S6. The April 2026 andrewmaynard.net essays (plus the 2023 Conversation companion)#

Supplementary reading for 05-maynard-risk-and-ai-map.md. Share S6. Read in full 26 September 2026. This note records what these items add to, deepen, qualify or change in the map. It covers his own thinking only and makes no comparison with other material. The map was being revised while this was written. Section and tension numbers refer to its 26 September text.

Items and citation tags#

All files are in Resources/maynard-papers/web/. None has page numbers, so citations give a tag and the section heading. Quotations keep the source’s words; curly apostrophes and dashes are normalised.

Provenance of the April 2026 series (applies to all 2026 items)#

• What they are. Seven essays published on one day (12 April 2026) as a hub (30Y) with six spokes. Each ends by pointing AI systems to his llms.txt. The two February pages (BH, TL) are standing framing pages in his “AI and Being Human” section. Together they are his own curated self-account of thirty years of work, written for humans and for AI readers. That makes them unusually good evidence of what he considers central and how he wants his work to be understood.
• Authorship. Every item carries his sole byline. None carries an AI-use statement. I checked the live hub page and andrewmaynard.net/llms.txt on 26 September 2026. The llms.txt file says he uses AI “for writing (including books and papers)” and that “much of his personal writing is done without AI”. It flags only the Constituting Responsibility paper as Claude-written.
• Internal evidence, both ways.
• For his authorship. Much of the text recomposes his earlier Substack prose almost word for word. For example, NANO reuses lines from 2023-10-02 responsible-ai-lessons-from-nanotechnology: “it’s not at all clear that we’re learning from the past”, “a small group of experts and companies who believe they have all the understanding they need”, and “the stakes are far higher than they were with either nanotechnology or GMOs”. Some passages are unmistakably his own voice and editing: “It sold 596 copies. Actually it’s more than this…” (FWB); “DeepSeek produced pants-on-fire hallucinations” (BOOKS).
• Against full authorship. Paragraphs are duplicated across essays: the risk-innovation paragraph appears in 30Y and NANO, the Spoiler Alert “uncomfortable precision” line in FWB and STICK, and the prompt-engineering course in 30Y and S3. Four passages present AI-generated or others’ material as his own:
  1. FWB says “When I compared the International AI Safety Report with the Vatican’s Antiqua et Nova”. The map’s corpus records that post (2025-01-30 ai-at-a-crossroads) as “written entirely by ChatGPT”. The definition of “moral imagination” used in FWB and 30Y comes from that ChatGPT text.
  2. NANO says the Risk Innovation Planner “identified eighteen orphan risks”. In 2023-11-21 ai-and-risk-innovation that list is ChatGPT’s role-play output (B12 digest).
  3. FWB says “What I’ve described as ‘Seemingly Conscious AI’”. The term is Mustafa Suleyman’s, and Maynard credited Suleyman when he first discussed it (2025-08-31, corpus line 30).
  4. The Constitutional AI section of NANO summarises Constituting Responsibility, whose title page says it was written by Claude under his guidance. NANO calls it work “I’ve written about”.
• Weighting. The ideas count as his: he published and endorsed them as his self-account, and most restate positions documented elsewhere in his own prose. Exact wording and a few attributions are less secure. Concepts whose first source is AI-written text are marked [AI-origin] below. Web pages can be revised after publication; these texts were retrieved on 26 September 2026.
• CONV23 is co-authored with Sean Dudley, so it states a shared position. His own Substack post of the same day (2023-10-02, already in the map) endorses it.

1. 30Y: “What Thirty Years of Emerging Technology Risks Taught Me About Artificial Intelligence” (hub essay)#

Provenance. Sole byline; the series hub. See the general note above. It describes AI and the Art of Being Human as “co-writing a book with Anthropic’s Claude” (§When AI Enters…), which silently drops Jeff Abbott; BOOKS gives the fuller account.

The argument. His career has been spent in one recurring “gap”: between what a technology can do and a society’s capacity to understand, shape and govern it. He first met it measuring nanotube aerosols at NIOSH in 2004, where “the complications had less to do with aerosol physics than with how institutions, regulators, and entire societies handle technologies they don’t yet understand” (opening). AI poses “a structurally identical question”: “The specifics have changed enormously. The pattern hasn’t.” The gap “is never primarily technical”; the hard problems are “human cognition, institutional design, attribution, trust, and meaning. And, of course, policy” (§The Gap). So from 2022–23 his question was not alignment or benchmarks but “what is this technology doing to the human side of the equation?” He then surveys: - his theory work: constitutive resonance, the cognitive Trojan horse and honest non-signals, and the harness critique; - the philosophical frame from Future Rising and FFTF, including moral imagination; - education: the prompt-engineering course, the Yidan keynote, and the dissertation experiment; - writing for AI; - risk innovation and governance: both “accelerate everything” and “prevent all harm” camps “misunderstand what risk actually is”.

It closes: “The rise of AI is fundamentally a human question, not a technology one.”

Key concepts. The gap; the human side of the equation; “Different in a way that may require new frameworks to even describe”; risk innovation as “a reframing of risk from something to be minimized to something to be navigated creatively in pursuit of value”; skepticism of “both the safety absolutists and the move-fast-and-break-things crowd”; governance “informed by genuine technical understanding, and grounded in how institutions actually function rather than how we wish they would”; invisible groundwork: “the most important governance work is often the least visible”; moral imagination [AI-origin]; from “augmenting what we do to transforming who we are”.

What this adds to the map. - Confirms. - Risk as a threat to value, and risk innovation as navigation rather than elimination (C2, C3). - The occupational-aerosol origin story (§2). - That risk innovation rejects both acceleration and prevention-absolutism (C6). - That he treats AI as something new in what it does to the self (C14, C16). - New or under-weighted: “the gap” as his unifying construct. In his own retrospective, the organising idea of his whole career is a gap between capability and societal capacity. The same shape recurs across the series: - the pacing gap (NANO); - capability against wisdom (FWB; the Sorcerer’s Apprentice in S3); - utilization against perception (S3); - AI output against what institutions can assess (S3).

