T8. How he thinks: epistemic stance and method#

A thematic synthesis of Andrew Maynard’s public writing, 2014 to September 2026, on one thread: how he reasons, what he counts as evidence, how he handles uncertainty and his own errors, how he uses stories and experiments, how he treats experts and publics, and the tone he adopts. It does not compare his work with anything else.

Evidence rules. - Only his own prose counts. Excluded: AI-generated text (including the Claude-written process account in 2026-09-24), guest posts, quoted material, and the standing exclusions (2025-11-24 start-here; the 2026-06-14 PhD-site note; the “Useful stuff” in 2026-05-15; notes 2–3 of 2026-01-31). - Modem Futura podcast posts are left out entirely, at the user’s request. - The 2026-09-24 King’s College lecture body was drafted by Claude from his transcript and then line-edited by him. Its ideas are treated as his, but exact phrasing is less secure. Its written introduction and the first part of its postscript are his own prose. - His own prompt text to an AI (2025-01-30) counts as his words, written as an instruction to a machine.

Citations give date plus (shortened) slug, or “FFTF p.X” for Films from the Future (2018). “Interpretation” marks my reading rather than his claim.

1. In brief#

Maynard reasons like a risk scientist who takes not-knowing seriously. Four habits run through the whole record.

1. An evidence conscience from occupational aerosol and nanomaterial science. Weight of evidence over single studies, hazard set in the context of exposure, suspicion of “brand” hype, and a physicist’s test of whether an idea respects how the world works.
2. A discipline of “plausible, rather than simply imaginable” futures (FFTF p.171). He applies it to hype and to doom alike. He later softens it into “informed speculation” held “within a context of humility” (2026-09-24).
3. Humility practised in public. He admits he may be wrong, posts addenda and updates, republishes old work to test whether it “still stands”, and names his reversals.
4. A commitment to thinking and experimenting in public. For him public writing is “integral to how I explore, test, and share new ideas” (2026-05-17). Since 2022 the experiments have mostly been with AI, which he treats as both an object of study and his own instrument, and to which he openly admits falling prey.

Around these sit a view of films and stories as “a common point of focus” rather than forecasts (FFTF p.25); a view of expertise as plural, in which leaving technology to experts is “an abdication of responsibility” (FFTF p.288); and a balanced, candid, self-implicating, non-demonising tone whose motto is “Don’t Panic”, but never complacency.

The core commitments are constant from 2014 to 2026. What changes is the object (nanomaterials to AI), his entanglement with it (observer to heavy user), the emotional register (by 2026 AI is “one of the scariest things I’ve ever seen”, 2026-09-24), and his confidence that engagement, literacy and responsible innovation will be enough.

2. His positions, in his own terms#

2.1 The risk scientist’s conscience: evidence, proportion and a “BS monitor”#

What he says. He presents himself first as a risk professional with a long empirical apprenticeship: thirteen years of workplace dust research at HSE and NIOSH, linked to his grandfather’s death from black lung (FFTF pp.118–122), the NNI environmental and health committee (FFTF p.215 fn149), and running risk centres. He uses this record as a credential and as a source of impatience: “I have less and less patience for how many people tend to think about risk” (FFTF p.22). Later: “I’ve been working in the field of risk for over thirty years” (2023-11-26 everything-youve-heard-about-ai-risk-is-wrong), and “I have 40 years of working as a scientist” (2026-07-04 just-how-good-is-anthropics-fable, footnote).

The habits that follow from this are explicit.

