Timeline: how Andrew Maynard’s thinking on risk, technology and AI evolved (2014 to September 2026)#

A phase-by-phase account of what preoccupied him, what was new, what persisted and where he changed his mind. It is built from the batch digests and notes (notes/B01–B32), the chapter notes on Films from the Future (book/notes-FFTF-A–F), and the consolidated list of concepts in concept-index.md. For definitions, see the concept index. This file is about sequence and change. It does not compare his work with anything else.

How to read this file#

What counts as evidence. Only his own prose is evidence. Excluded: - AI-generated text; - guest posts; - Modem Futura podcast posts, at the user’s request; - the standing exclusions listed in concept-index.md.

Weaker evidence: - co-written work (with Garbee, Boorsma, Lobo, and Abbott’s AI and the Art of Being Human); - the Claude-drafted text of the 2026-09-24 lecture.

Dates. Before 2019 the dates are approximate: migrated Conversation, Medium and 2020 Science posts carry Substack dates that can differ from first publication. Republished material is dated by when it was written where that matters. Examples: - the 2018 Ex Machina chapter, reposted 2023-04-16 and 2025-03-02; - the Minority Report chapter, reposted 2020-01-01 and 2020-09-26; - the orphan-risks article, reposted 2023-11-15.

Reposting old text unchanged is itself evidence that he still held the view.

Volume. Output is uneven. There are 14 high- or medium-relevance posts for 2014–17, 23 for 2018, 36 for 2019–22, and then 61, 65, 51 and 30 for 2023–2026 (the last to 24 September only). Read the thread counts at the end with that in mind.

Phases at a glance#

Phase 0. Before the record (to 2014): the making of a risk scientist#

This phase is known only from his later accounts. - Workplace aerosols. Thirteen years of workplace aerosol and dust research at HSE and NIOSH, shaped by his grandfather’s death from black lung. It gave him lasting convictions: technological sophistication does not produce safety; harm lands first on “the first tier of people”; scientific uncertainty can suit those who profit (FFTF p.118–122). - Nanotechnology. He was the first co-chair of the NNI’s environment and health committee and chief science adviser to the Project on Emerging Nanotechnologies. This is the source of his insider view of “brand nano” (FFTF p.214–218) and of the risks that “slip under the radar” once attention moves on (2016-02-01). - AI. He met Bostrom in 2008 through nanotech circles, when “AI wasn’t even on my radar” (FFTF p.168). - Governance proposals. WEF notes (2008) and a 2010 Davos proposal with Tim Harper for a Global Centre for Emerging Technology Intelligence (2023-12-22). - Civil society and geoengineering. The 2020 Science civil-society series (2009, FFTF p.191). Geoengineering ethics from 2009 (FFTF p.266). - Materials standards. The “sophisticated materials” work and his case against definitional regulation of nanomaterials (2010–11, restated in 2022-02-10). - Teaching and risk innovation. An entrepreneurial-ethics course from 2013 (2019-04-15). Risk-innovation work with Michigan students from 2013 (2020-11-05). - AI hardware. A 2014 Nature Nanotechnology piece on 3D-printed artificial minds (re-presented 2023-12-03).

Phase 1. 2014–2017: from nanomaterial risk to risk innovation#

What preoccupied him. - Whether nanomaterials are risky: quantum dots, carbon nanotubes, the two Vantablacks. He argues against hazard-only alarm (“taken in isolation”) and insists risks outlast public attention (2015-01-10; 2016-02-01). - Responsible development in a complex, connected world (2015-01-30). - Two critiques: of the WEF’s “Fourth Industrial Revolution” (2016-01-11 fourth-industrial-revolution), and of Musk as technically brilliant but socially naive (2016-03-12; 2017-04-10). - Neurotechnology acting on thought and moral judgement (2016-03-31). - Scientists’ responsibility for hype, citizen science, and universities’ public duty (2014-12-14; 2016-01-12; 2016-01-31).

What was new. - Risk innovation and risk as a threat to value (2016-01-11 thinking-innovatively). This is the most original move of the period. It comes with a risk “market” of constituents willing to invest in protecting what they value. - Future value (2016-03-02): not developing a technology is itself a risk. Alongside it, the split between tech-fixable risks and subtler social risks, and an early picture of conversational, agentic AI: open AI ecosystems that eavesdrop and “independently decide what’s best for you”. - Systemic fragility. Converging technologies form a “socio-techno-environmental system” that looks stable “until, suddenly, it isn’t” (2015-01-30). - The capability–responsibility gap “has continued to widen”, and Bill Joy’s warning “still haunts me” (2016-01-11 fourth-industrial-revolution). - Who defines risk: “who considers what a risk (and to whom)” (2016-03-31). - Social licence (2017-04-10).