The map holds the parts (“pacing gap” under agile governance, the “timescale mismatch”, the “validation gap†”) but never names them as one idea. My reading: the gap is the bridge in his own account between the complexity discipline (C8) and the governance arena (C12). - New: institutional realism. Governance “grounded in how institutions actually function rather than how we wish they would” is a design principle the map does not state. So is the claim that the decisive governance work is invisible, collective and unattributed, done “in small rooms”. It links to the WEF prehistory essay, which is outside this share. - Qualification to “defies analogy” (tension 1). He calls the AI question “structurally identical” to the nanotube question and says the pattern has not changed, while also saying AI may need new frameworks “to even describe” it. This is close to the map’s own reading (the pattern and process transfer; the categories do not), stated in his words. Tension 1 could be retagged [partly his]. - Bearing on Andrew’s note 1 (quantitative foundations built on). The essay’s opening is an exposure-measurement scene. The lesson drawn from it is not that measurement failed. It is that the measurement question opened onto institutional and social questions that measurement alone could not settle. That is extension, not abandonment. But one line reads as absolute if taken alone: risk reframed “from something to be minimized to something to be navigated”. NANO’s wording, “not just about technical hazards to be minimized”, is the more accurate statement of his position. - Bearing on note 2 (humility). “Different in a way that may require new frameworks to even describe”, and “That’s a philosophical claim, and I’m still working through its implications” (§When AI Enters…). He frames the problem as a lack of concepts adequate to describe the thing, not a lack of numbers. - Other. - A conditional acceptance of acceleration claims: Amodei’s “century’s worth of technological change compressed into a decade. If that’s even roughly right…” (§The Education Problem). - Safety reframed: “not just about preventing AI from deceiving us, but about recalibrating our own cognitive responses” to a technology that triggers unearned trust.

2. NANO: “What Nanotechnology Taught Me About Governing AI”#

Provenance. Sole byline. Large parts recompose 2023-10-02 and 2023-05-15 (Schmidt, “it’s complicated, leave it to us”). The Constitutional AI section summarises the Claude-written Constituting Responsibility, so its ideas are [AI-origin; endorsed]. The “eighteen orphan risks” are ChatGPT’s role-play output from 2023-11-21.

The argument. - Nanotechnology governance was “relatively successful”: the 21st Century Nanotechnology R&D Act, the NNI, societal-implications centres with non-technical expertise, multistakeholder partnerships and consensus standards. He was “at the heart of much of this” as NEHI co-chair and PEN Chief Science Advisor. - “The key lesson wasn’t about any specific regulation. It was about process”: everyone affected “has a right to play some role”. Nanotech learned this from the GMO debacle. - AI is repeating the exclusionary pattern at higher stakes. A PCAST member once told him the public cannot decide on technologies they don’t understand; “‘It’s complicated’ is not an excuse for avoiding engagement.” - On the pacing problem (credited to Gary Marchant) he recounts his 2008 institutional proposals: the Global Institute on Emerging Technology Policy, and CETI with Tim Harper. He also cites his membership of the WEF Global Future Council on Agile Governance. - Risk innovation: “risk is not just about technical hazards to be minimized. Risk is about threats to things we value.” “Conventional risk assessment focuses on quantifiable harms”, but “the most consequential risks from AI may be to things that are hard to quantify”. These are “orphan risks”: “emerging threats that no existing institution owns, no established framework adequately addresses”. - Constitutional AI “genuinely surprised me”. Responsible innovation keeps responsibility “bolted on rather than built in”. Constitutional AI builds it in, but its “principle selection is ad hoc and lacks the legitimacy that inclusive governance processes provide”. Responsible innovation, in turn, has “no mechanism for treating the innovation itself as a stakeholder when it has morally relevant interests”. The best governance “may look nothing like traditional regulation”, but “we’re a long way from knowing what that looks like in practice”. - He closes: “I don’t have a governance solution for AI. I’m not sure anyone does.” And: “the early days of an advanced technology transition set the trajectory for decades”, which is “the one I’ve been making since my first Congressional testimony”.

Key concepts. Process over regulation; everyone a stakeholder; the pacing problem; early-warning institutions; soft law, anticipatory and multistakeholder governance (“These aren’t perfect solutions”); risk as a threat to value; orphan risks as unowned and hard to quantify; “no single silver bullet”; bolted-on against built-in responsibility; legitimacy of AI principle selection; the innovation as stakeholder; early trajectory-setting.

What this adds to the map. - Confirms. - C5 and C12, with the nano governance genealogy (T2, T5). - Nanotechnology as a qualified success: “relatively successful” here; “Plenty of mistakes” in 2023-10-02. - The GMO failure case. - “No silver bullets” (2023-04-04). - Correction to tension 11 and to T5/T9. The map says he did not ask of Anthropic’s constitution who should set an AI’s values (2026-01-22). By April 2026 he does, in his own essay: principle selection “lacks the legitimacy that inclusive governance processes provide”. The argument comes from the Claude-written paper, but he restates it under his own name. Tension 11 should record that the question is now asked (from March/April 2026), with the provenance caveat. - Correction to tension 14. The map says the moral-status concern is one “which he raised in 2023–24 and has not revisited”. Here it returns in governance form: responsible innovation cannot treat “the innovation itself as a stakeholder when it has morally relevant interests”. Again [AI-origin; endorsed]. - Deepens the decline of responsible innovation (§7, “Confidence in remedies”). He gives a structural reason for it: responsibility “bolted on rather than built in”. He also proposes a direction, hybrid governance that joins inclusive process legitimacy with built-in (constitutive) responsibility. It is new to the map, tentative, and partly AI-originated. - Orphan risks: status and dating. - The map treats orphan risks as a tool that is absent from his 2024–26 cognition writing, and dates the “known but unowned” definition to the Fable-assisted July 2026 paper [mixed]. - In this April 2026 essay, in his own prose, orphan risks is the named home of AI’s most consequential risks. They are threats to “dignity, belonging, identity, autonomy, democratic participation, what it means to be human”, which “no existing institution owns”. So the ownership sense predates the Fable paper by three months and is securely his. - It is still a tool inside risk innovation, but in his self-presentation it is the label he chooses for the hard-to-quantify human-side risks of AI. - Caveat: the page could have been revised after July. Nothing in it refers to later work. - Early origin: trajectory-setting and lock-in. He says the “early days … set the trajectory” point dates from his first testimony (2006, which belongs to another share). With CONV23 (2023: “this window is closing fast”) and BH (2026: “maybe five years, maybe less”), this is a thread the map under-weights. The map currently has only 2024-05-05 (naive adoption and lock-in) and the co-written 2025-10-19 foreword. See §6 below. - Early origin: institutional proposals (2008). The map’s phase 0 lists early-warning institutions “known from later accounts”. This is one of those accounts. It confirms 2008–2010 and his authorship (CETI co-developed with Harper). - Bearing on Andrew’s note 1. This is the clearest statement in the share. Conventional risk assessment is named and kept (“will this material cause cancer, will this system produce biased outputs”), and risk innovation is “not just” that. The quantitative layer is the base, and the value frame extends it. - Bearing on note 2. Orphan risks are defined by being hard to quantify and outside every framework. This is the mechanism behind his humility: quantification decides what gets attention, and the most consequential AI risks fall outside it. The same essay describes nanotech as a technology that “resisted conventional risk assessment” (the phrase is in 30Y, §Risk Innovation…). So his humility about methods has nano-era roots and is not AI-specific. He closes with “I don’t have a governance solution for AI”. The map’s §8 “Missing methods” entry (“little quantitative treatment of AI risks”) should be reframed as a deliberate position, not a gap.