• Hazard is not risk. On quantum-dot toxicity, the cautionary evidence “taken in isolation” is “misleading” (2015-01-10 are-quantum-dot-tvs). “no cause, no risk”, and speculation without a causal pathway is not risk (2023-11-26).
• Weight of evidence, not the startling single study. “knee-jerk reactions to seemingly-startling results rarely result in socially beneficial outcomes”. “a single study can be useful, but inevitably is just one small piece of a much larger puzzle” (2019-03-05 should-we-be-treating-algorithms). The same instinct drives his anger at commentators who “take a single datapoint” from the Waymo arson and build a backlash narrative that is “both lazy and dangerous” (2024-02-18 setting-fire-to-self-driving-cars).
• Method critique before verdict. He credits OpenAI’s GPT-4o system card as “a sophisticated approach”, then adds: “I’m also not 100% convinced that the methodology used is likely to represent real-world use cases” (2024-09-01 is-chatgpts-new-voice-mode).
• A “BS monitor” for hype, including his own field’s. Having worked “in nanoscale science for nearly 30 years”, he admits “my BS monitor also gets a little twitchy”. He says brand-nano “fudged the science to sell the idea” (2018-02-21 the-bs-and-the-science-of-nanotechnology). On press releases, researchers carry responsibility for “a chain of trust between the researcher and reader” (2014-12-14 researchers-should-take-more-responsibility).
• A physicist’s reality check. “To the physicist in me, claiming that perpetual motion is possible is as fanciful as believing the earth is flat”. He faults Bostrom’s Superintelligence for favouring “philosophical speculation” over “practical reality” (2024-04-28 beyond-the-future-of-humanity-institute). Gray goo is tested against the one real self-replicating system, life on Earth (FFTF pp.204–205). The Drake equation is “back-of-the-envelope” reasoning “any physicist should feel immediately at home with” (FFTF p.283).
• Science is fallible and value-laden. Science corrects itself, “sometimes decades or centuries” late. Bias lies in “how we decide what data to collect, and how we end up interpreting and using it” (FFTF p.75). Scientists are believers too, “seekers of truth, but skeptical of the truth” (FFTF p.241).

How firmly. Very firmly, and it is an identity claim as much as a method. He will say “I know my stuff” (2026-05-17 the-nonsense-i-write) and is frustrated by “people who assume a mantle of expertise on safety” without knowing the field (2024-06-20 ilya-sutskevers-safe-superintelligence). Yet he also accepts that risk is “ultimately a social construct” (2023-11-26). His empiricism is about causation and evidence, not a claim that numbers settle what matters (his risk-as-threat-to-value thread).

2.2 Disciplined imagination: plausible versus imaginable#

What he says. The 2018 book states the discipline most fully: - “humility alone isn’t enough. There also has to be some measure of plausibility” (FFTF p.168); - “what is plausible, rather than simply imaginable, is vitally important” (FFTF p.171); - Occam’s Razor becomes a way to rank futures. Gray goo and superintelligence “depend on a house-of-cards stack of assumptions”, and funding them over evidence-based materials harms is “more an act of faith than of reason”. But that is a lower probability, “not a zero probability” (FFTF p.281; 2018-11-01 contact-occams-razor); - exponential extrapolation is “beguiling” and error-amplifying, and Moore’s Law is “a self-fulfilling prophecy” (FFTF pp.199–200).

The stakes are practical. Speculation hurts people “when make-believe is treated as plausible reality”, through techno-terrorism (the ITS bombing of nanotechnologists) and also through policy, advocacy and investment that forgo beneficial technology (FFTF pp.205–206). The “fatal flaw” of drastic action is “the assumption that we can predict with confidence what the future will bring” (FFTF p.199). In sum: “as soon as we start to believe our own fantasies, we have a problem” (FFTF p.206).

He names the technique as his own in 2024. First let imagination run, then rein it in. This is “a technique that I’ve used in the past to explore the plausible boundaries of emerging technologies”, so that ethics is “grounded in plausibility rather than hyperbole” (2024-11-17 navigating-the-ethical-dilemmas-of-bci). He uses scenarios the same way. Scenarios “allow speculative boundaries to be placed around plausible AI futures”, but they tend to treat AI as “something that happens to society” (2025-11-30 postscript-letters).

How it develops. The 2018 version mostly deflates. From 2023 the same discipline justifies preparing for tails. He accepts a risk “of potentially existential proportions” while refusing the FLI pause (2023-04-04 what-are-the-alternatives-to-calling). He calls AI 2027 “speculation — no more”, yet worth planning for “just on the off chance that there’s a sliver of truth here”, and turns his critique of exponentials into a diagnosis of human blindness: “it always will feel like an intellectual exercise until it’s too late” (2025-04-06 responsible-innovation-and-ai-acceleration).

By 2026 the formula is procedural. When technology outpaces data, “don’t disallow speculation, but do it within a context of humility”. That means recognising it as speculation, looking at “possible futures rather than real futures”, seeking data and “bringing in different voices” (2026-09-24 being-an-academic, Q&A note). He calls AGI-and-singularity speculation “blinkered and naive”, and “nothing new under the sun” dismissal also speculation (same note).

How firmly. Plausibility is his most consistent epistemic discipline and one he returns to unprompted. What moves is where he sets the threshold. Interpretation: Occam’s Razor, which discounts stacked assumptions, sits uneasily with his complexity thinking, which warns of blindsides from non-linear systems (FFTF p.260). By 2025 the complexity side is winning.