What persisted from here. Nearly everything: threat-to-value, anti-hazard-only reasoning, responsible innovation (Stilgoe et al.), adaptive governance (DOT’s 2016 AV policy as a model, 2016-04-01), convergence, and the stance that “we can’t afford to slam the breaks” [sic] on innovation (2015-01-30).

AI’s place. One member of a converging set. The early named risks are autonomous weapons and machines that don’t respect human values (2015), AI inequality and eavesdropping ecosystems (2016), and neural lace coupled to cloud AI (2016). His register is anti-Hollywood: real failures look like “teens troll supercomputer” (2016-03-02).

Changes of mind. Between 2015 and 2016 he moves from citing responsible innovation as an outside framework to proposing his own risk-innovation reframing.

Phase 2. 2018 to mid-2019: the book era (Films from the Future)#

What preoccupied him. The book (November 2018) was written as a tour of his whole field, so this phase is the best single snapshot of his worldview before generative AI. Around it are tech-company-facing posts: a social-risk reboot (2018-09-03), orphan risks for startups (2018-12-13), an ethics reset (2019-04-15), and algorithms as chemicals (2019-03-05).

What was new, or first set out in full. - Responsible innovation as the organising question. How to “reap the benefits of innovation without running into serious problems along the way” (2018-10-12). Could vs should (FFTF p.36–39). - Myopically benevolent science. A non-demonising model of developers, with self-implication (“even myself at times”, FFTF p.220). “It’s good to talk”: non-engagement is “an extremely high-risk strategy” (FFTF p.226–227). - Permissionless innovation and hubris. The critique, the lure, “isn’t necessarily reckless” (FFTF p.159–168). Rising irreversibility: “playing with fire in a world made of kindling” (FFTF p.167). - Artificial manipulation. His headline AI risk. Plato’s Cave and the “human club”, and the need for tests of “when we are being played by machines” (FFTF p.153–177). It is already present as “heuristic manipulation” in the ten-risk Risk Bites list (2018-05-12). - Superintelligence agnosticism. “something of an agnostic”. Superintelligence is “currently scientifically implausible”, intelligence is “a term of convenience”, and he prefers bounded optimality (FFTF p.168–173; 2018-10-04). - Plausible vs imaginable. Occam’s Razor for futures (“a house-of-cards stack of assumptions” versus nanomaterial harms; 2018-11-01). The critique of exponential extrapolation, and speculative fear feeding real violence (2018-11-15). - Worth and dignity. The “convenient lie” (FFTF Ch. 3). Engines of persuasion and algorithmic bias (FFTF Ch. 4). Intelligence obsession and “Being smart doesn’t make you good” (FFTF Ch. 5). - Power, inequity and labour. Technology amplifies power (Elysium). The disposable workforce. Automation takes away “choice” (FFTF Ch. 6). - Identity under augmentation. Ghost hacking and malleable identity. “who owns you”. The pull-the-plug fallacy. “Here be monsters” maps (FFTF Ch. 7). - Convergence and base code. Bits, bases and atoms, and cross-coding (FFTF p.183–189). Neo-Luddites reclaimed as a demand for justice (FFTF p.191–193). - Immoral logic and the right to act unilaterally. The Honest Broker as his chosen role (FFTF Ch. 11). - Value-based resilience (FFTF Ch. 12). Everyone is a stakeholder, and public indifference is a governance gap (FFTF Ch. 13–14). - Orphan risks are named as “a new” term for startups and investors (2018-12-13). - Algorithmic exposure and exposure–response. His explicit transfer of chemical risk assessment to algorithms, with the caveat “an algorithm is not a chemical” (2019-03-05). - De-marginalizing the future (2019-03-31).

What persisted from Phase 1. Risk innovation (FFTF p.22–23: “where much of my current work lies”), threat-to-value (FFTF p.23–24), convergence, responsible innovation, and “Don’t Panic” with the obligation to innovate (FFTF p.287–290).

AI in this phase. Manipulation is plausible and superintelligence implausible, and he flags a machine-learning and NLP “tipping point” (FFTF p.170). He criticises OpenAI’s founding logic that the answer to innovation is more innovation (2018-12-15). Ethics boards are “smoke-and-mirrors” at worst (2019-04-15).