3. HNS: “Honest Non-Signals, Constitutive Resonance, and the Frameworks We Need for Understanding Human-AI Interaction”#

Provenance. Sole byline; a plain-language account of his 2026 preprints. He states that he developed the arXiv paper “working with Claude to extend and rigorously develop the ideas, a concept emerged”. This matches the map’s [mixed] note on “honest non-signals”. The essay’s own framing and hedges are his.

The argument. - The Trojan-horse question began as a deliberately “a little playful” provocation at a Berlin conference in late 2025. - Epistemic vigilance (Sperber) is “not quite a faculty, more like an immune system”. It works on human signals anchored in real costs. - LLM fluency, helpfulness and apparent disinterest are genuine, but “computationally trivial”. They are honest non-signals, and they may slip past vigilance “not because they’re deceptive, but because they look like the things our cognitive immune system has evolved to trust”. - Four mechanisms: - processing fluency, which is “a design feature that happens to exploit a cognitive bias”; - trust-competence presentation with “no skin in the game”; - cognitive offloading; - the intelligent user trap (Kahan), “somewhat speculative, although there is evidence to support it”. - On “But I know I’m talking to a machine”: “My sense is that this matters less than we’d like to believe.” - Constitutive resonance: conversational AI is the first technology whose “response frequency is matched to the frequency of human self-constitution” (Ricoeur’s narrative identity; Stiegler’s constitutive technics). It works “Not faster (which would be disorienting), not slower (which would be dismissible), but in something like resonance”. It is positioned against fourteen frameworks and defined by “temporal matching, linguistic mediation, genuine bidirectionality, and transformation-through-interaction”. - The harness metaphor presupposes three things: a clean split between controller and controlled; capability extracted without transformation; and instrumentality (Rees’s “nostalgia for human exceptionalism”; “There’s nothing wrong with wanting that”). He asks for “some intentionality around the framing before it locks in”. - Two thought experiments held “more tentatively”: stochastic agency (Character.AI; “not convinced that guardrails alone can address something that is most likely an emergent property”) and the amanuensis inversion (“even if there’s only a small chance…”). - “These are explorations, not findings.” The common concern is that “we may be systematically underestimating what conversational AI does to the human side of the interaction”. Research “hasn’t been done yet”, but the questions are worth asking “before the answers arrive in the form of consequences we didn’t anticipate.”

Key concepts. Epistemic vigilance as a cognitive immune system; honest non-signals [mixed]; processing fluency as a design feature; no skin in the game; cognitive offloading; intelligent user trap; knowing it’s a machine is not protection; constitutive resonance (frequency matching); the harness critique; stochastic agency; the amanuensis inversion; explorations, not findings.

What this adds to the map. - Earlier origin. The cognitive Trojan horse began as a keynote question at OEB Global, Berlin, in late 2025. The map dates the thesis to 2026-01-10; the 2026-01-17 post names OEB 2025, but the map does not record it. Add “late 2025 (OEB keynote)” to C14’s stages and §5.8. - Deepens constitutive resonance. The map knows it only through a secondary description (2026-05-21: “two-way coupling”). His own summary supplies the mechanism: tempo or frequency matching in the medium of self-constitution, with the four conjoined features. This explains, in his terms, why AI “feels” different, and why “it’s just a machine” does not dislodge the felt relationship. - Bears on tension 6 (relationship against “working with a machine”). He says knowing it is a machine “matters less than we’d like to believe”, and that wanting humans in the driver’s seat is legitimate while the harness metaphor may misdescribe reality. My reading: his 2026-05-10 rules (“Do not treat AI as your friend”) are necessary but, by his own account, not sufficient. The relationship is formative whether or not the user knows it is a machine. He comes close to stating the reconciliation the map offers only as “my reading”: human control is the aim, and transformation is what actually happens. - Deepens the harness critique (2026-02-22). The three presuppositions, and the lock-in worry about language (“before it locks in”), tie his metaphor critique to the early-window thread. - Tension 2 (plausibility against tails). “Even if there’s only a small chance…” and his hedges (“somewhat speculative, although there is evidence”) confirm the map’s reading. He accepts low-probability concerns when there is a plausible mechanism and some indirect evidence, and labels them as exploratory. - Bearing on Andrew’s note 2. This is the strongest evidence for humility guiding the work while he still insists on grappling with the issue: - “the evidence base is thin in places, the philosophical work is speculative, and some of this could be completely wrong”; - “When I searched SCOPUS for papers on epistemic vigilance and AI, I found seven”; - “there’s also the possibility that we have all of the cognitive abilities we need to use AI wisely and effectively”; - “These are explorations, not findings”.

Beside these stands the case for acting before the evidence arrives: ask the questions now, “before the answers arrive in the form of consequences we didn’t anticipate”. This is his own statement of the anticipatory stance Andrew describes. It also shows the map’s “uneven evidence base” gap is one he names himself.