2.3 Humility as method, and changing his mind in public#

What he says. He hedges by habit, and the hedges are meant: “Here, I freely admit that I may be wrong” on superintelligence (FFTF p.170); “I may be wrong there” on scaling (2024-10-08 ai-captures-this-years-nobel-prize); “week by week … less sure what the future holds” (2025-03-30 reimagining-education).

His fullest statement is the 2023 risk essay. It says, “the more I study artificial intelligence, the less certain I am that we even know how to formulate the problems we face around AI”. The “understanding-vacuum” is being filled by “dogmatic overconfidence”, and “only the foolish would claim to understand with certainty”. It then turns on itself: “Maybe I was a little hubristic with my statement”. It closes by asking whether we will have “the collective humility and imagination to learn together” (2023-11-26). A day later he adds an addendum because “It’s been bugging me” that he had left out the hazard–exposure formulation.

For him humility is procedural and collective, not only a personal virtue. The remedy is to “drop our egos, recognize the limits of our own understanding, and have the humility to listen to and learn from others” (2023-11-26). His proposed AI “world congress” should pick people who can work together “without ego getting in the way” (2023-04-04). Scientific advocacy should be collective, because “it’s hard to justify one person being the sole arbiter of truth” (FFTF p.247). By 2026 he wants “a collective form of epistemic vigilance” by humans-in-the-loop (2026-01-17 i-cracked-and-wrote-an-academic-paper).

Changing his mind in public. He flags reversals, often by date: - Deepfakes. He once trusted common sense to spot fakes. “Today I am far less sure”, so he signs a letter he would normally avoid (2024-02-25 ai-rollercoaster-of-a-week). - Technology apologetics. He questions “a form of technology apologetics that’s been part of my professional life for decades” (2024-03-31 we-have-a-technology-problem-and). - AI fiction. “A week ago I would have said no way”, then “Or so I though” [sic] on ChatGPT’s fiction (2024-09-28 chatgpt-as-author-part-1). - Agentic social AI. Revised into “stochastic agency” within a week (2024-10-27). - Musk. A “rather naive perspective on Elon Musk” in his 2018 text (2025-03-02 the-lure-of-permissionless-innovation). - Parasocial communication. “Clearly I read the tea leaves wrong back in May!”, while he still stands by the core idea (2025-11-19 parasocial-relationships-problematic). - Early GPT. He thought early GPT “a toy. It’ll never catch on.” “I was wrong.” (2026-09-24). - AI prose. From being “blown away” by it to finding it “superficially profound yet substantively hollow” (2026-07-19 publish-or-perish).

He also corrects in place: - an update after Gregor Wolbring pointed out that “suffering” from quadriplegia implied a judgement that was “clearly not right or appropriate” (2024-01-30 first-in-human-trial-of-neuralink-bci); - an update telling readers to “read with caution” once the MIT productivity paper he had relied on was disavowed (2024-11-10 is-ai-poised-to-suck-the-soul-out-of-science); - a footnote that reverses the claim it annotates: “Re-reading this, I’m not so sure” (2026-04-26 why-im-falling-out-of-love-with-claude).

Republishing is itself a test. He reposts his 2018 chapters and says they are “more relevant now”, or that “much of what I wrote back then still stands” (2025-03-02). He brings back the 2018 Risk Bites list of AI risks in 2026 to check whether it holds (2026-09-15 will-ai-really-kill-us-all).

How firmly. Firmly as a stance, but it is not relativism. He accepts that he may be wrong and still keeps his judgement. In 2026 he floats the possibility that he is “an anachronism from a previous age, lost in his own myopic hubris”, and then says: “OK, so I actually don’t believe this” (2026-07-19). He even allows “a change [chance] of course that the LLMs are right and I’m wrong”, yet he still defends his standard of writing on the grounds of lived human experience.

2.4 “Don’t Panic”: against both hype and doom#

What he says. The book closes with “Don’t Panic” and “we shouldn’t be complacent—far from it” (FFTF pp.289–290). The twin errors are fatalism and becoming “so enamored by the tech itself” (FFTF p.290). His positive stance is an “obligation” to innovate, since renouncing technology “from a position of privilege” denies others choices (FFTF p.288). He still calls himself “a technology geek” (FFTF p.17), who will “revel in their potential” while being “deeply cognizant” of harm (2023-10-19 marc-andreessen-ditch-sustainability).