Changes of mind. A candid reassessment of “brand nano”: the field he helped build was sold with fudged science, though it had real interdisciplinary value (2018-02-21). His superintelligence scepticism is held provisionally: “I freely admit that I may be wrong” (FFTF p.170).

Phase 3. Mid-2019 to 2022: consolidation, neurotechnology, Future Rising, and the approach of generative AI#

What preoccupied him. - The Risk Innovation Nexus in practice. Brain–machine interfaces (2019-07-23; 2019-11-01; 2020-08-28; 2020-10-15), astrobiology (2020-07-30), the Nexus’s history and end (2020-11-05). - Responsible innovation for US entrepreneurs (2019-08-13, co-written). - Prediction and pre-justice. The Minority Report chapter was republished twice in 2020. - Future Rising (2020). Responsibility to the future, and stolen futures (2020-12-10). - Trustworthiness in governance (2020-12-15). - COVID-era materials risk: graphene masks (2021-03-28) and advanced-materials standards (2022-02-10). - AI moves toward the foreground. The AI-risk research gap (2021-08-03), Tesla Bot (2021-09-07), Alexa (2022-02-12), Midjourney (2022-09-16) and ChatGPT (2022-12-08).

What was new. - Structural vulnerabilities. Tight coupling, latency and value mismatch. Responsibility has to be culturally embedded in innovator communities through “creating mutual worth” (2019-08-13). - Orphan risks become a method: a list of 18, landscape mapping, and clusters converging into blindsides (2019-11-01). - Precaution handled carefully. It is “often misunderstood”, yet the COMEST version is “a sound philosophy”. Alongside it: “There’s more to risk than probabilities” (2020-07-30). - The turning point is scaling, not science. A “synergistic scaling of ability, accessibility, and use” outruns the narrow frames of experts. Enhancement risks the “technological equivalent of indentured servitude” (2020-08-28; 2020-10-15). - AI risk is mundane, not apocalyptic. It is “far more mundane” but “no less serious for this”, and AI sits on “a knife edge” (2020-11-12). - Base code extended. First to “bricking” the world, then to a social base code of norms and ideas (2021-02-25). The solution problem, and an inversion of timescales in which consequences outpace fixes (2021-04-09). - The materials risk method, in two moves. Irresponsibility judged by process (2021-03-28). Terms of art versus terms of science: “nature doesn’t care what we call a material” (2022-02-10). - Whose future? Tech visionaries imposing futures (2021-09-07; 2022-06-13). Predictive inference that turns users into dependent super-consumers (2022-02-12). - First reaction to generative AI. Co-creation with Midjourney, with “humans in the loop somewhere” (2022-09-16). ChatGPT as a leap in machine-augmented communication, with risks of attachment, gullibility and disinformation (2022-12-08).

What persisted. Risk innovation, threat-to-value, responsible innovation, the plausibility discipline, the non-prohibitionist stance (“Not that I think this should be taken as an excuse not to build” brain–machine interfaces, 2020-08-28), and the prediction/bias critique.

Changes of mind. - On Neuralink: “we were somewhat naïve” to assume it put medical uses ahead of enhancement, and he is “more convinced than ever” that entrepreneurial BMI work outruns governance (2020-10-15). - His first reaction to generative AI (2022) is strikingly positive. That positivity sits alongside his criticism of AI companies’ market logic.

Phase 4. 2023: after ChatGPT, governance, transitions and the language turn#

Context. He launched the Substack in April 2023 after realising he had written almost nothing “neatly citable” on AI risk (2023-04-04 welcome-to-the-future-of-being-human). The year falls into four stretches: - January–March. Hands-on ChatGPT experiments. - April. A rapid burst after GPT-4 and the FLI pause letter. - May–July. Responses to institutional events: Schmidt, the Senate hearing, Bengio, the CAIS statement, frontier-AI regulation, the EU AI Act, x.AI. - July–December. Education, consciousness, public engagement, nanotech lessons, the NSF comments, Andreessen, the Executive Order, the first-principles risk essay, and orphan risks re-aimed at AI.

What preoccupied him. Who governs AI and on whose expertise. The existential-risk debate. The place of past technology transitions. Language, relationship and mental health. AI in the classroom. Machine consciousness.