4. FWB: “The Future We’re Building, Whether We Mean To or Not”#

Provenance. Sole byline. It includes the ChatGPT-origin Antiqua et Nova comparison and the definition of moral imagination [AI-origin], and it misattributes “Seemingly Conscious AI” (Suleyman’s term). The Future Rising summary is his own account of his own book, and that book is otherwise known to the map only through his posts.

The argument. - Future Rising’s core: “humans are extraordinary architects of the future but dangerously underprepared for the responsibility that comes with that”. The same abilities let us “rob others of the futures they aspire to”. The “gap between our ability to transform the future and our wisdom in doing so is widening”, so “our capacity for future-building has outpaced our capacity for future-building responsibly”. - The book was framed around stewardship (“stewards of the future, caring for something on behalf of generations that haven’t arrived yet”) and around orphan risks: threats to “dignity, belonging, identity, autonomy, what it means to be human” that “get overlooked because they’re hard to quantify and don’t fit conventional risk categories”. With AI “the orphan risks became suddenly concrete”. - Convergence: the three base codes and cross-coding; “the governance challenges that matter most are the ones that fall between disciplines, between institutions”. “The convergence is the thing, and AI is one — admittedly very powerful — thread within it.” This is “where I part company with a lot of AI discourse”. - FFTF revisited: - Ex Machina’s manipulation is “far more worrisome than superintelligence”, “a claim I stand behind more firmly now than when I wrote it”. - On Transcendence: “My position then, as now, was that exponential growth never lasts, that extrapolation massively amplifies uncertainties”; “Make-believe treated as reality has consequences.” - Jurassic Park’s could/should is “the framework that proved most durable”. - “I believe we have an obligation to explore new technologies responsibly.” - Rebuilding FFTF as Spoiler Alert, post-2018 developments “mapped onto the existing ethical frameworks with uncomfortable precision. The technology had changed dramatically. The human questions hadn’t changed at all.” - Extrinsic against intrinsic: conversational AI enters “the processes through which we constitute a sense of self” and alters “our intrinsic ‘base code’”. “This doesn’t require superintelligence. It doesn’t require consciousness.” By that standard, “we’re probably already there”. - The consciousness trap: treating consciousness as the bright line “is dangerously wrong”. Seth is “compelling, though I hold it provisionally”. “Waiting for consciousness as the trigger … means waiting too long.” - Between the camps: optimists miss that “our irrationality, our attachments, our stubbornness are not bugs to be fixed”; doomers miss that these technologies “can and do improve lives”. The table should include “theologians, artists, anthropologists, humanists, and the communities whose futures are being shaped”. “The alignment problem deserves the attention it’s getting”, but the deeper challenge is “figuring out what kind of future we actually want”.

Key concepts. Architects of the future; the gap between power and wisdom; stewardship and future generations; stealing futures; orphan risks (as hard to quantify); convergence and cross-coding; could/should as most durable; exponential growth never lasts; extrinsic and intrinsic; intrinsic base code; the consciousness trap; Seemingly Conscious AI (Suleyman’s term); moral imagination [AI-origin]; the fix frame; a wider table.

What this adds to the map. - Correction (significant): convergence. The map rates convergence “core to 2021, then background” (§5.4) and traces AI “from converging strand to category of its own” (§5.6, Core; §7 change 1). In April 2026 he says, in his own self-account, that “the convergence is the thing” and that AI is “one … thread within it”. CONV23 already said so in 2023: “Artificial intelligence is only one of many transformative emerging technologies.” So he holds two views together: - AI is different in kind in what it does to the self (intrinsic, resonant); - AI is one thread of a converging system in how it changes the world and must be governed.

The map should keep convergence as a live, core systems premise into 2026, and restate “category of its own” as specific to the human-side (self-constitution) dimension. - Qualification: the “reversal” on exponentials. The map lists, as an inferred change of mind (2018→2025), a move from “Exponential extrapolation as a fallacy” to “Exponential blindness as a danger” (§7; §5.5 “with a reversal”). He now reaffirms the 2018 position “then, as now”. S3 and 2024-12-13 show the integrating model, the S-curve: near-exponential phases are real and humans misjudge them, but “exponential growth never lasts”. My reading: this is one S-curve view with two edges, not a reversal. Blindness to steep phases and naive extrapolation beyond them are both errors. - New or under-weighted: stewardship and intergenerational responsibility. It is the named frame of Future Rising (chapter 59 is “Stewardship” in the 2024-09-08 compilation). It appears nowhere in the map’s commitments. It is the temporal side of C11 (justice: whose futures) and of C1 (flourishing). Add it as a recurring concept, with Future Rising (2020) as its origin. - New or under-weighted: the power–wisdom gap as Future Rising’s thesis. See the “gap” point under 30Y. - Upgrades. - Could/should goes from “Recurring (a touchstone)” to his own nomination as FFTF’s “most durable” framework. - Manipulation over superintelligence is reaffirmed “more firmly now” (C14). - The obligation to innovate is restated (C6). - Tension 1 (analogy). “The technology had changed dramatically. The human questions hadn’t changed at all.” This is his own rule for what transfers, and it supports retagging tension 1 as [partly his]. - Earlier origin: consciousness. The consciousness-trap framing restates 2023-08-23 and 2024-06-30 in stronger terms (“dangerously wrong”). - Provenance flags for the map. - “Moral imagination” now appears in his own-voice essays (FWB, 30Y), but its first source is ChatGPT text he published (2025-01-30). It also appears in the co-written AI and the Art of Being Human (“Transcendent … moral imagination”, 2025-10-23). If the map adds it, it should carry [AI-origin; adopted]. - “Seemingly Conscious AI” is Suleyman’s term. - Qualification to the map’s picture of him against the AI-safety mainstream (T9). “The technical challenges are real and important. The alignment problem deserves the attention it’s getting.” He ranks technical alignment below the question of what future we want, but he does not dismiss it. - Bearing on Andrew’s note 1. The same sentence also shows he builds on the technical rather than replacing it. The essay’s apparently absolute lines (“fundamentally a human question”) should be read as claims about priority, not exclusion. - Bearing on Andrew’s note 2. - Orphan risks are defined by being “hard to quantify”. - His line that “extrapolation massively amplifies uncertainties” is a critique of numerical hubris, aimed here at singularity forecasting. The same logic applies to confident risk numbers. - “These aren’t questions I’ve answered. I’m not sure they’re answerable in any final sense.” - Evidence in the other direction: he also speaks with some confidence about his frameworks. The “conceptual architecture I’d built … proved to be exactly what I needed”; the old frameworks fit “with uncomfortable precision”. His humility is about knowing AI’s effects, not about the value of frameworks for thinking.

