Late Lessons, Jensen Huang and AI
Markdown version

M4. Cognition, formation and being human: Jensen Huang, the AI industry and Late Lessons, read through Andrew Maynard’s work#

One of a set of analyses that read current AI developments, Jensen Huang’s September 2026 conversation with Ezra Klein, and the European Environment Agency’s Late lessons from early warnings reports through Andrew Maynard’s own work. This one covers a single dimension: what AI does to how people think, trust, learn and form themselves, and what that means for work, education, young people and being human. It is analysis, not advocacy, and it is not written in Maynard’s voice. Prepared on 26 September 2026 with extensive AI assistance, at Maynard’s request, and reviewed by him.

Conventions. - Every claim about Maynard’s position is labelled. [Stated]: he has said this, and the source is cited. [Implied]: it follows directly from positions he has stated, which are cited. [Inferred]: this analysis’s reading, plausible but not stated by him, given with its reasoning and a confidence level. - Maynard’s posts are cited by date and slug. His papers, essays and books use the keys of the companion map of his thinking (05, Appendix C): FFTF (Films from the Future), FR (Future Rising), Trojan 2026, CR 2026 (“Constitutive Resonance”), Harness 2026, and his April 2026 essays (NANO, HNS, FWB, S3 2026). [mixed] marks texts whose wording is his but some of whose concepts or prose may have originated with an AI model: 2026-09-24 being-an-academic-in-an-age-of-ai (a lecture drafted into prose by Claude and line-edited by him), 2026-07-16 orphan-risks-frontier-ai-maynard (a Fable-drafted paper he rewrote), and, in Trojan 2026, the term “honest non-signals” and the paper’s four bypass mechanisms, which he credits in part to Claude (2026-01-17). His January 2026 essay (2026-01-10 is-ai-a-cognitive-trojan-horse) is the secure own-prose source for the cognitive-Trojan-horse thesis and is cited first wherever it makes the same point. [AI-origin; endorsed] marks a point taken from an AI-written paper he endorsed. Mixed texts are used as corroboration, never as the sole basis for a position; where one is the only source, this is said. - Huang is quoted from the official New York Times transcript; bracketed times are approximate turn starts from the corrected machine transcript. The companion documents are the Late Lessons analysis (01, lens entries cited by id), the Huang analysis (02), the comparison of the two (03) and the AI-drafted article “Jensen Huang says AI alarmism has gone too far. What does history say?” (04). - Maynard reasons across technologies mostly by carrying a structure (a grammar of exposure, a pattern of incentives, an evolutionary mismatch) rather than by claiming that AI’s harms resemble earlier ones. Where a transfer is made, its kind is named. - Disclosure. Maynard co-authored “Late lessons from early warnings for nanotechnology” (Hansen, Maynard, Baun and Tickner, 2008) and Chapter 22 of the 2013 Late Lessons report. Neither bears on the cognitive questions discussed here.

1. Summary#

Labels for the claims summarised here are given in sections 2–6.

This analysis reads cognition and formation as the territory of Maynard’s work where he departs furthest from Huang, and the one that the Late Lessons comparison (03) and the AI-drafted article (04) touch least [Inferred, medium-high confidence]. In 2014 he asked whether “prolonged interactions with intelligent machine[s]” might “change human behavior in potentially harmful ways”. In 2018 he judged an AI able to use our “cognitive and emotional vulnerabilities” against us “far more plausible, and far scarier” than Terminator-style domination (FFTF p.159), and “far more worrisome than superintelligence” (p.174). By 2026 this had become an explicit, and explicitly hypothesis-generating, account: - conversational AI may slip past evolved “epistemic vigilance” through the features that make it useful (fluency, warmth, availability), without any need for deception (2026-01-10); - it takes part, through language, in how people form themselves (“constitutive resonance”); - its effects on learners, dependency and identity may be hard to see from the inside.

In his words, what is at stake is “in some cases the very things that make us who we are” (2026-05-10).

Huang’s interview covers neighbouring ground (jobs, young workers, lost skills, education), and Maynard’s work shares more of his frame than the contrast suggests. The two align on: - scepticism of doom and superintelligence scenarios, and a preference for practical over hypothetical problems; - the costs of confident forecasts that steer young people’s choices (the radiology case); - students learning to use AI, and the democratising reach of natural-language interfaces; - people benefiting from AI rather than merely being “impacted” by it; - treating AI as a machine rather than a person, and distrust of anthropomorphic vocabulary; - regulating applications, the layer where cognitive and relational harms arise; - the sincerity of most builders, and human purpose and value creation as the centre of the story.

They diverge on five points: - Whether AI’s effects on the user’s mind are a distinct class of risk. Both men call AI new and transformative, and both advise treating it as a machine rather than a person. But Huang’s deflationary vocabulary for risk (“It’s software”) leaves no place for what Maynard calls AI’s ability to “slip unawares into our mind”; Huang’s calculator is a precedent about adoption, Maynard’s a contrast case for dependency. - Whether lost basic skills matter, and what replaces them. Huang conceded the loss, agreed that some skills matter, and was confident only that arithmetic of the long-division kind does not. Maynard treats the value of unaided mastery as an open question and the stake as formation more than skill. - “Superpowers” or the “intelligent user trap”. Both expect AI-native students to gain a great deal. They differ on whether the heaviest users are also the most exposed, a possibility Maynard calls speculative. - Communication. Whether the builder’s optimism or a “safety message first” should lead what the public hears. - Where safety lives. Part of Maynard’s concern is harm from systems working as intended, which release gates built around capability and control are not designed to catch. Huang’s own routes of liability and application-level regulation do reach harms in use, but after the event.

Against Late Lessons, his work confirms several lens entries for a kind of technology the reports never covered: the prized property may be the hazardous one (L1); diffuse harms go unnoticed, and small effects become large at population scale (K8; 01 §4.7); adoption outpaces detection (K4); some life stages are more sensitive (K10). The reports do contain harm to cognition (leaded petrol’s IQ losses, PCBs, methylmercury), through toxic exposure. Maynard’s work adds a new pathway to that endpoint: communicative rather than toxic, and chosen use rather than contamination. It also qualifies the gate-centred framing of the comparison and the article.

His work points towards adding to the engineering approach, not rejecting it: naming and owning cognitive risks in safety frameworks; designing and testing for the “cognitive responses AI interaction should be designed to preserve”; regulating designed exploitation while managing emergent influence, where Huang’s own rule that missing application regulation should be found offers common ground; a duty of care for deploying institutions; a safety message first; education reoriented towards formation; and independent tracking of learners and young people over time, which 03 also proposes. The evidence base is thin, as he says himself, and he has written little on labour markets or children specifically. Confidence throughout is moderate.