What was new. - Advanced technology transitions becomes his umbrella frame (2023-04-12; 2023-09-25). There is a built-in tension: past lessons are ignored and each wave tends to “re-invent the wheel”, yet the present is unprecedented. - The ethics-to-risk reframe. A governance genealogy (rDNA to ELSI to nano and synthetic-biology soft law), with risk innovation offered as the corrective to AI’s return to ethics (2023-04-04 alternatives). - Who gets heard. The “insights vacuum” and loud versus quiet voices (2023-04-10). “leave it to the technical experts” “never plays out well”, and everyone has the right to play some role (2023-05-15). - The language turn. - “Language plays a large part in how we develop these relationships” (2023-04-05). - Self-replicating ideas spread by machines “adroit at manipulating language” (2023-04-26). - LLMs’ “seductive mastery of language” (2023-05-31). - AI as “part of the mix of influences that inform and modulate our thinking” (2023-08-21). - Catastrophe as mass loss of value. He declines the CAIS extinction framing: “too narrow and absolute a framing, and too human-centric”. Conventional risks are “shavings off the tip of the AI iceberg” (2023-05-31). - Harm through apparent authority. “authoritative rather than accurate” (2023-05-22). Doubts about separating technology from use (2023-05-17; 2023-07-12). - Alignment as a question of power. Whose values, and who decides (2023-05-25). Curiosity is not benevolence (2023-07-19). - AI compared with nuclear. AI is “hidden, dispersed, readily accessible” (2023-07-25). - Machine consciousness and moral status. He endorses computational functionalism and names the “economic expediency of denying consciousness” (2023-08-23). - Engagement and nano lessons. “intentional two-way engagement” (2023-09-04). Nano and GMO lessons published as process lessons, not hazard lessons (2023-10-02). - Technological foreshortening (2023-10-19). - Risk from first principles. “no cause, no risk”, risk as “ultimately a social construct”, and the hazard–exposure paradigm tested on AI’s influence on thought (2023-11-26). - Value vs values. The Planner, and orphan risks re-aimed at AI developers (2023-11-15; 2023-11-21). - AI as socio-technical. “suck it and see” versus tipping points (2023-11-29). - Governance lineage and the good-intentions fallacy. The WEF-era lineage behind AI governance and the “AI divide” (2023-12-22). Tech leaders’ good intentions as “sheer fantasy” (2023-12-15).

What persisted. Manipulation over superintelligence: the 2018 chapter was reposted as “more important today” (2023-04-16). The ten-risk list, still “not bad” (2023-04-24). Permissionless-innovation critique and Luddites reclaimed (2023-05-12). Risk innovation, “perhaps more relevant than ever” (2023-08-08). Plausibility. Engagement. Myopic benevolence (2023-09-08).

Changes of mind and shifts in tone. - Existential risk. He now accepts a risk “of potentially existential proportions”, stronger than in 2018, while refusing the pause letter because there are “no silver bullets” (2023-04-04 alternatives). - Regulating use, not technology. He publicly doubts his own field’s nanotech-era consensus on this (2023-05-17; 2023-07-12). - Consciousness. Within a week he moves from “a long way from machines that have consciousness” (2023-08-18) to near-term plausibility (2023-08-23). He does not reconcile this with his republished 2018 hardware scepticism. - Peak educational optimism. ChatGPT is “a profoundly effective catalyst for engaged and creative thinking” (2023-08-14). A course designed with ChatGPT is on a “printing press” scale (2023-07-16). - His own writing. “I typically don’t use ChatGPT myself when I write” (2023-09-20). - A tentative opening to a pause: “a chance to take a breath (a pause even)” (2023-11-18).

Phase 5. 2024: being human, relational AI, and transitions as a field#

What preoccupied him. - Being human. Extrinsic vs intrinsic technologies (2024-01-01). The oxygen analogy (2024-03-31). - Global governance and risk perception. UN and WEF (2023-12-22; 2024-01-14; 2024-01-17). - Tech leaders. Neuralink (2024-01-30; 2024-03-21; 2024-09-18), Altman and Her (2024-05-21), Musk (2024-08-04), Amodei and Altman on climate (2024-10-06; 2024-10-13). - Relational and agentic influence. 2024-05-15; 2024-07-13; 2024-09-01; 2024-09-22; 2024-10-20; 2024-10-27; 2024-11-03. - Safety and existential-risk culture. 2024-04-28; 2024-06-20; 2024-06-23; 2024-06-30. - Transitions as a field. 2024-08-11; 2024-08-18; 2024-08-25. - Education. 2024-03-10; 2024-03-17; 2024-05-05; 2024-12-13. - Embryo screening (2024-04-14) and ChatGPT as author (2024-09-28).