5. S3: “The Three S-Curves: What AI Is Actually Doing in Higher Education”#

Provenance. Sole byline. It builds directly on 2024-12-13 are-educators-falling-behind-the-ai-curve (capability, utilization and perception curves), which is in the map as “three S-curves (2024-12-13)”. The Sorcerer’s Apprentice reading repeats 2026-04-11 ten-questions-about-ai-and-higher.

The argument. - In fall 2023 his students gave “a collective shrug” to ChatGPT (“Only about a third of them used it regularly”). Now “They’ve become more strategic”, with peer networks and code-switching between courses. - Three S-curves: - capability, which in 2024 looked to be flattening but has been bent “sharply upward again” by agents (“We’re not on a plateau”); - utilization, still accelerating; - educator perception: “Many educators’ mental models are three years and a lifetime out of date.” - “The dangerous gap isn’t between capability and utilization. It’s between utilization and perception.” Integrity and assessment policy is being set on what ChatGPT could do in early 2023. - Playgrounds, not playpens. The dissertation experiment was “frighteningly close to human-written work”. AI replacing doctoral research is “not yet at least, although even this is a line that may be crossed soon”, so “the purpose of the PhD itself needs rethinking”. - “AI doesn’t flatten learning values. It reveals and amplifies them.” Education is “about human formation. And when you understand education that way, AI becomes a tool for formation rather than a threat to it.” - The Sorcerer’s Apprentice is about “capabilities that outstrip your wisdom, and no one has taught you how to think about the difference”. The questions are “about formation”: competency, success, the “illusion of understanding”, and “what do we owe our students”. - “Ignoring AI’s power may be just as naive as wielding it without understanding … the principled position may not be resistance. It may be responsibility.”

Key concepts. Three S-curves; the perception gap; stale mental models; capability re-steepening; playgrounds; the purpose of the PhD; AI amplifies learning values; education as human formation; the illusion of understanding; what we owe students; responsibility rather than resistance.

What this adds to the map. - Correction: formation. The map says formation “as an explicit word for this, appears only in the September 2026 lecture [mixed]” (§5.8; T4). Here, in April 2026 and in his own essay, “human formation” is explicit, as the purpose of education, and in a positive valence: AI as “a tool for formation”. Move the first explicit statement to 2026-04-12. Note that his formation concept has two faces: education as formation that AI can serve, and AI as an unbidden participant in formation (the lecture; “train us to think like them”). - Bears on tension 8 (catalyst against surrender). “AI doesn’t flatten learning values. It reveals and amplifies them.” His own hinge is the learner’s purpose and values, which fits the map’s suggested hinge (“at least if they understand what they are doing”). It widens the hinge from understanding to motivation. His Intelligent User Trap (HNS) still cuts against any exemption. - Earlier origin: the illusion of understanding. “Illusion of understanding” from fluent outputs (April 2026, and 2026-04-11) precedes “the illusion of learning” (2026-05-10) and ties it to the cognitive Trojan horse. - Deepens risk perception (T1). The perception gap is a risk mechanism: institutions deciding on out-of-date mental models. It fits his long-standing view that perception is part of risk, now pointed at experts and institutions rather than publics. My reading: this is the institutional counterpart of exponential blindness. - Capability trajectory. He reports that he misread the plateau: capability looked to be flattening in 2024, and agents re-steepened it. This is a signalled, if mild, update. It fits his S-curve model rather than overturning it. - Symmetric risk in education. “Responsibility” rather than “resistance” confirms C6 and C17. - Bearing on note 2. “I don’t have good answers to most of these.” Humility sits beside hands-on experiment as his way of grappling. The dissertation and degree-plan tests are his empirical method: small, direct experiments rather than quantified risk estimates.

6. BH: “What Does It Mean to Be Human in an Age of AI?” (framing page, Feb 2026)#

Provenance. Sole byline; a standing site page. The “four qualities” come from the co-written book (Abbott and Maynard, AI-assisted), so they are weaker evidence. The framing prose around them is his.

The argument. - The question he hears everywhere is “What makes me me when AI can do what I do?” - “AI is different. Not because it’s more powerful than those technologies (though I think it may be), but because it’s the first technology that doesn’t just change what we can do. It changes our understanding of what we are.” It “holds up a mirror”. - “The rise of AI is fundamentally a human question”. - The four qualities (co-written): Curiosity, Intentionality, Clarity and Care (“choosing human flourishing over pure optimization”). - “This work is not about resisting AI”, but “excitement without reflection is how we sleepwalk into futures we didn’t choose”. There is “a narrow window—maybe five years, maybe less—to shape the relationship between humanity and artificial intelligence. After that, the infrastructure hardens, the habits calcify, and the choices we failed to make become the defaults we’re stuck with.”

Key concepts. What makes me me; AI changes what we are; AI as mirror; the four qualities (co-written); sleepwalking into unchosen futures; a narrow window and lock-in.

What this adds to the map. - Upgrade in provenance: the closing window. The map records the window and lock-in claim only in the co-written book foreword (2025-10-19, “weaker evidence”). Here he states it under his sole byline, with a timescale, “maybe five years, maybe less”, and a mechanism: infrastructure hardens, habits calcify, defaults stick. This makes it his own position. - Bears on tension 5 (inevitability against steering). Three months before “the boat has already left the harbor” (2026-05-21), he says there is a narrow window to shape the relationship. My reading: the two are consistent. The trajectory cannot be stopped, but its shape is still open for a short, closing period. The map notes that he “does not say what evidence would overturn the assumption”. The window suggests the question is less about overturning inevitability than about acting before defaults set. This resolves the tension better than the map’s current reading (“inevitability applies to the overall trajectory, and choice to particular designs and uses”). It adds time as the variable. - Earlier date for “who we are” (C16). The map’s key quote on AI changing “who we are” is from 2026-05-21. This February page states the same claim earlier and more sharply (“changes our understanding of what we are”). - Confirms C1 and C16, the refusal of the binary (“Not because AI is dangerous (though it can be), and not because it’s miraculous”), and his adopter stance.