He rejects the optimist/pessimist frame outright. Asking whether he is one or the other is “a bit like asking if I’m an oxygen pessimist or optimist” (2024-03-31). In 2026: “I’m neither an AI optimist nor an AI pessimist. But I do have really, really bad days with this technology” (2026-09-24).

The symmetry shows in his practice.

• Against hype. He calls out brand-nano (2018-02-21), university press releases (2014-12-14), Musk’s “telepathy” hyperbole (2024-03-21), “technological foreshortening” in Andreessen’s history (2023-10-19), and AI developers “acting as if they’re the first people to notice” long-known risks (2026-09-15). His “pencil and paper” test for “hyped up educational tech” asks whether we “manufacture problems” to justify new tools (2024-12-29 fantasy-top-ten-lists-2025, fn4).
• Against doom, without dismissing catastrophe.
• He did not sign the FLI pause letter (2023-04-04) or the CAIS extinction statement. Extinction is “too narrow and absolute a framing, and too human-centric” (2023-05-31 existential-risks-of-ai).
• Yet “simply ignoring the possibility of potentially catastrophic events … is in itself a risky strategy”. He asks for ways to hold low-probability, high-impact risks “in context” that “open up conversations rather than closing them down” (2024-06-23 existential-risk-jay-baruchel).
• Existential risk from AI is “not that likely” but should not be dismissed. It should be handled “without running around like headless chickens” (2026-09-15, fn5).
• Against polarised framings as such. Polarised books sell, but “dialogue and discussion are far more important than preaching” (2023-07-07 moviegoer-guide-chapter-1). He wants to move “away from stances and statements that have all the sophistication of a bumper sticker” (2023-07-25 oppenheimer-and-ai). An AI documentary left him “drowned in opinions that were only loosely tethered to reality — whether from the techno-doomers or techno-optimists” (2026-03-22 are-you-an-ai-apocaloptimist). Technology moral panics are “rarely cut and dried” and should not be mocked, because they signal threats to what matters to people (2025-06-01 vibe-coding-moral-panic).

He also insists that naming risk is not fear-mongering: “it never ceases to amaze me how many people equate talking about risk with fear mongering” (2026-09-15, fn1). He puts “the safety message first” and complains that “Talking about the potential risks of AI is not popular. It gets you branded as a technology-pessimist, or even a Luddite” (2026-05-10 do-not-do-this-with-ai). He jokes about “having the temerity to be balanced” in an attention economy (same post).

How firmly. This is a constant temperament, from the 2015 “we can’t afford to slam the breaks” [sic] (2015-01-30) to 2026. What moves is the centre of gravity: - 2018 mostly deflates exaggerated fears; - 2023–24 accepts potentially existential stakes and calls for serious engagement; - 2026 says AI “scares the life out of me sometimes”, adding “Some of this is rational, some of it is irrational” (2026-09-24).

Interpretation: the balance is held with increasing effort, and he now reports his own emotional state as data to “calibrate” against (“just so you can calibrate”, 2026-09-24).

2.5 Scholarship and experimentation in public#

What he says. Public writing is not outreach for him. It is how he does scholarship. He started with a nanotechnology blog in 2007, and public writing has “become central to my conviction that the work I do as an academic should be as accessible as possible”. It is “Not as an add-on to my research and scholarship, but as something that’s integral to how I explore, test, and share new ideas and insights” (2026-05-17). He grounds this in a duty: “the privilege of academic scholarship and research comes with an obligation to ensure that the knowledge we unearth is accessible to anyone who can benefit from it” (2024-09-04 succeeding-at-science-on-youtube, on Risk Bites, launched 2012 “with limited time and even less talent”). He began the Substack in 2023 after discovering he had written almost nothing on AI risk “that’s neatly citable” (2023-04-04 welcome-to-the-future-of-being-human). Looking back in 2026, he describes choosing the Substack over “higher-prestige platforms” as the cost of deciding to “put public good before academic prestige” (2026-05-17).

Thinking in public means publishing work in progress: a governance thread that is “a little rough” but “worth posting” (2023-04-04); an essay that is “part of my process of trying to marshal my thoughts” and ends “I’m sorry to disappoint, but I don’t have one” (2024-03-31).