What was new. - Intrinsic technologies. They may change “what we are”, possibly “without our agreement or permission”. Language is part of the base code of identity (2024-01-01). - Technology is part of us. It is constitutive, not “something we do” (2024-03-31). “pilots” are needed for transitions. Neighbouring fields, responsible innovation included, lack breadth. - Hyper-anthropomorphism (2024-05-15). Relational persuasion and benevolent persuasion, with “who decides what is good for society” (2024-09-01). Stories “hard not to trust, and yet are not trustworthy” (2024-09-22). - The economic gradient toward manipulation. A dual-use quadrant in which individuals become “engines of value creation rather than the primary recipients” (2024-07-13). - Agentic social AI, then stochastic agency. Harm can arise without corporate irresponsibility (2024-10-20; 2024-10-27). Hidden AI-mediated influence (2024-11-03). - Safety is social. “There is no such thing as absolute safety”, and the question is who decides what “safe” means (2024-06-20). - Against the x-risk ideology, not x-risk. A physicist’s critique of FHI (2024-04-28). But ignoring catastrophe “is in itself a risky strategy” (2024-06-23). - Conscious-seeming AI. “instinctively incapable” of acting on what we know. The thermodynamic doubt about AGI (2024-06-30). Counterfeit minds (2024-03-03). Relinquishing decisions to machines (2024-07-21). - Two transitions models. Avoid/Adapt/Extend/Embrace (2024-08-18), and the threat/opportunity model built explicitly on threat-to-value (2024-08-25). - The “fix” frame. Fixing people, and who decides “normal”. The deficit model “debunked decades ago” (2024-10-06; 2024-10-13). - Agency and joy. The human amanuensis (2024-11-24). The soul of science and joy as values (2024-11-10; 2024-12-29). - Imagination and humility. Plausibility named as his own technique (2024-11-17). Wariness of historical analogies for AI (2024-05-05). - Risk innovation meets biotech. Value vs values, and internal vs reciprocal threats (2024-12-17).

What persisted. Manipulation (the Ex Machina concern is “even more relevant now”, 2024-06-30). Orphan-risks landscape (“still stands up reasonably well”, 2024-03-21). Inevitability, stated as “manage and channel the inevitability of technological change” (2024-08-18). Plausibility. Who decides. Could vs should.

Changes of mind. - Deepfakes. He once trusted common sense, is now “far less sure”, and signs a letter for criminalisation and developer liability (2024-02-25). - Technology apologetics. He questions his decades-long accommodation of the view that technology is separate from us, and names responsible innovation among fields too narrow (2024-03-31). - Pace. GPT-4o arrived “faster than I could have imagined” (2024-05-15). - Tipping points. He now entertains embracing them, against “my original thinking” (2024-08-18). - Humanoid robots. He moves from laughing at the idea to taking the bet seriously (2024-08-07). - ChatGPT as a creative tool. On its creative capability and its “tool” status: “Now I’m not so sure” (2024-09-28). - Agentic influence. Goal-directed agentic social AI is revised a week later into stochastic agency (2024-10-27). - A narrow, conditional pause. “pausing — or even rethinking” emotion-exploiting companion bots (2024-10-27). This goes further than his 2023 refusal of the pause letter. - Scaling. He doubts it (“I’m not even convinced”, 2024-10-06) yet calls powerful AI “a realistic possibility” (2024-10-13). He weights a plateau (2024-12-13), then takes agentic trajectories seriously after o3 (2025-01-07 universities-need-to-step-up). - Tone toward leaders. Warmer toward Musk (“I get where he’s coming from”, 2024-08-04) and toward Neuralink (2024-03-21), and sharply critical of OpenAI’s conduct (2024-05-21), while enthusiastic about the ASU–OpenAI partnership (2024-01-18).

Phase 6. 2025: AI in scholarship, permissionless politics, care, and everyday risk#

What preoccupied him. - AI and research. Deep Research and the future of scholarship (2025-02-04; 2025-02-09; 2025-02-16). Manus and agents (2025-03-22; 2025-03-27). - The political turn toward permissionless innovation (2025-02-23; 2025-03-02; 2025-07-23). - Care (2025-03-09). - Education and AI’s categorical novelty (2025-03-15; 2025-03-30). - AI acceleration and responsible innovation (2025-04-06). Agent governance (2025-05-04). - The Trajectories Tool (2025-05-18). Parasocial communication (2025-05-25; 2025-11-19). - The manipulation framework (2025-07-06; 2025-07-13; 2025-08-31). - Everyday AI risks (2025-09-07; 2025-10-05; 2025-10-26; 2025-11-09). - AI and the Art of Being Human (co-written, October 2025), and the story Letters from the Department of Intellectual Craft with its Postscript (November 2025).