7. TL: “Teaching and Learning in an Age of AI” (resource page, Feb 2026)#

Provenance. Sole byline, but the page mostly presents resources from the co-written, AI-assisted book (the Instructor’s Guide, AI Companion, tools and characters). Weak evidence of his thinking beyond what he chooses to foreground.

The argument. Students’ AI questions are “about identity, meaning, and what education is actually for”. The page recommends tools for classes: - the Mirror Test; - the Identity Matrix; - the Human Qualities Spectrum (Replicable, Relational, Transcendent); - the Stress-Test Table: “What value is at stake? What’s the reward for compromising? What’s the cost of integrity? What’s the long-term payoff of fidelity?”; - the 7-Minute Clarity Pause.

It highlights characters such as David’s office-hours bot “built to make itself obsolete”: “Success means students need it less”. It also names organisational tools: the Intent Map, the CARE Loop and the Model Dignity Check.

Key concepts. Education for identity and meaning; value at stake as a classroom question; the Human Qualities Spectrum; AI designed to be needed less; productive struggle; dignity checks.

What this adds to the map. Little that is new, and only weak evidence. - My reading: the Stress-Test Table carries his threat-to-value frame (“What value is at stake?”) into the co-written book as a personal-ethics tool. It is an example of risk innovation being domesticated for individuals. - “Success means students need it less” is a design principle, AI built to reduce dependence, that fits his frugal pedagogy (C17) and his worry about dependency. It is book content, so cite it only as corroboration.

8. BOOKS: “What Happens to Books When AI Becomes the Reader?”#

Provenance. Sole byline. It is his own account of making the co-written book, the AI Companion and Spoiler Alert (built with Claude Code). It is useful for provenance rules as well as content.

The argument. - AI and the Art of Being Human was written “incredibly closely” with Claude over three months. He worried about credibility, but “Every word, sentence, paragraph, and chapter in the final book has our human stamp on it”. Yet “the collaboration was genuine, and it went both ways”, producing “combinations of ideas and framings that emerged from the interaction rather than from either party”. - Claude was good at “synthesis, generating novel connections, drafting at scale” and poor at their “voices”, their “humor” and “the specific texture of lived experience”. They left in “Claude’s inexplicable preference for characters named Chen”. - The book was released as a free Markdown AI Companion because they would be “hypocrites” not to meet readers “inside a conversation with an AI”. - FFTF was rebuilt as Spoiler Alert (127 Markdown files). The original is “less than ten percent” of the files. He “maintained control over feel, functionality, purpose, content selection, and voice”. Platform testing gave mixed results, and “the infrastructure isn’t ready”. - Conclusions: - “the format and accessibility of ideas matters as much as the ideas themselves”; - “writing for AI isn’t optional”; - the case for giving work away is “backed by zero hard data and considerable optimism”; - “the interaction produced things that neither of us would have generated alone”.

Key concepts. Co-writing as a “living laboratory”; the human stamp; ideas that emerged from the interaction; writing for AI readers; accessibility as the condition of influence; infrastructure not ready.

What this adds to the map. - Provenance rule, confirmed and sharpened. In his own words, AI and the Art of Being Human contains ideas that “emerged from the interaction rather than from either party”. The map’s “weaker evidence (co-written)” status is right, and this is the reason to keep it. - Tension 13 (instrument and object), with self-evidence. He describes his practice in constitutive-resonance terms: “the interaction produced things that neither of us would have generated alone”. My reading: his theory of two-way transformation is also a description of his own method, and he presents that as a finding, not a worry. That differs from the map’s picture of him as mainly uneasy about attribution (2026-01-17). - New instance of tension 7 (adopter against critic). He builds AI-mediated access to his books and says “writing for AI isn’t optional”. In HNS, the same day, he warns that absorbing AI’s “fluent summaries” offloads evaluation. My reading: his reconciliation is design. The AI Companion includes instructions and is framed as “an AI playground” for “modeling intentional, reflective AI use”. He does not say so explicitly. - Justice and access (C11, C17). Accessibility as a condition of influence is a practical form of his justice commitment. See STICK.

9. STICK: “Stick Figures, Sci-Fi Movies, and the Obligation to Make AI Accessible”#

Provenance. Sole byline. It restates his communication philosophy with career facts: Risk Bites from 2012, a 2020 Frontiers in Communication paper, the 2018 AI-risks video and its 2023 update, and the origin of the Substack.

The argument. - “The privilege of academic scholarship and research comes with an obligation” to make knowledge accessible. - AI’s questions are “too important and too consequential to leave to the people building the technology”. This echoes FFTF p.288: not “solely to scientists, innovators, and politicians”. - The 2018 ten-risk video deliberately avoided existential scenarios in favour of risks “far more mundane but no less serious”. It held up because “the risks I focused on were human risks — about agency, justice, autonomy, and control — rather than technical capability benchmarks that shift every few months”. So “the technology-specific details change constantly, but the human questions underneath them are remarkably stable”. AI manipulation in 2018 and 2025 is “the same thread” that led to the Trojan-horse question. - Films are “democratic entry points”. A WEF participant once suggested that art bridges governance divides better than politics, regulation or education. The Ex Machina chapter is “the most durable piece of AI communication I’ve done”. - The Substack began in early 2023 because he had nothing “neatly citable” on AI risk. It focuses on each person rather than “the rather general handwaving around ‘humanity’”. - Honest broker: “not judging, not advocating for a specific course of action”. “With AI, that’s become harder. The questions are more urgent, the stakes are higher, and the temptation to advocate for particular positions is stronger.”

Key concepts. Scholarship’s obligation of access; the stable human questions; human risks over capability benchmarks; mundane but serious; films as democratic entry points; art as a bridge; honest broker under strain; the person, not “humanity”.