Experimentation, especially with AI. From 2022 his main way of knowing AI is to use it hard and report what happens: Midjourney and ChatGPT (2022-09-16; 2022-12-08); annual reviews, where it “slightly worries me” that he already thinks of ChatGPT “as a colleague and a collaborator” (2023-01-31); “over 2,000 conversations” from his own course (2023-08-14 chatgpt-stimulates-creativity); an author–editor loop and then “an honest conversation between a human (me), and a machine” that was “generative for me” (2024-09-28; 2024-09-29); a Deep Research “dissertation” and the warning that unaugmented scholarship may look “somewhat quaint” (“if you find that a little disconcerting, you probably should”, 2025-02-04; 2025-02-09); agent tests (2025-03-22); home-built evaluation tools with prompts on GitHub (XENOPS, 2025-07-13); a Claude-assisted arXiv paper (2026-01-17); a 169-film corpus with open data (2026-05-15); Fable 5 as research partner (2026-06-12; 2026-07-04); and an anonymised comparison of his edit against Fable’s paper, put to readers (2026-07-19).

Three norms recur in how he reports these experiments.

1. Disclosure and labelling of limits. - He gives process accounts, the models used, and who did what. - He is plain about scale: “admittedly with an “n” of one” (2026-07-19). The films corpus “doesn’t rise to the level of a publication-quality study yet”, with Claude as “my not-always-reliable research assistant” (2026-05-15). - He flags credit: “the concept of honest non-signals came from Claude” (2026-01-17).
2. Self-implication as data. He was “suckered by Claude at the very moment I was writing about the risks of being suckered by Claude!”, and sums up: “The reasoning was impeccable. The advice unfounded.” (2026-02-08 beeswax-hallucinations). He turns his epistemic-vigilance thesis on his own paper: “how do I know I’m not an unwitting victim here?” (2026-01-17). His five “do not do” rules are “personal because these are all things that I find myself having to think about” (2026-05-10). He admits a working relationship with an LLM while avoiding anthropomorphism: “I strenuously avoid anthropomorphizing AI apps … But I do have a relationship with the LLMs I use” (2026-04-26).
3. Refusing AI where meaning matters. - “I typically don’t use ChatGPT myself when I write”. To hand that over “would be to diminish myself” (2023-09-20 what-do-college-students-think-about-chatgpt). - His Action Plan analysis was “very intentionally not an AI-generated first take”, because summaries miss “meaning, implications, subtexts” (2025-07-23 americas-ai-action-plan, fn1). - The Letters correspondence was written without AI as “an intentional meta-reflection” (2025-11-30). - When he breaks his own rule he says so and gives reasons: “I’ve resisted doing this for so long” (2025-01-30 ai-at-a-crossroads).

How firmly, and how it changes. Public scholarship is a lifelong, firm commitment, though at the start of his 2026 sabbatical he doubts whether it does public good: many see him as “just a commentator”, and “to write without care for your readers is a very academic trap” (2026-05-17). His use of AI as an instrument moves from refusal (2023) to partnership (2025) to “I cracked” (2026-01) and partial disenchantment: “AI used well doesn’t necessarily make things faster if you’re going for quality” (2026-09-24 postscript, his own prose). Experimenting as a way of knowing is constant; his confidence in what the experiments show is not.

2.6 Stories, films and imagination as ways of knowing#

What he says. Films from the Future rests on an epistemic claim: science-fiction films are “pretty bad at predicting future technologies” but good at showing how technology and society intertwine (2018-10-12 Q&A). They are useful “not because they are accurate or prescient, but precisely because they are not tethered to scientific accuracy”, when “seasoned with feet-on-the-ground thinking” (FFTF p.288). Art is “a common point of focus” across divided audiences, and imagination is valuable “as long as that imagination is grounded in reality where it matters” (FFTF p.25). Reason and imagination need each other: “Critical thinking alone is almost inhuman in its cold impartiality. On the other hand, creativity on its own leads down a path of fantasy and delusion” (FFTF p.282).

In his teaching he stresses emotion and shared experience: “transformative learning has to be felt” (2021-01-15 can-watching-sci-fi-movies). Stories persuade where argument fails: “Preach to someone about the future, and most people will shut down”, but “tell them a story that resonates with them, and you open their mind” (2024-01-21 how-can-stories-unlock-pathways). He reads a “doom and gloom” monoculture as removing agency (same post, endorsing the film’s line) and the dystopian default as a hazard of its own (FFTF p.289).

He also writes fiction as method, because “there are affordances in fiction that allow complex ideas to be explored with a nuance and sophistication that all too easily elude more literal pieces” (2025-11-23 letters-from-the-department-of-intellectual-craft-prelude; also “Soul Update”, 2026-02-11). He values ideas he does not believe for their “creative jolt to the imagination” (2025-12-21), treats play as a way of learning (2024-03-17, on physics as play; 2026-08-02), and tests the dystopia trope itself: most AI films in his corpus are not purely dystopian (2026-05-15).