What was new. - The artisanal intellectual. Coined in a footnote (2025-02-09). Human-only scholarship may look “intellectually limited and somewhat quaint” (2025-02-04). - Hard care as a governance-grade concept. Care against control and against aggregate thinking (2025-03-09). - The reversibility test. Experimentation is fine in reversible, linear systems but not when breaking “people, governance, society, and the planet” (2025-03-02). - AI’s categorical difference. Treating AI as a learning aid is “a categorical error” (2025-03-15). What we do / who we are / where we live, and the value-creation model of education (2025-03-30). - The timescale mismatch and exponential-growth blindness. Responsible innovation may “seem futile” (2025-04-06). - Agentic AI and non-linear effects. A working definition of an AI agent, with risk “closer to exponential than linear” (2025-05-04). Six cause–effect models, with hysteresis and chaos. The companion bot as his paradigm of irreversible harm (2025-05-18). - Manipulation, structured. Motive, means and opportunity (2025-07-06). Emergent versus designed manipulation, universal vulnerability, and channelling the flood (2025-08-31). - Institutions and everyday risk. A duty of care for institutions deploying AI, and why warnings and literacy classes fail (2025-11-09). Relational connective tissue (2025-09-07). Memory as informant (2025-10-05). - Fiction as scenario. Intellectual craft, the Great AI Reset, human-adjacent values and the crisis of abundance (2025-11-24 part-1-letters to 2025-11-30). - Expert crowds underrate AI. Mainstream experts “do not grasp how disruptive the technology may turn out to be” (2025-01-19).

What persisted. Threat-to-value, now merged with care: “effect” is what we value “or what we care for” (2025-05-18). Permissionless-innovation critique (“more relevant now than it was then”, 2025-03-02). Hubris. Plausibility. Inevitability (“guide and steer”; “The AI genie is out of the bottle”). The nanotech governance precedent (2025-07-23). The Jurassic Park chapter reaffirmed unchanged (2025-04-13).

Changes of mind. - Writing with AI. From the 2023 personal rule to publishing a wholly ChatGPT-written analysis: “I’ve resisted doing this for so long” (2025-01-30). By mid-February he finds it “hard to imagine a future of effective intellectual labor that is not AI-enhanced” (2025-02-16). Yet letting AI write as him makes him uneasy (2025-03-16). - Musk. Revised from humble AI-risk voice (2018) to naive permissionless innovator (DOGE) (2025-03-02). - Exponentials. From a fallacy to debunk to a blind spot in human cognition and planning (2025-04-06). - Responsible innovation under acceleration. It would fare “not that well” (2025-04-06). - Prompt engineering. He dropped his course as outdated (2025-08-17). - “Parasocial.” “I read the tea leaves wrong”, while he still stands by the concept (2025-11-19). - AI capability. “stochastic parrot” language (2025-02-23) coexists with “most definitely not the generative AI of 2022” (2025-02-09). By August, GPT-5 is “good—really good” (2025-08-10).

Phase 7. January to September 2026: cognition, formation and the frontier#

What preoccupied him. - Epistemic vigilance and cognition. 2026-01-10; 2026-01-17; 2026-02-08; 2026-05-10. - What AI is. 2026-01-22; 2026-02-22; 2026-04-26. - Formation and identity. 2026-03-08; 2026-05-21; 2026-06-10; 2026-07-19; 2026-09-24. - Higher education and his own university. 2026-03-29; 2026-04-11; 2026-04-14; 2026-05-03; 2026-08-30. - Frontier AI as research partner. 2026-06-12; 2026-07-04; 2026-09-04. - How AI companies select risks. 2026-07-16. - The AI risk landscape and existential risk. 2026-03-22; 2026-09-15. - Play. 2026-08-02. - His sabbatical (from May) and his move to Thunderbird.