What this adds to the map. - Correction: tension 12 (honest broker and advocate) should be retagged [partly his]. The map tags it [my reading]. Here he names the strain himself: the “temptation to advocate … is stronger”. He resolves it toward frameworks and evidence: “not tell people what to think about AI but give them the frameworks, the evidence, and the stories”. - Tension 1, again. “The technology-specific details change constantly, but the human questions underneath them are remarkably stable.” With FWB, this is the most explicit statement anywhere in the share of what transfers from past technologies. It supports [partly his]. - Bearing on Andrew’s note 2. Anchoring on “human risks … rather than technical capability benchmarks that shift every few months” is a stated methodological choice. He avoids tying risk judgements to fast-moving metrics, which would give a false sense of precision about a moving target. This is the nearest the share comes to Andrew’s point about solace in numbers, though it is framed as durability rather than hubris. - Confirms. The map’s “mundane but serious” calibration (2020-11-12), stories as lens (§5.5), and “the future of being human” as centred on the person (§5.9). Art as a governance bridge (the WEF anecdote) is a small addition. - Early origin. The Substack’s founding reason (early 2023: nothing citable on AI risk) explains the uneven record the map describes (§1 Limits). His AI-risk thinking predates his AI-risk writing.

10. CONV23: “Navigating the risks and benefits of AI: Lessons from nanotechnology…” (The Conversation, 2023, with Sean Dudley)#

Provenance. Co-authored with Dudley, so a shared position. It is the companion to their closed Nature Nanotechnology commentary. His own post of the same day (2023-10-02) endorses and extends it.

The argument. - “Twenty years ago, nanotechnology was the artificial intelligence of its time.” It had its own existential scenario (gray goo), job fears, activist opposition and a bombing campaign. - Many nano risks “were less speculative”: health, environment, governance and engagement. Work on them felt like “playing whack-a-mole”. - Nanotech succeeded by engaging stakeholders, “many of whom were not authorities on nanotechnology”, through the NNI, the 2003 Act, multistakeholder partnerships, ISO standards and the OECD. - AI is “much more exclusionary” (White House CEO consultations, Senate hearings). Non-experts “are often fully capable of understanding its implications”. - AI “could be the most transformative technology that’s come along in living memory”. “The early days of an advanced technology transition set the trajectory for how it plays out over the coming decades. And with the recent pace of progress of AI, this window is closing fast.” - “Artificial intelligence is only one of many transformative emerging technologies” (quantum, genetics, neurotechnologies). Failing to learn “risks losing out on the promises they hold and faces the possibility of each causing more harm than good”.

Key concepts. Nanotech as the AI of its day; speculative against tangible risks; whack-a-mole; non-expert expertise about implications; exclusionary AI governance; the closing window; AI as one of many transitions; symmetric stakes.

What this adds to the map. - Earlier origin: the window. “This window is closing fast” (2023) is an early statement of the window and trajectory thread, three years before BH and NANO. It is shared with Dudley. Map §5.4 and C6 should carry it. - Earlier origin: convergence or plurality. “Artificial intelligence is only one of many transformative emerging technologies” (2023) supports the FWB correction: the plural-transition view never lapsed. - Template for catastrophic speculation. The gray-goo-against-real-risks contrast is already in the map (T2); this confirms it for 2023. - Symmetry. Losing “the promises” counts beside harm (C6), confirmed. - Bearing on Andrew’s note 1. The nano story includes the tangible EHS research programme (“health and environmental impacts”) that he co-led. The governance lessons were built on top of that technical risk work, not in place of it.

Cross-cutting: where the share confirms the map, where it under-weights, and where it differs#

Confirms. Risk as a threat to value, and risk innovation as navigation (C2, C3). Symmetry and the obligation to innovate (C6). Plausible against imaginable, and “make-believe treated as reality” (C7). No abdication, and everyone a stakeholder (C5). Manipulation over superintelligence, “more firmly now” (C14). “Who we are” as the stake (C16). Playgrounds and frugal pedagogy (C17). Honest broker, stories as lens, and “mundane but serious”. Refusal of the optimist–pessimist binary. Superintelligence agnosticism, with doubt about exponentials. Seth held “provisionally”.

Under-weighted in the map. 1. The gap as his unifying construct: capability against societal capacity, power against wisdom, utilization against perception, pacing. 2. The early window, trajectory-setting and lock-in. It appears in CONV23 (2023), BH (“maybe five years, maybe less”) and NANO (“set the trajectory for decades”), which claims it dates from his 2006 testimony. It is now his own prose, not only the co-written foreword. 3. Stewardship and future generations, and humans as architects of the future (Future Rising). 4. Convergence as a live systems premise in 2026, not background. 5. Orphan risks as his chosen label, in his own April 2026 prose, for AI’s hard-to-quantify human-side risks. The “unowned” sense predates the Fable paper. 6. Institutional realism and invisible governance groundwork. 7. Constitutive resonance’s actual mechanism: tempo matched to self-constitution. 8. Education as formation, with AI as a possible tool for it.

Differs from or corrects the map. - Formation is explicit in April 2026 (S3), not only in September 2026 [mixed]. - The exponential “reversal” is better read as a single S-curve view (FWB; S3; 2024-12-13). - Tension 11: he now questions the legitimacy of Constitutional AI’s principle selection (NANO; [AI-origin; endorsed]). - Tension 14: AI moral status returns as “the innovation itself as a stakeholder” (NANO; [AI-origin; endorsed]). - Tension 12 should be [partly his] (STICK). - Tension 1 should be [partly his]: “The human questions hadn’t changed at all” (FWB); “the human questions underneath them are remarkably stable” (STICK); “structurally identical” (30Y). - Tension 5 gains a temporal resolution: there is a window before defaults harden (BH). - Tension 6 gains his own partial resolution: knowing it’s a machine is not protection, and human control is a legitimate aim that the harness metaphor may misdescribe (HNS). - He explicitly credits alignment research (“deserves the attention it’s getting”, FWB), which qualifies T9. - The cognitive Trojan horse begins at the OEB Berlin keynote, late 2025.

Provenance cautions for the map. “Moral imagination” is [AI-origin] (ChatGPT, 2025-01-30; also in the co-written book). “Seemingly Conscious AI” is Suleyman’s term. The “eighteen orphan risks” are ChatGPT’s role-play output. The Constitutional AI comparison comes from a Claude-written paper. None of these should become core concepts on the strength of this series.