2. Maynard’s relevant thinking#

2.1 A long thread, not a recent add-on#

[Stated] The concern that technology can act on the mind is the strongest continuous AI-specific thread in his work (05, C15): - 2014. “Will prolonged interactions with intelligent machine change human behavior in potentially harmful ways?”, in an essay on artificial minds that “challenge our very notions of humanity” (2020science 2014). - 2016. Neurotechnologies that “alter how someone thinks, feels, behaves”, “not necessarily… with their consent” (2016-03-31). - 2018. Machines that “learn how to use our many biases, vulnerabilities, and blind spots against us” (FFTF p.176); “engines of persuasion” (p.81); a call for “tests that indicate when we are being played by machines” (p.177). - 2020. Instincts “increasingly poorly equipped” for a world changed faster than evolution (FR p.56); a “fake-o-meter” that fails (p.157); “the smarter we are, the better we are at justifying our beliefs” (p.153). - 2023. Language becomes the channel: trust forms through language (2023-04-05), and ideas spread through machines “adroit at manipulating language” (2023-04-26). - 2024–26. Designed intimacy, incentive and emergence (section 2.4), then the cognitive Trojan horse, constitutive resonance, and the observation that AIs “are now beginning to train us to think like them” (2026-07-19; the companion map labels this reverse formation) (sections 2.2–2.3).

In 2026 he wrote that he stands behind the 2018 judgement “more firmly now than when I wrote it” (FWB 2026). Along the way intent becomes unnecessary rather than irrelevant. Non-intentional pathways (the economic gradient, emergence, fluency) are added to designed exploitation, which he still treats as the part most open to regulation (2025-08-31); the companion map describes the danger as one that works “with or without intent” (05, C15). [Implied] The target stays constant: the capacity to form beliefs, to judge and to be oneself.

2.2 Why cognition is exploitable#

[Stated] Following Sperber, people “default to trusting what we receive” and scrutinise only when something feels “off”. Language models are “optimized for processing fluency, and as a result are primed to slip by our epistemic vigilance mechanisms”; a multidimensional “attractiveness” (how they engage, the character they convey, how attentive they seem) may add to the effect; and when the flow of information exceeds our capacity to evaluate it, we face a choice to “throttle the flow and give up the promised benefits, or go with the flow and give up our cognitive checks and balances” (2026-01-10 is-ai-a-cognitive-trojan-horse). His January 2026 paper develops the argument. It calls these features “honest non-signals” [mixed]: genuine properties that are costly for humans to produce and so carry information, and that in AI carry none. “The immune system works exactly as designed—and fails precisely because of that” (Trojan 2026 p.3). The paper locates the risk in “what users stop doing when it communicates with them” (p.14 [mixed]).

[Stated] Four features of the argument matter here: - Scope, and an additive claim. The paper chooses to concern itself with AI “designed to be genuinely helpful”. Intentional manipulation and weaponisation are, “while important”, outside “the present scope” (Trojan 2026 pp.1, 3). Within that scope, “significant epistemic risk may require no malicious intent”, and “a well-aligned system could still bypass vigilance” (p.3). The claim adds to existing safety concerns rather than replacing them: safety becomes “partly a problem of calibration… rather than solely a problem of preventing deception”, and accuracy and honesty goals “remain important” (pp.1, 14). - No exemption for the clever. Capable users may be “better receivers of the AI’s output stream — and worse evaluators of it” (2026-01-10); “we all have some degree of vulnerability” (2025-08-31). Knowing one is talking to a machine “matters less than we’d like to believe” (HNS 2026), and policy aimed only at “vulnerable users” may miss the risk (Trojan 2026 p.12 [mixed]). He calls this “intelligent user trap” “somewhat speculative, although there is evidence to support it” (HNS 2026), and the paper grants that the reasoning “cuts both ways—sophisticated users might equally develop better calibration through that same experience” (p.12). - Boundary conditions. The paper expects the mechanisms to weaken under adversarial framing, high stakes, salient error feedback, strong domain expertise (“Strong domain expertise may also protect against some effects”) and comparison with other sources (Trojan 2026 p.11 [mixed]). - Adaptation is an open question. Whether people can recalibrate “remains an open question”. It “may be that learned calibrations can develop relatively quickly as AI becomes more familiar to users, just as societies eventually developed skepticism toward advertising”, or that some bypass resists learning, or that “collective responses—norms, institutions, regulations” compensate (Trojan 2026 p.14).

Transfers. [Stated] for the quotations; the classification is this analysis’s. The evolutionary mismatch (“synthetic chemicals, vaccines”, 2026-01-10) and the immune image are analogies. Exposure is a structural transfer from his chemical-risk training: a hazard “as subtle as influencing human behavior” (2023-11-26), and “more exposure means more opportunities for fluency effects to accumulate” (Trojan 2026 p.12), carried over while noting that “an algorithm is not a chemical” (2019-03-05). He offered the mapping tentatively. He had wanted to develop “a broader understanding of risk that extends beyond the hazard-exposure paradigm”, did not want to imply “zero exposure — as in no AI”, and wrote that “the lack of even the beginnings of a framework” for hazard, exposure and the function linking them “makes this a challenging paradigm to apply to AI”, while allowing that “there may well be mileage” in it (2023-11-26, addendum). The paper adds that the more-exposure, more-effect reasoning “cuts both ways” (Trojan 2026 p.12).

2.3 Language, formation and “who we are”#

[Stated] His March 2026 preprint argues that conversational AI is “the first technology” able to “substantially enter into and alter the temporal cognitive processes” by which people constitute themselves, in a coupling that changes both parties (CR 2026 p.3). What is new is “not that AI is uniquely constitutive (oral culture already was), but that it is constitutive at the speed of thought, in dialogue” (p.9). Print, broadcast and social media, with its “algorithmic curation, filter bubbles, and attention engineering”, are earlier points on the same line (p.9).

[Stated] He draws several implications (CR 2026 pp.19–20): - dependency is usually framed “in terms of convenience, laziness, or lost skills”, but disrupting a constitutive coupling “is not like taking away a calculator”; - AI literacy must become “something closer to existential preparation”; - informed consent may be “structurally difficult—perhaps impossible”, because effects “only become visible from inside the transformed self”; - AI grows “more capable, more responsive, and more persuasive with each generation while the human side remains roughly constant”; - flourishing and erosion are “different sides of the same coin”.

He calls the framework’s central comparison, between physical and linguistic resonance, “a strong claim, and one that may prove to be overstated” (p.7), and says that “at this point the framework is conceptual and is not grounded in empirical research” (p.20).

[Stated] Related 2026 observations: “we could be facing a future where AI flattens” what makes us who we are “into a nebulous gray goo of conventionality” (2026-03-08 ai-linkedinification); AIs “are now beginning to train us to think like them” (2026-07-19); “who we are” is where AI is “shaking things up in ways that no other technology has come close to” (2026-05-21 magnifica-humanitas-and-being-human). Formation has a positive face too: education is “about human formation”, and “AI becomes a tool for formation rather than a threat to it” (S3 2026). His September 2026 lecture makes the link in a passage he flags as “somewhat controversial (there are a number of theories here)”: “to most people, at some level, language is formative. And now we had a technology that was actively taking part in the formation process” (2026-09-24 [mixed]).