What was new. - The cognitive Trojan horse. A second-order evolutionary mismatch in which AI bypasses epistemic vigilance through fluency, attractiveness and volume, with the Intelligent User Trap (2026-01-10). - Honest non-signals. Adopted from Claude. Calibrated trust-cues as AI safety, and “a collective form of epistemic vigilance” (2026-01-17). - AI defies analogy. It is “at once deeply human and deeply alien” (2026-01-22). - Relationship, not command. The harness critique, bidirectionality and constitutive resonance (2026-02-22; 2026-05-21). Relational technology and character constancy (2026-04-26). - LinkedInification (2026-03-08) and reverse formation: “the AIs we have trained to ‘think’ like us are now beginning to train us to think like them” (2026-07-19). - Safety message first, with five don’ts and five dos. AI is “the first technology of it’s kind” [sic] to slip unnoticed into the mind (2026-05-10). - Two-sided hubris, and “near-frictionless access to power that transcends our understanding” (2026-04-11). - “Who we are” as the domain where AI is unprecedented, and AI is not “just a tool” (2026-05-21). - Formation. From outputs to formation in education (2026-06-10). The validation gap and the epistemic precipice (2026-06-12). - The frontier-AI risk-selection paper. Four filters, the safety differential, severity floors, apertures, an orphan-risk register, unequal conversion channels, and sincerity inside an incentive field (2026-07-16). - Play and learning. “play without purpose”, and learning with “no learning expectations” (2026-08-02). - The KCL lecture (2026-09-24): - intelligence scarcity as the university’s hidden business model; - “Language is formative”, with AI “actively taking part in the formation process”; - language as a lever, with humans as “just another cog in the works”; - no pause; - universities as “an accelerator and a catalyst”; - the academic crab bucket.

What persisted. The 2018 ten AI risks “still hold”, with new additions: environmental, children’s development, “psychological/cognitive disruption” (2026-09-15). Existential risk is taken seriously but “without running around like headless chickens”. Threat-to-value and risk innovation (2026-05-03; 2026-07-16). Plausibility, now framed as informed speculation about “possible futures rather than real futures” with humility (2026-09-24). The being-human frame. The universities thread going back to 2016.

Changes of mind. - AI-assisted scholarship. “I cracked”: a Claude-drafted paper he calls “genuinely insightful and generative” (2026-01-17). - AI’s nature. From “stochastic parrot” language (2025) to rejecting the label: “Rather, they are different” (2026-01-22). - Moltbook. Alarm about emergent agent behaviour is tempered within two weeks, and the concern moves to agents nudging humans (2026-01-31; 2026-02-11). - AI prose. A year after being “blown away” by Claude’s eloquence, he finds it “superficially profound yet substantively hollow” (2026-07-19). - Authorship. He names an AI as sole author (2026-09-04). - Early GPT. “I was wrong” to think early GPT APIs were “a toy” (2026-09-24). - A darker register. AI is “one of the scariest things I’ve ever seen” (2026-09-24). - Orphan risks. Moved from an enterprise tool to a public-accountability framework (2026-07-16). The concept is absent from the September lecture, where his risk map comes from cognition, formation and identity.

Long arcs across the record#

1. Risk as a threat to value (2016–2026): steady, widening. It began as an argument to startups and funders (2016–2021). It became his definition of AI catastrophe (2023-05-31), the axis of his transitions model (2024-08-25), his account of education’s purpose (2025-03-30), part of his language of care (2025-05-18), and a frontier-AI governance framework (2026-07-16). It never changed its core definition.

2. From brains to language to formation (2016–2026). This is his most developed AI-specific thread: - direct neurotechnological action on thought (2016-03-31); - AI that exploits cognitive vulnerabilities (the 2018 book); - language as the medium of trust and influence (2023); - designed attachment, political economy and emergence (2024); - a structured manipulation risk (2025); - the bypass of epistemic vigilance, with no manipulative intent needed, and AI as a participant in human formation (2026).

The unit of concern shifts from malicious or goal-directed manipulation to ordinary, even “honest”, features of fluent machines.

3. AI’s status: from one strand to a category of its own. One member of a converging set (2015–2021). A technology at a tipping point for self-understanding (2023). “a categorical error” to treat as a tool (2025). Beyond analogy (2026). Unlike “any other technology in human history” (2026-09-24).

4. Superintelligence and existential risk: stable agnosticism, rising willingness to plan for tails. Implausible and low-probability, “not a zero probability” (2018). “potentially existential proportions” but not extinction (2023). Critique of the x-risk ideology alongside refusal to dismiss catastrophe (2024). Edge-case planning (2025). AGI “ill-defined” and “irrelevant” to the real risk map (2026).