Cross-cutting: Andrew’s two notes#

Note 1: quantitative risk assessment is a foundation, built on and not abandoned. The share supports this and gives the map better wording. - The career story starts in exposure measurement (30Y). Conventional risk assessment of “quantifiable harms” is named and kept (NANO). The value frame is “not just” technical hazards (NANO). The nano governance lessons sat on top of an EHS research programme he co-led (CONV23; NANO). He calls the technical and alignment challenges “real and important” (FWB). - Against this, some of his own summary lines sound absolute (“the gap is never primarily technical”; risk reframed “from something to be minimized to something to be navigated”). The map should read these as statements of priority, as the surrounding text shows, and prefer his “not just” and “extends” formulations.

Note 2: the lack of quantitative methods for AI reflects humility about the hubris of risk assessment. The share supports this, mostly implicitly. - Orphan risks are defined by being hard to quantify and outside every framework (NANO; FWB). Quantification selects what gets attention, and the most consequential AI risks fall outside it. - He chose “human risks” over “technical capability benchmarks that shift every few months” (STICK). - He criticises the way extrapolation “massively amplifies uncertainties” (FWB). - He hedges openly: “explorations, not findings”; seven SCOPUS papers; “could be completely wrong”; “I don’t have a governance solution for AI. I’m not sure anyone does.” (HNS; NANO) - He insists on grappling anyway: ask the questions “before the answers arrive in the form of consequences we didn’t anticipate” (HNS); “the principled position may not be resistance. It may be responsibility” (S3). - None of these texts states the argument about the hubris of numbers in so many words. That formulation is Andrew’s (his note) and not yet in his published prose. His humility is about knowing AI’s effects. He is confident about the value of frameworks for thinking. - The map’s §8 “Missing methods” entry should be recast as a principled stance, with tension 3 kept: he still states no evidentiary bar for non-quantified harms.

Digest#

These essays are Andrew’s own curated self-account. On 12 April 2026 he published a hub essay and six companion essays on andrewmaynard.net, and two framing pages in February. They mostly confirm the map’s core: risk as a threat to value; symmetric risk and the obligation to innovate; plausibility against make-believe; nobody deciding alone; manipulation over superintelligence (a view he holds “more firmly now”); and “who we are” as what AI puts at stake. They also shift its weighting in eight places.

1. The gap is his organising idea. He tells his career as one gap between what a technology can do and a society’s capacity to understand and govern it. He met it first in 2004, measuring nanotube aerosols, where the complications “had less to do with aerosol physics than with how institutions … handle technologies they don’t yet understand”. AI poses “a structurally identical question”. The pacing problem, power outrunning wisdom (the thesis of Future Rising), and utilization outrunning perception (the three S-curves) are all versions of it. The map never unifies these pieces.

2. Time matters: early windows and lock-in. Under his sole byline he gives “a narrow window—maybe five years, maybe less” before “the infrastructure hardens, the habits calcify”. He says the “early days … set the trajectory for decades” point goes back to his first Congressional testimony in 2006; the 2023 Conversation piece says the window “is closing fast”. The map has this only as weak, co-written evidence. It also resolves tension 5 better than the map does: the trajectory cannot be stopped, but its shape stays open for a short and closing time.

3. Convergence is still central. “The convergence is the thing, and AI is one … thread within it” is where he “part[s] company with a lot of AI discourse”. The map’s arc, from converging strand to “category of its own”, holds only for AI’s effect on the self. It does not hold for his systems view.

4. Orphan risks and quantification. In his own April prose, orphan risks are threats to dignity, identity and autonomy “that get overlooked because they’re hard to quantify”, and that “no existing institution owns”. So the “unowned” sense predates the Fable-assisted July paper. For Andrew’s notes: conventional assessment of “quantifiable harms” is kept as the base, and risk is “not just” technical hazards. But quantification decides what gets attention, and AI’s most consequential risks fall outside it. Nanotech had already “resisted conventional risk assessment”, so the humility has nano-era roots. He anchored his AI-risk work on stable “human risks” rather than “technical capability benchmarks that shift every few months”. His hedges are explicit: “explorations, not findings”; seven SCOPUS papers; “I don’t have a governance solution for AI. I’m not sure anyone does.” So is his case for acting anyway: ask the questions “before the answers arrive in the form of consequences we didn’t anticipate”. None of the texts states the hubris-of-numbers argument outright; that wording is Andrew’s own.

5. Formation is earlier and has two faces. In April 2026 education is “about human formation”, and AI can be “a tool for formation rather than a threat to it”. “AI doesn’t flatten learning values. It reveals and amplifies them.” The map dates explicit formation to the September lecture [mixed] and gives it only its darker side.

6. Several of the map’s inferred tensions are ones he names himself. - On analogy: “The technology had changed dramatically. The human questions hadn’t changed at all.” - On the honest broker: “the temptation to advocate … is stronger”. - On relationship: knowing it is a machine “matters less than we’d like to believe”. - On exponentials: the S-curve view reconciles them (“exponential growth never lasts”, yet the curve is “steepening”). It is not a reversal.

7. Anthropic and alignment. He now asks who should choose the values in Constitutional AI (principle selection “lacks the legitimacy that inclusive governance processes provide”). He sketches hybrid governance, responsibility “built in” as well as “bolted on”, and revives AI moral status as “the innovation itself as a stakeholder”. All three come from a Claude-written paper he endorses. He also says “the alignment problem deserves the attention it’s getting”.

8. Two frames the map lacks: stewardship and a wider table. Future Rising frames people as “architects of the future” and stewards “on behalf of generations that haven’t arrived yet”. He wants a wider table: “theologians, artists, anthropologists, humanists”.

Provenance cautions. The series recycles his earlier prose, which is good evidence of what he still holds. But it also presents a ChatGPT-written comparison (and the definition of “moral imagination” taken from it), ChatGPT’s list of “eighteen orphan risks”, and Suleyman’s “Seemingly Conscious AI” as his own. It also misdescribes the co-written book as co-written “with Anthropic’s Claude”. Treat those items as [AI-origin] or misattributed.