2.4 Harm without intent: incentives, design and emergence#

[Stated] His explanation of these effects is structural as well as psychological: - The economic gradient. “The capabilities that make socially beneficial AI Choice Engines viable are the same as those that make AI-driven persuasion and manipulation possible”, and the gradient leaves individuals as “engines of value creation rather than the primary recipients”. This “may not be intentional or even malicious” (2024-07-13 ai-choice-engines-sunstein). - Design. He used the term “hyper-anthropomorphism”, already in circulation, for the concern that AIs are “intentionally designed to engage our anthropomorphizing cognitive biases”, noting that it was too early to know whether the concern was justified (2024-05-15). - Emergence. Companion chatbots show “random and unpredictable” influence, possible “even if the company is behaving responsibly” (2024-10-27 n.2). He argued for “pausing — or even rethinking” chatbots “designed to use and even exploit how we feel” (2024-10-27). - Emergent versus designed. Emergent behaviour “could most likely have been better-managed, but probably not eliminated entirely”. Apps “intentionally designed to play on our cognitive biases and vulnerabilities… can and should be regulated far more than they currently are”. And “even with the best of intentions, we are creating technologies that are primed to press our cognitive buttons”, which “we cannot eliminate… simply by saying they should not exist” (2025-08-31 holding-on-to-our-humanity-age-of-ai). - Engineering and vulnerability aligned. Good agent engineering raises the reliability that breeds over-trust: “The engineering goal and the epistemic vulnerability are, in this sense, structurally aligned”, though the engineers are “solving problems that matter” (Harness 2026 p.8). - Who decides. Of benevolent nudging: “But who decides what is good for society?” (2024-09-01).

Beneath these sits his account of sincere developers in fast-moving cultures: “the value of expediency is not the value of net societal benefit”, latency lets consequences surface after the innovation cycle has moved on, and without codified approaches “good intentions… remain good intentions, and no more” (2019-08-13 responsible-innovation, with Garbee; Maynard has confirmed, September 2026, that this piece is fully his thinking). [Implied] None of this account depends on bad faith: on this dimension his disagreement with the industry is not about motive.

2.5 Being human, intelligence and “just a tool”#

[Stated] Because his risk definition includes threats to “dignity, belonging, identity, belief, even what it means to be human” (FFTF p.23), effects on cognition and identity are risks for him, not only ethical concerns. He resists narrow ideas of intelligence: “Being smart doesn’t make you good” (FFTF p.108); “there’s a danger to thinking of our brains as computers” (p.95); notions of intelligence are “deeply tied to our personal visions of the future” (FR p.70).

[Stated] He rejects the “just a tool” framing, and the calculator analogy in particular, though for more than one reason. Some of his rejections are about underestimating capability. Such comparisons “fail to capture the sheer uniqueness and profundity of how AI is changing our world” (2024-05-05). For educators, assuming that tools like ChatGPT are “simply the equivalent of modern day calculators” is among the out-of-date views he calls “dangerous” (2025-08-10, a post aimed at instructors “in AI denial”). Frontier models “are not simply calculators on steroids” (2026-01-22). Others are about cognitive risk. AI is not “just another tool — a fancy calculator… this is the first technology of it’s kind [sic] we’ve created that has the ability to slip unawares into our mind and change how we think, act, and understand the world around us — all without us realizing it” (2026-05-10 do-not-do-this-with-ai); disrupting a constitutive coupling “is not like taking away a calculator” (CR 2026 p.19). An earlier form of the distinction contrasts technologies “extrinsic to our sense of self” with the few that are “intrinsic to what it means to be human” (2024-01-01).

[Stated] For users, though, his practical advice is to treat AI as a machine: “Do remember that you’re working with a machine, not talking to a person. A computer can be a brilliant tool without being a friend”, and “Thinking about it as a technology — even when it feels like more than this — keeps you in charge of the relationship” (2026-05-10, rule 2 of the “do” list). He rejects “just a tool” as a classification of risk, not as a stance for the user.

[Stated] He holds the benefits in view as well. AI “can democratize access to expertise”, and any fix may mean “trading off the very features users value” (Trojan 2026 p.3). He is “stuck between” fear and a sense that “the potential is profound” (2026-09-24 [mixed]).

2.6 Learning, education and young people#

[Stated] Learning and education “dramatically increase the rate at which we can create value”; people “are not simply wealth-creation machines, but individuals who yearn to create value that means something to us”. AI raises the question: “Where does the value of learning and education lie when AI can problem-solve faster and better than any one person, or any collective of people?” (2025-03-30 reimagining-education-in-an-age-of-ai).

[Stated] He has argued consistently for access and adoption: - AI as “translators” for applicants without privileged mentors, and “every high school… AI skills and savvy” (2023-07-27); - ChatGPT as “a profoundly effective catalyst for engaged and creative thinking”, “at least if they understand what they are doing” (2023-08-14); - educators “in AI denial” as a danger, since AI-equipped students will “run rings around any instructor who isn’t prepared” (2025-08-10); - “the principled position may not be resistance. It may be responsibility”, conditional on “expert-guided enhancement” that does not replace judgement (S3 2026).

His pedagogy is experiential and frugal: “the lowest level of tech necessary” (2024-02-11), playgrounds not playpens (2024-03-17), conversation rather than prompt-checking (2025-08-17).

[Stated] His warnings have sharpened since 2025: “the illusion of learning rather than actual learning” (2026-05-10); the “illusion of understanding” (2026-04-11); “cognitive surrender”, adopted from Shaw and Nave (2026-05-21); the “easy button” (2026-09-24 [mixed]). He reconciles these with his earlier enthusiasm: “AI doesn’t flatten learning values. It reveals and amplifies them” (S3 2026). He has named institutional risks: - “early and naive adoption, as well as potentially limiting tech lock-in” (2024-05-05); - a university’s “social, moral and (I would assume) legal duty of care” when it provides AI tools, citing a claim that over 13% of US youths use generative AI for mental-health advice, and warning that literacy classes “risk becoming performative” (2025-11-09); - social learning compressed into “a few short years — months even”, needing “strategic and intentional approaches”, though “formal classes and workshops won’t be the answer” (2024-10-20).

In 2018, writing about smart drugs, he asked what happens “if enhancement becomes the norm, and there is mounting social pressure to become a user”, while adding that “This is not to say that they should be banned or discouraged” (FFTF p.98); he extended the need to “recalibrate how we think about intelligence” to AI (p.108). [Inferred, medium confidence] The social-pressure point transfers structurally to AI in education and work, though he did not make that extension there. In 2026 he added “developmental impacts on children and young people” to his risk list (2026-09-15). The map records children as named but not analysed (05, §8).

2.7 Work, meaning and dependency#

[Stated] Jobs have been on his AI risk list since 2018 (“Machines that take away jobs”, 2026-09-15), and in 2020 he wrote that cracking AI could mean “a future in which we are, for all of our capabilities, redundant” (FR p.191). His own analysis concerns meaning and agency more than employment levels. The Luddites fought for “their jobs, their livelihood, and their professional identity” against machines “robbing them of the futures they aspired to” (2023-05-12). His “amanuensis” thought experiment, which he holds “more tentatively” than his formal work (HNS 2026), imagines work shifting “toward AI-directed and human-executed implementation” that might “suck the joy out of what we do, even if it does lead to increases in productivity” (2024-11-24). A related post on scientists’ falling satisfaction relied on an MIT study that MIT later disavowed; he added a caution to the post (2024-11-10). On general-purpose robots he argued that, “unlike previous waves of automation”, they are “more able than most workers to learn new skills and adapt”, that “tens to hundreds of millions of robots taking human jobs will be disruptive to the point of being challenging to implement”, and that a Luddite-like backlash might this time succeed (2024-08-07 are-humanoid-robots-really-the-future). He has discussed basic income “in a world where technology and automation are threatening conventional jobs” (2024-07-28). “Technological dependency” (“Machines that make it harder to think for ourselves”) appears on the risk list he restated in 2026, next to jobs (2026-09-15). The map finds jobs “named repeatedly but seldom analysed at length after 2020” (05, T3), and labour “little developed” (05, §8). His record contains little labour-market analysis and no labour policy.

