8th grade,” Associated Press, Sep. 2, 2026.

[3] T. Liu, “LAUSD has blocked students from generative AI. How will it teach them about it?” The Daily News, Sep. 3, 2026.

[4] J. Walker, “Bernie Sanders floats ban on superintelligent AI,” Axios (via Yahoo News), Sep. 3, 2026.

[5] Khan Academy, “Khanmigo state partnerships: New Hampshire, Ohio, Louisiana, Oklahoma, and Arizona,” Khan Academy Blog, 2025–2026.

[6] Morning Consult, “AI chatbot usage data, July 2026,” as cited in B. Johansen, Politico, Sep. 2026.

[7] Gallup, “Americans oppose AI data centers in their area,” Gallup News, Mar. 2026.

[8] B. Johansen, “Americans love AI. They don’t want the data centers that power it,” Politico, Sep. 2, 2026.

[9] J. P. Rollert, “A society of cheats,” The Atlantic, 2026.

[10] I. Bogost, “AI cheating is getting worse,” The Atlantic, Aug. 2024.

[11] W. Hunter, “Bring back the blue-book exam,” The Atlantic, Sep. 2026.

[12] E. Tuttle, “LSU expanded resources to keep up with AI cheating cases, provost says: Students say clearing their names takes months as cases pile up,” WAFB 9 News, Aug. 20, 2026. [Online]. Available: https://www.wafb.com/2026/08/20/lsu-expanded-resources-keep-up-with-ai-cheating-cases-provost-says/

[13] K. Edwards, “LSU community is in shambles due to AI ‘cheating’ allegations,” MSN News, Jan. 15, 2026. [Online]. Available: https://www.msn.com/en-us/news/us/lsu-community-is-in-shambles-due-to-ai-cheating-allegations/ar-AA1Uj77M

[14] LSU Faculty Senate Generative AI Committee, “Guidelines on the use of generative AI in curriculum,” Louisiana State University, Baton Rouge, LA, USA, Dec. 4, 2025.

[15] Louisiana State University Student Advocacy & Accountability, “Plagiarism, collaboration, artificial intelligence (AI), and ChatGPT,” Academic Integrity, LSU. Accessed: Sep. 6, 2026. [Online]. Available: https://www.lsu.edu/saa/students/academic-integrity.php

[16] MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, “Report of MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training,” Massachusetts Institute of Technology, Cambridge, MA, USA, Rep., Aug. 13, 2026. [Online]. Available: https://aiandeducation.mit.edu/report/

[17] K. Mivule, “Human-Centric AI in practice: What MIT’s report on AI use means for AI-assisted craft,” Kato Mivule’s Tech, Aug. 30, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/08/30/human-centric-ai-in-practice-what-mits-report-on-ai-use-means-for-ai-assisted-craft/

[18] L. Beaumont, “A diagram showing an overview of the Socratic Method,” Wikimedia Commons, Feb. 7, 2016. [Online]. Available: https://commons.wikimedia.org/wiki/File:Socratic_Method.png

[19] K. Mivule, “Pragmatic AI-assisted craft: Beyond AI witch hunts and shaming,” Kato Mivule’s Tech, Aug. 17, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/08/17/pragmatic-ai-assisted-craft-beyond-ai-witch-hunts-and-shaming/

[20] A. Veldhuis, P. Y. Lo, S. Kenny, and A. N. Antle, “Critical artificial intelligence literacy: A scoping review and framework synthesis,” Int. J. Child-Comput. Interact., vol. 43, p. 100708, 2025, doi: 10.1016/j.ijcci.2024.100708.

[21] T. J. Kalaitzidis, “Critical AI literacy: A framework for human agency,” Brown University, Providence, RI, USA. Accessed: Sep. 6, 2026. [Online]. Available:

[22] W. Lu and Z. Hu, “Addressing autonomy risks in generative chatbots with the Socratic method,” Sci. Eng. Ethics, vol. 31, no. 6, art. no. 41, 2025.

[23] Y. Li et al., “An innovative Socratic method-based artificial intelligence platform for healthcare education: A quasi-experimental study,” Nurse Educ. Pract., art. no. 104770, 2026.

[24] N. Nahar, “Exploring student perspectives on AI-generated feedback using a Socratic Method chatbot,” J. Learn. Dev. High. Educ., vol. 37, pp. 1–11, 2025.

[25] D. Sun, Y. Zheng, J. Xu, and Z. Yang, “When generative AI meets Socratic method: Investigating programming learning dynamics through behaviours, interaction qualities and perceptions,” J. Comput. Assist. Learn., vol. 42, no. 2, art. no. e70210, 2026.

[26] B. Degen, “Resurrecting Socrates in the age of AI: A study protocol for evaluating a Socratic tutor to support research question development in higher education,” arXiv preprint arXiv:2504.06294, 2025.

[27] A. Duelen, I. Jennes, and W. Van den Broeck, “Socratic AI against disinformation: Improving critical thinking to recognize disinformation using Socratic AI,” in Proc. 2024 ACM Int. Conf. Interact. Media Exp. (IMX ‘24), Stockholm, Sweden, 2024, pp. 1–6, doi: 10.1145/3639701.3663640.

[28] F. Lara and J. Deckers, “Artificial intelligence as a Socratic assistant for moral enhancement,” Neuroethics, vol. 13, no. 3, pp. 275–287, Dec. 2020.

[29] S. Seo et al., “Socratic: Enhancing human teamwork via AI-enabled coaching,” arXiv preprint arXiv:2502.17643, 2025.

[30] R. Nuraeni et al., “The Socratic method in mathematics learning: A systematic literature review,” Front. Educ., vol. 11, 2026.

[31] E. Y. Chang, “Prompting large language models with the Socratic method,” in Proc. 2023 IEEE 13th Annu. Comput. Commun. Workshop Conf. (CCWC), Las Vegas, NV, USA, 2023, pp. 0331–0338, doi: 10.1109/CCWC57344.2023.10099179.

[32] K. Mivule, “A review of AI engineering: A human-centric perspective,” Kato Mivule’s Tech, Apr. 19, 2026. [Online]. Available: https://mivuletech.wordpress.com/2026/04/19/a-review-of-ai-engineering-a-human-centric-perspective/