5. Responsible innovation: from confidence to doubt to augmentation. An endorsed framework (2015–2019). “fiendishly hard to operationalize” (2023). Too narrow, and companies’ lip service (2024). Possibly “futile” under acceleration, and politically out of fashion (2025). To be augmented by care, capacity-building and institutions (2025–26).

6. Governance: soft law as a constant, government confidence falling, universities rising. Adaptive, multi-stakeholder soft law throughout, with hard law for specific harms (deepfakes 2024; designed manipulation 2025). Democratic governments lack “imagination, vision, or agility” (2025-01-07 universities-need-to-step-up), and universities should fill the gap (2026-09-24).

7. Past lessons versus novelty. Structural lessons from nano, GMOs, rDNA, chemicals and nuclear are used heavily in 2019–2023 (the past-technologies tag falls from about half of posts before 2019 to 10% in 2025–26). Claims of discontinuity strengthen from 2024. He keeps using analogies as mindsets while denying that any analogy captures AI.

8. AI and thinking: from catalyst to surrender. “a profoundly effective catalyst” (2023-08-14). Diminished critical thinking as the long-term threat (2024-08-25). Cognitive surrender and the “easy button” (2026-05-21; 2026-09-24). He never explicitly retracts the 2023 view; the emphasis moves.

9. AI literacy: from remedy to insufficiency. For all majors and “every high school” (2023). Ways of thinking (2024). Classes “risk becoming performative” (2025). Literacy alone will not change risk behaviour, so the safety message comes first (2026).

10. His own practice: from refusal to partnership to disenchantment. He does not write with ChatGPT (2023-09-20). Co-creation experiments (2024). AI-written pieces and a “synthetic research partner” (2025). AI-assisted papers (2026-01). AI as sole author, disenchantment with AI prose, and quality needing time (2026-07 to 09).

11. Tech leaders. Visionary but socially naive (2016–2019). Sharper criticism of lip service and capture (2023–2024). Engagement justified by influence (2024). A structural reading, “sincerity almost always operates inside an incentive field” (2026).

12. Inevitability. The obligation to innovate (2018). “manage and channel the inevitability of technological change” (2024-08-18). “guide and steer” (2025-03-30). “the boat has already left the harbor” (2026-05-21). “We can’t pause it” (2026-09-24).

Signalled changes of mind (consolidated)#

Thread tags by year#

Method. Counts are computed from the list of high- and medium-relevance posts compiled from the batch reading, 280 posts in all, 150 of them high. Tags were assigned by the batch readers, and a post can carry many tags. Each cell gives total posts with the tag (of which high relevance); “–” means none.

Caveats. - Years before 2019 are thin, and their dates approximate. - 2026 runs only to 24 September. - The list includes one Modem Futura promo post (2024-11-06 ai-in-a-world-of-trump). It is counted because its written framing is his. - Films from the Future itself is not counted; only posts are. - Several 2018–2020 and 2023–2025 posts republish book chapters.

Supplementary view: the share of each period’s posts carrying a tag. Years are grouped to avoid tiny denominators. Only threads with five or more posts are shown.

What the counts show, and what they don’t. - Rising. The AI-specific threads rise steadily: nature of AI, agentic AI, cognition and epistemics, and language and formation, which jumps to 40% of posts in 2026. Being human becomes the most frequent frame from 2024. AI in scholarship and writing becomes about half of posts in 2025–26. - Falling. Past technologies falls from about half of posts before 2019 to 10% in 2025–26. Responsible innovation falls from half to a fifth. The explicit risk-concept and risk-innovation tags fall after 2020, apart from the 2026 revival around the frontier-AI paper. - Tag frequency understates the risk concepts. Threat-to-value reasoning runs through posts tagged being-human, education, and complexity and transitions without carrying the risk tags: the 2024-08-25 transitions model, the 2025-03-30 education keynote, the 2025-05-18 Trajectories Tool. Its underlying weight in his thinking is larger than the counts suggest. - Orphan risks is a minor tag. It appears on 16 posts across eleven years, never more than 17% of a period’s posts. That matches the concept index’s judgement: it is a recurring tool, not the organising idea. - Existential risk peaks in 2018 and 2023. These are the book’s superintelligence chapter and the pause and CAIS debates. It returns modestly in 2026 with the extinction-alarm wave (2026-09-15). It never dominates.