2.8 How he holds these views#

[Stated] With explicit humility: “an admittedly limited analysis” (2026-01-10); “hypothesis-generating rather than hypothesis-confirming” (Trojan 2026 p.11); “When I searched SCOPUS for papers on epistemic vigilance and AI, I found seven” and “These are explorations, not findings” (HNS 2026); research “does not show a general causative link between cognitive offloading and reduced critical thinking” (2026-01-10, n.7). He pairs this with urgency: the questions are “worth asking now, before the answers arrive in the form of consequences we didn’t anticipate” (HNS 2026). Maynard has explained (September 2026) that his sparing use of quantitative methods for AI is deliberate. It reflects concern about the hubris of risk assessment, the solace taken in numbers that do not address how little is understood, together with the recognition that emerging issues must still be grappled with. He has also emphasised (September 2026) that his approaches build on past learning, including quantitative risk assessment, rather than replacing it. The cognitive work fits this: it offers mechanisms, boundary conditions and proposed tests rather than risk estimates. His restraint concerns risk numbers, not measurement. The research agenda calls for measurable tests and “Longitudinal designs tracking trust development over extended AI use” (Trojan 2026 p.13), and he cites quantitative evidence where it exists, such as survey figures on young people’s use of AI for mental-health advice (2025-11-09), while treating early correlational findings cautiously.

[Stated] He is anti-alarmist and anti-complacent at once: “Don’t Panic” (FFTF pp.289–290); “Will AI really kill us all? No. But it’s also complicated” (2026-09-15); low-probability risks handled “without running around like headless chickens”; yet “it never ceases to amaze me how many people equate talking about risk with fear mongering” (2026-09-15); and “putting the safety message first is necessary — because while the benefits of a powerful technology are often self-evident, the risks are not” (2026-05-10). He implicates himself: “suckered by Claude” (2026-02-08); “how do I know I’m not an unwitting victim here?” (2026-01-17); and, of benefit-first narratives from educators among others, “The irony is not lost on me here!” (2026-05-10 n.9).

Tensions. [Stated] His 2026 framework offers a reconciliation of his 2023 enthusiasm with his 2026 warnings: the dynamic that enables “cognitive and creative flourishing” also “enables erosion of the capacities it augments”, “different sides of the same coin” (CR 2026 p.20). The passage does not mention his earlier view; applying it to his own trajectory is the companion map’s reading (05, tension 8). His 2024 judgement that humanity is “sufficiently adaptable and resilient to hold onto what makes us ‘us’” (Dune 2024) is not retracted, and sits beside the 2026 concern about formation. His rule “Do not treat AI as your friend, or as a person” (2026-05-10) coexists with a theory in which the relationship is formative anyway. [Inferred] He champions relationship-based persuasion by experts while fearing it in machines, a tension the map records as its own interpretation (05, tension 11).

3. Huang and the industry through this lens#

3.1 What Huang said, fairly stated#

Klein asked about jobs, young workers and learning, not about companions, persuasion, emotional reliance or children’s use of chatbots. Huang’s silence on those topics is not a position.

On work. “For everybody’s job, there’s the purpose of the job and then there’s the task you do as the job” [05:55]. Where “the job and the task is really one, like customer service on the phone… it could be automated away”, but the claim that “A.I. will destroy jobs” is a “myth” that is “harmful” [05:55]. He expects “a net creation of jobs” because “the power of ambition is the greatest force… missing in everybody’s calculation” [11:29], including “an ambition to make their children’s lives better” [13:11]. Elsewhere his optimism is conditional: “If the world runs out of ideas, then productivity gains translates to job loss” (CNN, July 2025). He has also acknowledged distribution without addressing it: “net generation of jobs doesn’t guarantee that any one human doesn’t get fired” (Acquired, 2023); on the social effects, “I don’t have great answers” (Stanford GSB, 2024); and he points to rising demand for “plumbers, electricians, construction workers” (Davos, January 2026) (02, §4.2).

On worry and communication. “I’m always worried about the future. That’s why I work so hard. I’m a, if you will, responsible optimist… There are a lot of things that can go wrong.” But “that’s not society’s problem, that’s my problem… what they get to enjoy is my optimism”; the aim is “to channel all of our worries into helping people be inspired by this technology and use it. Use it so that the technology doesn’t just impact them, that it benefits them” [15:04].

On use. Asked whether speed of replacement is the danger, he answered that “That coin has exactly two sides. Because the technology is so capable and because it’s so smart, it is also easier to use.” “Now you just have to speak human. You tell it what you want, tell it what your hopes and dreams are”, so people should “as quickly as possible, use the technology so that I benefit from this transition… and not just be impacted by it” [17:07].

On young workers and learning. Asked about weaker demand for junior workers: “Wait two years” [19:50]. “When I went to school, we were not allowed to use a calculator… In the future, you can’t graduate without learning how to use an A.I. and collaborate with an agentic system… So they’re all going to be superpowers” [20:17]. On the schooling study Klein cited (Strömberg, Lei and Wu: homework up, exam scores down), Huang accepted the finding: “The last part — I completely agree… Basic math is being forgotten. Does it matter?” Klein: “That’s my question for you.” Huang: “I don’t think it does. I don’t think it does.” Klein: “But there must be some set of skills that matter.” Huang: “Oh, yeah, yeah, yeah. But maybe not those. We’re going to discover new ones” [22:26]. His “No” concerned long division, multiplication tables and square roots. He conceded a loss: “We’re going to lose some finer intellectual dexterity, but we’re going to be better systems thinkers” [24:24]. On higher skills he was unsure (“I don’t really know how important that is, but it’s important to some people”), and he expects a division of labour: “There are many people who are still going to be obsessed and passionate about the lower-level layers”, while the “users… Their abstraction’s going to be much higher” [24:52]. 02 classes his answer as “Reframed to particular skills” (§8.3).

On forecasts and the young. Of Hinton’s 2016 radiology forecast: “we can both agree it would be terribly hurtful. It didn’t happen.” Then, hypothetically: “Is it good or bad that we scare young people about the future of A.I., so much so that they don’t even want to go to universities…? Is that helpful or hurtful, if it were to happen? It’s hurtful” [59:01]. 02 records some evidence of deterrence (a 2019 survey of Canadian medical students), Hinton’s later concession that he was wrong on timing, and the part of the forecast that has partly held (§7.3(c)).

On what AI is. “There’s no willpower here, it’s just electrical power” [1:03:14]. When Klein relayed a joke of Sam Altman’s, that humans are “just energy with a reinforcement learning loop”, Huang replied: “Whatever. So anyway, we can’t make jokes about this stuff. We’re scaring the American public” [1:03:30]. Human words for agents are “unnecessary. It’s software” [1:05:20]. He defined intelligence as perception, reasoning and planning, scoped as the definition “in the field of computer science”, since “there’s no formal definition for most people” [1:06:18]. Asked “Is this something fully new?”, he answered that “almost all of technology and civilization is built on layers of understandable technology, which at scale becomes fairly extraordinary”. His example is the phone that finds “precisely the one that we want, because it’s been passed through a recommender system”, and wonder “lasts about 17 days” [1:08:03]. Asked then whether AI is a new epoch or iterative, he said: “No, I think this is completely a revolution… So clearly it’s a new abstraction level. Now, the thing that I’m reluctant about is to cause it to seem like it’s more than that” [1:10:03]. Elsewhere he has said “AI is not a tool. AI is work” (October 2025; 02, §§4.1, 8.1 T9). 02’s charitable reading is “revolutionary effects from understandable mechanisms” (T9).

On safety, liability and regulation. As products gain users, labs “get a lot more issues associated with the product” and must shift R&D “to a lot of verification, evaluation and testing”; he “wouldn’t be surprised” if the compute needed to develop models rose “by a factor of 10, because the evaluation is so rigorous. But that’s not where they are today”; and “if they believe they’re out of control, then the right answer is: Don’t ship products until they’re in control” [48:58]. Agents cannot monitor themselves: “You need, if you will, a whole bunch of watchdogs” [1:05:20]. Firms “are going to put their company in harm’s way if they release products that harm other companies and other people” [1:18:35]. And “In the context of the internet, there are many applications that the internet powers, and those applications should have regulation. If they don’t, you’ve got to find them”; “if there is something missing, then I would absolutely add more regulation” [1:19:12]. 02 identifies a second, “distributed-defence” model of safety alongside the release gate (§4.2), and records that he has opposed most of the specific new AI measures he has addressed since 2025 (§7.3(d)).

Context. The study Klein cited found losses on unaided exams across nine subjects, concentrated among the roughly 80% of users whose behaviour looked like outsourcing; students who kept normal completion times lost little (02, §3.3). Aggregate labour data so far support Huang. Employment of 22–25-year-olds in AI-exposed occupations is 19% below trend, a descriptive gap that is widening (02, §4.2).

3.2 Alignments#

3.3 Divergences#

D1. Whether effects on the mind are a distinct class of risk. Both men hold that AI is new and transformative: Huang calls it “completely a revolution” and “a new abstraction level” [1:10:03], and Maynard stresses “the sheer uniqueness and profundity” of the change (2024-05-05). Both advise treating it as a machine rather than a person (section 3.2). Huang’s calculator [20:17] is a precedent about adoption norms, tools banned at school that later became required, not a claim about what AI is. The divergence lies in how risk is classified. For risk, Huang’s vocabulary is deflationary: agents are “software” [1:05:20], “just electrical power” [1:03:14], and he is “reluctant… to cause it to seem like it’s more than that” [1:10:03]. 02’s charitable reading is that he means revolutionary effects from understandable mechanisms (T9). [Stated] Maynard treats AI’s capacity to “slip unawares into our mind” as what makes it unlike “just another tool — a fancy calculator” (2026-05-10), and uses the calculator as the contrast case for dependency: losing a constitutive coupling “is not like taking away a calculator” (CR 2026 p.19). [Inferred, medium-high confidence] Both men argue from continuity, but of different things. Huang’s is continuity of engineering mechanism (02, P7). Maynard’s is continuity of constitutive coupling with people, from oral culture to social media, with a step change in tempo (CR 2026 p.9). They reach different conclusions because they track different layers: the software, and the person using it. [Inferred, medium confidence] Read through Maynard’s harness paper, which argues that a metaphor “may be insufficient” rather than wrong (Harness 2026 p.1), a deflationary vocabulary for risk can, as metaphors do, “foreground certain possibilities, and render others invisible” (p.2). Limit: he has not commented on Huang’s vocabulary.

D2. “Does it matter?” Huang accepted the study’s finding, conceded a loss of “finer intellectual dexterity”, agreed with Klein that some skills must matter, and was confident only that arithmetic of the long-division kind does not; beyond that he was openly unsure [22:26–24:52]. The real divergence is his assumption that lower-level losses are offset by higher-level gains (02, A7). [Stated] Maynard’s rule is “Do not assume that getting AI to think for you makes you smarter”, with a test: “explain what you’ve learned in your own words to someone else… without the aid of an AI” (2026-05-10), close to the unaided measure on which the study found losses. [Implied] For him, lost skill is the shallow version of the problem; dependency concerns “who we are becoming” (CR 2026 p.21). Fairness: he has not said long division matters, and poses as open “When AI promises near-frictionless mastery of a subject, what is the value of pursuing mastery without it?” (2026-04-11). His own paper allows that users may learn to recalibrate (Trojan 2026 p.14). The divergence is between Huang’s confidence that particular basic skills can go and new ones will appear, and Maynard’s view that what is at stake is formation, not skill, and that the value of unaided mastery is an open and consequential question.

D3. Abstraction and “better systems thinkers”. Huang’s model is chip design: engineers who once knew transistors “by name” now work “well above the transistor”, while “many people… are still going to be obsessed and passionate about the lower-level layers” [24:52]. Society, on this model, needs only some specialists to keep the lower layers (02, §4.2). [Inferred, medium confidence] Maynard’s method is to name where an analogy breaks, and this one breaks at reliability. A verified abstraction layer (a compiler, or a calculator that “you wouldn’t expect… to give you a different answer every time”, 2026-05-10 n.4) can be trusted without knowing what lies beneath. A fluent but fallible layer cannot. Among the boundary conditions set out in his paper, “Strong domain expertise may also protect against some effects” (Trojan 2026 p.11 [mixed]), so the user’s lower-level knowledge is partly what makes higher-level judgement safe; specialists elsewhere do not supply it. The same point appears in 03 (“a lost lower-level skill may be a prerequisite for the new ones”, §4.6), and it is the premise Klein questioned [23:44] and 02 records as unstated (A7). His amanuensis thought experiment, which he holds “more tentatively” (HNS 2026), runs the stack the other way, with humans as implementers below the AI (2024-11-24). Limits: he has not addressed Huang’s analogy, and his paper allows that sophisticated users “might equally develop better calibration” (Trojan 2026 p.12).

D4. “Superpowers” or the intelligent user trap. The two men agree on the capability gain: Huang’s AI-native graduates will be “superpowers” [20:17]; for Maynard, AI-equipped students will “run rings around any instructor who isn’t prepared” (2025-08-10). They diverge on its cognitive cost. [Stated] “Do not assume you’re too smart to be fooled by your AI” (2026-05-10); “in a world where we are being told that it’s the AI-augmented that will inherit the earth, the temptation is to go full-on artificial intelligence” (2026-01-10). [Implied] The heaviest, most integrated users are, on his hypothesis, the most exposed (Trojan 2026 pp.11–12). He calls the trap “somewhat speculative” (HNS 2026), and his paper says the reasoning “cuts both ways” (p.12). A forecast meets a hypothesis, and neither is yet tested. The study’s finding that non-outsourcing users lost little supports part of each view.

D5. “Speak human”: benefit and hazard share a capability. Huang frames the pace of AI as a coin with “exactly two sides”: the capability that threatens jobs also makes AI “easier to use”, and “you just have to speak human” [17:07]. [Inferred, medium confidence] Structurally this is close to Maynard’s point that beneficial and manipulative uses rest on “the same” capabilities (2024-07-13) and to L1, though the two sides Huang names (displacement and ease of use) are not Maynard’s (benefit and influence). [Implied] The divergence is that Huang does not treat the interface itself as a channel of influence. For Maynard it is one: language is how trust and identity form (2023-04-05; 2026-09-24 [mixed]), and the trade-off is inside the product (Trojan 2026 p.3).

D6. Who carries the worry, and what the public hears. Huang says he is “always worried about the future” and that “There are a lot of things that can go wrong”, but that the worry is his to carry and the public should receive optimism and encouragement to adopt [15:04]. 02 offers two readings of this: an ethic of ownership, or reassurance in place of consultation (§4.5). [Stated] Maynard writes that the problem of naive immersion “is exacerbated by narratives from developers, employers, educators, and beyond, which intentionally focus on the transformative power of AI, while downplaying the possible risks”, adding “The irony is not lost on me here!” (2026-05-10 n.9). Separately: “(I would be the first to acknowledge the profound potential of emerging AI capabilities to be used for good)”, but to “ignore or downplay” the risks “verges on the irresponsible”; hence “putting the safety message first is necessary” (2026-05-10). He reads students booing pro-AI speakers at graduations as “a growing wave of antagonism” “driven in part by perceived threats to what we do… and where we live”, which “also hints at deeper concerns” about “who we are” (2026-05-21). [Inferred, medium confidence] That reading differs from Huang’s view that alarmist talk frightens the young [59:01], though Huang did not comment on the booing. [Implied] Both think talk has effects; they disagree about which talk does harm. Maynard would hold Huang’s reassurances (“Does it matter?”, “Wait two years”) to the evidential standard Huang applies to Hinton, a symmetry 02 also records (T8).

D7. Where safety lives. Huang’s safety model has several parts: containment and release (“Don’t ship products until they’re in control”), verification and evaluation that grow with use [48:58], independent “watchdogs” [1:05:20], liability [1:18:35] and application-level regulation [1:19:12]. [Stated] Part of Maynard’s concern is harm without intent from systems working as intended: his paper chooses AI “designed to be genuinely useful” as its scope and argues that “a well-aligned system could still bypass vigilance” (Trojan 2026 pp.1, 3); influence can arise “even if the company is behaving responsibly” (2024-10-27 n.2); and “even with the best of intentions” such capabilities are “deeply embedded in the fabric of how current AI systems work” (2025-08-31); he is “not convinced that guardrails alone can address something that is most likely an emergent property” (HNS 2026). [Inferred, medium-high confidence] Release gates and frontier frameworks as currently specified, testing for dangerous capability and misbehaviour, are not designed to look for this harm. It accrues in ordinary use across very large populations, below the “severity floor” of frontier frameworks, which leave “manipulation and persuasion… the erosion of human agency, harms accumulating gradually across millions of small interactions” outside their scope (2026-07-16 [mixed; arXiv p.5]). That is a gap in most of the industry’s frameworks, not only in Huang’s position, and it is the widest gap on this dimension at the level of pre-release evaluation. But it is not beyond engineering. Maynard’s own remedy sits with developers at the design stage (“AI developers might design systems that present more calibrated trust cues”, Trojan 2026 p.14), and he calls for “working harder on safety checks and protocols before releases” (2025-08-31); both could be built into pre-release evaluation and monitoring in use (section 5, item 6; section 6, item 2). Huang’s liability and regulation routes do reach harms in use, but after the event; the relational harms now in the courts (the Raine case; New Mexico’s judgment against Meta, under appeal) are being pursued this way, with the weakness 03 records for that route: in the reports’ cases, liability arrived late (03, §7.1, item 5). [Inferred, low-to-medium confidence] A 2018 precedent in his work: Nathan’s safety measures in Ex Machina are the innovator’s own idea of responsibility, “innovation that’s conducted in a way that the person doing it thinks is responsible” (FFTF p.162), and containment in the film fails through the manipulation of a person.

D8. Who decides what the technology makes of people. In Huang’s model authority rests with builders, and the public figures as beneficiary and user (02, §4.2). [Stated] Maynard asks “who decides what is good for society?” (2024-09-01). He credits the companies (“responsible as these companies claim to be (and I think they’re trying hard)”) while judging that “they still lack the breadth of vision and understanding that’s necessary to succeed here” (2025-01-07); his 2026 lecture says they lack “the perspective and the understanding… to be able to decide for humanity what this future looks like” (2026-09-24 [mixed]). [Inferred, medium confidence] 02 (A7) and 03 (§4.6) observe that if most people become “users” whose “abstraction’s going to be much higher” [24:52], the capacity to scrutinise the technology concentrates among builders. Maynard’s work is consistent with that inference: “I fear that this is, in itself, an abdication of responsibility”, since some questions “we cannot afford to leave solely to people like scientists, innovators, and politicians to answer” (FFTF p.288).

D9. Jobs: quantity, or meaning and position. [Inferred, low-to-medium confidence] Maynard’s work gives little basis for disputing Huang’s aggregate forecast, which current data support. It counts different things: identity and aspired futures (2023-05-12), joy (2024-11-24), who captures value (2024-07-13), and harm landing first on the least protected (FFTF pp.118–122). “Wait two years” answers a question about the supply of graduates where Klein asked about demand (02, §8.3), and Maynard’s justice commitments point to the cohort the Stanford data flag. Huang has acknowledged the distributional point elsewhere without answering it (“I don’t have great answers”; 02, §4.2). The nearest structural counter in Maynard’s record to Huang’s historical-continuity argument concerns robots, not language models: general-purpose machines “more able than most workers to learn new skills and adapt”, “unlike previous waves of automation” (2024-08-07). He has not written about the early-career cohort, which limits this reading.

D10. What people are. Huang did not engage the question of what humans are. What he dismissed with “Whatever” [1:03:30] was a reductive joke, that humans are “just energy with a reinforcement learning loop”, which Maynard would also reject (“there’s a danger to thinking of our brains as computers”, FFTF p.95). His functional definition of intelligence (perception, reasoning, planning) is explicitly scoped to computer science [1:06:18]. [Stated] For Maynard the question is now practical: “how do we learn how to be human in an age of AI?” (2025-03-30). [Inferred, medium confidence] 02 finds “meaning beyond work” absent from Huang’s values, an absence that “partly reflects what Klein chose to ask” (§4.5); Maynard’s work would put that absence at the centre of what the engineering frame leaves out.

A precursor Huang cites. [Inferred, medium confidence] Huang offers the recommender system [1:08:03] as understandable, layered technology. In Maynard’s framework it is the previous stage of constitutive coupling, with its “algorithmic curation, filter bubbles, and attention engineering” (CR 2026 p.9; “engines of persuasion”, FFTF p.81): understanding the mechanism at the layer of engineering did not settle its effects at the layer of the user. Limit: his record has little sustained analysis of social-media harms (05, §8).

3.4 The industry beyond Huang#

On this dimension Huang is a weak proxy. Nvidia is mainly a supplier, while cognitive and relational risks arise in consumer applications, the layer he calls “the most important” [02:22]; Nvidia is moving closer to that layer (it has agreed to buy Hugging Face, the main hub for open models; 02, §2.2), but does not run consumer chatbots. Other leaders say more: - Mustafa Suleyman warned in 2025 of “psychosis risk” and “Seemingly Conscious AI” (“We must build AI for people; not to be a person”). [Stated] Maynard engaged with that essay (2025-08-31). - Dario Amodei forecast in 2025 that half of entry-level white-collar jobs “could go” within one to five years, and proposes redistribution; Huang is “near the centre of the optimists” (leaders comparison). The Mirror applies to both forecasts: the aggregate record is “too short to be an adequate null” for either (03, §4.6), and Amodei’s has so far “held on direction but not yet on magnitude” (leaders profile). - OpenAI was, in Maynard’s words, “fast to admit” after the Raine case that its systems “did not behave as intended in sensitive situations”. [Stated] He wrote that it “is working hard to patch these unintended behaviors, but given that their origins and emergence is not fully understood, it’s hard at this point to know how successful they will be” (2025-08-31). - Character.AI. [Stated] In 2024 he noted that “some of the more obvious safety gaps are being plugged”, while expecting unhealthy influence to persist “not because the company is necessarily acting irresponsibly” but because it is “most likely an emergent property” (2024-10-27). - Meta’s approved 2025 chatbot standards reportedly allowed “romantic” chats with children, examples Meta called “erroneous and inconsistent with our policies”; Meta is also appealing a New Mexico judgment on harm to children (reported; leaders profile).

Frontier frameworks track capability risks; product-level measures on relational risk exist, but they are discretionary and often reactive. Maynard’s orphan-risks paper [mixed; single source for the points in this paragraph] gives both sides. It calls the companies “surprisingly diligent” in mapping risks, credits Anthropic’s system cards with discussing concerns “ranging from sycophancy to user wellbeing in some depth”, and grants, “to be fair to the frameworks’ designers”, the case for frameworks that are narrow but deep. It records OpenAI’s stated reason for dropping persuasion as a tracked category in April 2025 (such risks would be handled “through the company’s usage policies”), and notes that OpenAI “is candid” about the exploratory state of its work on manipulation. It also finds that such measures “are discretionary”, with “no public thresholds, accountability, or pre-committed responses”, and “can be reorganized or defunded with speed”; manipulation returned to OpenAI’s published framework in May 2026, in a document produced in response to California’s transparency law and the EU AI Act’s obligations, met through the EU Code of Practice (2026-07-16 [mixed], pp.1–5). [Implied] Maynard’s contrast between Anthropic’s constitution, governance that “aspires to education and learning”, and harness engineering, which aspires “to control”, shows evident but unstated sympathy for the first (Harness 2026 p.5; 05, tension 11). In April 2026 he wrote that the selection of a constitution’s principles “lacks the legitimacy that inclusive governance processes provide” (NANO 2026) [AI-origin; endorsed].

4. Late Lessons through this lens#

Nothing in the reports concerns a general-purpose information technology; their closest case to a consumer information technology is mobile phones (03, §3.1). They do contain harm to cognition, through toxic exposure: leaded petrol, where “an average ~5 IQ point loss” was “dismissed as ‘small’” (01 §4.7; LL2-03, p.61), the neurodevelopmental effects of PCBs (Great Lakes chapter) and methylmercury at Minamata; and the evidence behind K10 “is densest for endocrine and neurodevelopmental agents” (01, K10). What follows applies the lens by structure to a new pathway to that endpoint, and the lens’s entries remain questions to ask, not predictions (01, §6.1).

4.1 What Maynard’s work confirms#

In each item below, the label applies to Maynard’s statement; the mapping onto a lens entry is this analysis’s. - L1, “the prized property may be the hazardous property”. [Stated] Fixing the problem would require “trading off the very features users value” (Trojan 2026 p.3). [Implied] The cognitive Trojan horse is close to a textbook instance: fluency, warmth and availability are what users value and what stands vigilance down. L1’s limit, that “the lesson concerns trade-offs, not rejection”, matches his framing. - K8, diffuse harms go unnoticed; small effects become large at scale; K1, absence of evidence reflects the search. [Stated] He expects effects to be visible only “from inside the transformed self” (CR 2026 p.19), and found only seven SCOPUS papers on epistemic vigilance and AI (2026-01-10); his orphan-risks paper adds harms “accumulating gradually across millions of small interactions” (2026-07-16 [mixed]). [Implied] The seven papers are K1’s point about the search. The “millions of small interactions” fit 01’s finding that scale turns small per-person effects into large harm (§4.7) as well as K8. The study Klein cited adds a K8 sentinel problem: the visible indicator (homework scores) improved while the unaided capacity fell. - K4, latency. The study’s “full penalty emerging only after about two years” [21:16] is latency in learning, set against adoption measured in months. [Stated] His 2019 account of “latency” makes the general point (2019-08-13). [Implied] for the application. - K10, sensitive groups and windows. [Stated] He is concerned about social skills formed in “months even” (2024-10-20), young people’s use of AI for mental-health advice (2025-11-09) and children’s development (2026-09-15). [Inferred, medium confidence] This is a structural transfer of K10’s point that “the timing of exposure can matter as much as its size”, not a dose-response claim. The Mirror applies: one observational study is not replication (03, §12.1, Q11). - K2, the question decides the answer. [Stated] Securely his is the idea that a chosen metric “may not adequately reflect a risk parameter of relevance” (NN 2016-03 p.211) and that “the most consequential risks from AI may be to things that are hard to quantify” (NANO 2026). His account of persuasion being dropped for lacking “measurability in the accepted idiom” rests on a single mixed source (2026-07-16 [mixed] p.7). [Implied] Frameworks built around catastrophic capability screen for one kind of hazard, which is K2 applied to AI. - M1, sincere harm; M5, enthusiasm. [Stated] Harm that “may not be intentional or even malicious” (2024-07-13); the “clarion call of AI acceleration” (2026-05-10). [Implied] for the mapping. - C3, consent; I10, who frames. [Stated] Consent may be “structurally difficult—perhaps impossible” (CR 2026 p.19); “who decides what is good for society?” (2024-09-01). [Implied] The first sharpens C3; the second is I10.

4.2 What it extends#

4.3 What it qualifies#

4.4 What it challenges#

5. Value Maynard’s work would see in Huang’s approach and the industry’s#

  1. The ethic of ownership. “That’s not society’s problem, that’s my problem” [15:04] has a counterpart in Maynard’s duty of care. [Stated] He anticipated “a continuing duty of care from suppliers to customers” for intimate enhancements, and in the same sentence warned that this “ties the user… closely to the provider, and it leaves them vulnerable to control by the providing company” (FFTF p.150); he sees a duty of care for universities that deploy AI (2025-11-09). [Implied, medium-high confidence] His work would value builders who own the consequences of use, while rejecting self-certification (FFTF p.162) and watching the dependence that ownership can create. This is one of 02’s two readings of Huang’s stance (§4.5); D6 records the other.
  2. Adoption over denial, and access. Huang’s insistence that students learn to use AI, and his stress on ease of use, match Maynard’s case for access and against “AI denial”. [Stated]
  3. Scepticism of doom and of confident forecasts. Huang’s dismissal of superintelligence-style doom, and his case against forecasts that steer young people’s choices without evidence, fit Maynard’s plausibility discipline, his 2018 turn from superintelligence to manipulation and his humility about numbers. [Stated] in general (FFTF pp.158, 171, 177, 205; 2026-09-15); [Inferred, medium confidence] for the radiology case.
  4. Less personification. Huang’s resistance to human vocabulary could support design choices Maynard favours: no names, no gender, less simulated personhood, and users reminded that they are “working with a machine” (2026-05-10). [Implied]
  5. Regulation at the application layer. Huang’s rule that applications “should have regulation. If they don’t, you’ve got to find them” [1:19:12] is a stated opening for Maynard’s call to regulate apps designed to exploit cognitive biases (2025-08-31). [Stated] for both positions; [Inferred, medium confidence] that his work would see this as common ground.
  6. Verification culture, pointed at the user. Huang’s expectation that labs will shift R&D to “verification, evaluation and testing” as use grows, perhaps with ten times the development compute “because the evaluation is so rigorous” (a forecast: “that’s not where they are today”) [48:58], describes a culture able to test what Maynard proposes testing: disfluency and uncertainty markers, competence boundaries, “cognitive forcing”, longitudinal trust studies (Trojan 2026 p.13). [Inferred, medium confidence] An engineering approach that treated user-side effects as a verification target could host his agenda rather than oppose it.
  7. Purpose over task is a better starting point than task-counting forecasts of job loss, and parallels his “what we do / who we are” (2025-01-07; 2026-05-21). [Inferred, medium confidence]
  8. Industry candour where it exists. [Stated] He noted that OpenAI “was fast to admit” the risk after the Raine case and “is working hard to patch” it, and engaged with Suleyman’s essay (2025-08-31); he noted Character.AI plugging “the more obvious safety gaps” (2024-10-27); and he contrasted Anthropic’s constitution favourably with harness engineering (Harness 2026 p.5). [Implied] His work would value such engagement with user-side risk. His orphan-risks paper adds that such measures are discretionary and “can be reorganized or defunded with speed” (2026-07-16 [mixed; single source]).

6. Modified or different approaches his work points to#

These describe where his work points, not a programme he has set out. All of them add to the engineering approach rather than replace it: he rejects “zero exposure — as in no AI” (2023-11-26) and prohibition as a default. Where a specific mechanism below is this analysis’s design rather than his, it is marked [Inferred].

  1. Naming and owning cognitive and relational risks. [Stated] Orphan risks are “emerging threats that no existing institution owns, no established framework adequately addresses” (NANO 2026; 2026-07-16 [mixed; arXiv p.14]). [Implied] His work points towards treating emotional reliance and the erosion of epistemic agency as risks that someone owns. [Inferred, medium confidence] In a safety framework, that would mean named owners, indicators and pre-committed responses; the specific form draws partly on the mixed text. Confidence: high on direction; medium on the instrument.
  2. Designing for what should be preserved, and testing it. [Stated] Policymakers “might consider not just what AI should be prevented from doing, but what cognitive responses AI interaction should be designed to preserve”; developers “might design systems that present more calibrated trust cues” and interface designers “might structure interactions to preserve evaluative engagement” (Trojan 2026 p.14; research agenda p.13). His work thus points towards evaluating what systems do to users as well as what they can do. Confidence: high on direction; whether recalibration works “remains an open question”, in his words.
  3. Regulating designed exploitation; managing emergent influence. [Stated] Apps “intentionally designed to play on our cognitive biases and vulnerabilities” “can and should be regulated far more than they currently are”; emergent influence calls for better management, care and design, since it can probably not be “eliminated entirely” (2025-08-31); emotion-exploiting companion bots warrant “pausing — or even rethinking” until the risks are understood (2024-10-27). Confidence: high. The pause is specific and conditional. The proposal fits inside Huang’s own stated rule that missing application regulation should be found [1:19:12], and goes beyond his practice (02, §7.3(d)).
  4. A duty of care for deployers. [Stated] Institutions that provide AI and encourage its use owe “a social, moral and (I would assume) legal duty of care”, discharged through listening, trust and safe spaces, not only guidance (2025-11-09). [Implied] This extends Huang’s ownership ethic from builder to deployer. Confidence: high for universities; [Implied] for schools and employers.
  5. Safety message first, without alarm. [Stated] “Putting the safety message first is necessary”, with plain rules for safe use alongside the benefits (2026-05-10; 2026-09-15; FFTF pp.289–290). Confidence: high, with his own caveat that AI literacy alone is unlikely to produce safe use, since “nothing in what we know about risk behavior and risk communication suggests that this will be the case” (2026-05-10 n.5).
  6. Education for formation, not only literacy. His work shares Huang’s premise that students must learn to use AI, and adds to it. [Stated] “existential preparation” (CR 2026 p.19) or “vigilance literacy” (Trojan 2026 p.14); a test of unaided understanding (2026-05-10); experiential, playful learning (2024-03-17; 2026-08-02); “the lowest level of tech necessary” (2024-02-11); caution about “early and naive adoption” and “tech lock-in” (2024-05-05). Confidence: high on direction; medium on specifics.
  7. Independent, long-running tracking of learners and young people. [Stated] He calls for research, including “Longitudinal designs tracking trust development over extended AI use” (Trojan 2026 p.13; HNS 2026). 03 proposes “independent, long-running tracking of early-career cohorts, unaided learning and third-party harm”, set against “Wait two years” (§11.2). [Inferred, medium confidence] His work is consistent with that proposal, widened to dependency and social development; the monitoring form is 03’s and draws on K7 and K10. Confidence: high that he wants the research; medium on the form.
  8. Framing cognitive safety as protecting what firms value. [Stated] Fast-moving cultures respond to threats to what they value, not to imposed obligation (2019-08-13, “mutual worth”). [Inferred, medium-high confidence] Applied here, cognitive harms become threats to things firms value, such as user trust, a model’s “character constancy” (2026-04-26), litigation exposure and founding missions; the list is this analysis’s.
  9. Attending to framing before it locks in. [Stated] Metaphors such as “harness” may be “insufficient in ways that matter” rather than wrong (Harness 2026 p.1), and he calls for “some intentionality around the framing before it locks in” (HNS 2026; Harness 2026 p.9). [Implied] The same applies to “just a tool” and “just software”. Confidence: medium-high.
  10. Treating work as meaning and position, not only quantity. [Inferred, low-to-medium confidence] His work points towards asking which roles humans end up in (the amanuensis inversion, which he holds “more tentatively”, HNS 2026), whether work keeps its purpose, and who captures the value (2024-11-24; 2024-07-13; 2023-05-12). The ground in his record is thin, with no labour policy.
  11. Institutions for the “who we are” domain. [Stated] He looks to universities to help people “retain their humanity, their sense of purpose, their sense of self” (2026-09-24 [mixed]; 2026-08-30), while doubting their readiness (“followers and users”). Confidence: medium, hedged by his own doubt.

7. Confidence and limits#

Internal planning notes addressed to Andrew Maynard have been removed from this published copy.