Cautions he attaches. - Entrepreneurs take the “cool tech” and miss the social message, so tech leaders should keep watching sci-fi “only if they take on the broader message” (2024-05-26 should-tech-entrepreneurs-be-banned-from-scifi). - He adopts Virk’s “Sci Fi feedback loop” as a warning (2024-09-18). - He knows nuance does not sell. By most standards the book “was a failure”, in an age where “certainty rather than subtlety sells”, and colleagues called the popular book “professionally embarrassing” (2023-10-08 a-guide-to-responsible-innovation).

How firmly. Core and constant, from the 2018 course and book through the 2026 lecture’s use of the Ring of Power and the Tesseract as a structural archetype of power that changes its user (2026-09-24).

2.7 Experts and publics#

What he says. He holds two claims together. Expertise is real and needed: “Naturally, we still need technical experts, laws, and policies” (FFTF p.25). But it is plural and bounded: “Scientists and technologists don’t have a monopoly on expertise and insights”, and “most people have a pretty high level of expertise in what’s important to them and their communities” (FFTF p.222). “one thing we’re all qualified to do” is to think about what technology’s consequences “might mean to us and the people we care for” (FFTF p.26).

Leaving these questions to scientists, innovators and politicians “is, in itself, an abdication of responsibility” (FFTF p.288). Public indifference hands the field to whoever cares, so “we collectively need to give a damn” (FFTF p.286).

Across the posts: - Against scientific elitism. He criticises the attitude that “nonscientists” should “revere, but not interfere with, science”, using Flint as a case where experts and officials were wrong and citizen science was right (2016-01-12 can-citizen-science-empower). - Institutions too. Public-good work at his Michigan centre happened “despite the institution we were a part of” (2016-01-31 public-universities-must-do-more). - Against “leave it to the experts” in AI. “it’s complicated” is not an excuse, and “everyone has the right to play some role” (2023-05-15 erik-schmidt-ai-regulation). Policy risks being set by “loud (but not necessarily informed) voices” while “quiet voices” are ignored: governance scholars, developers buried in companies, affected communities, and “ordinary people who sometimes have extraordinary insights” (2023-04-10 as-ai-goes-to-washington). - Against narrow expertise, including expert crowds. Cutting-edge AI experts are strong on AI but light on governing emerging technologies (2023-05-17). Crowd-sourced expert surveys “regress to the mean” and underrate what mainstream experts do not understand (2025-01-19 wef-global-risks-2025). - Against the deficit model. It has “been repeatedly shown not to be effective” (2025-05-25), and Amodei’s use of it was “debunked decades ago” (2024-10-13). Relational, even parasocial, communication asks experts to “make themselves vulnerable” and to value people “whom they might otherwise dismiss” (2025-05-25). - His role. Pielke’s honest broker: “This is the role I try to carve out for myself in my public-facing work, trying not to judge others or advocate for a specific course of action” (FFTF p.246). Where silence would amount to consent he qualifies this, preferring collective and institutional advocacy (FFTF pp.246–247).

How it develops. The commitment is constant. His answer to who should act narrows over time. By 2026 publics are “critically important”, but “you cannot hand a problem of this magnitude over to everyday people”, and universities should fill the gap as “an accelerator and a catalyst” (2026-09-24). His criticism of his own tribe grows sharper. He describes the academic “crab bucket” from chairing ASU’s promotion and tenure committee (2026-09-24), and doubts that universities will lead: “Sadly, this has been my experience so far” (2026-08-30 do-universities-have-a-place-in-bill).

2.8 Tone and the ethics of engagement#

What he says, and how he writes. Several features are consistent enough to count as an ethic.

• Non-demonising diagnosis. Developers usually fail through absorption, not malice. Scientists “care deeply” (FFTF p.17) and are rarely self-consciously unethical (FFTF p.36). The failure is “myopically benevolent science”. He includes himself, “the single-mindedness and benevolent myopia I see in many of my peers, and even myself at times” (FFTF p.220), and generalises: “All of us, it has to be said, have a bit of Sidney Stratton in us” (FFTF p.227). In 2026 he refuses a cynical reading of AI firms: “sincerity almost always operates inside an incentive field” (2026-07-16 orphan-risks-frontier-ai, his framing).
• Self-implication. He confesses his PhD rule-bending (“it’s shocking how quickly I sloughed off any sense of responsibility to get the data I needed”, FFTF p.161). He admits he was “quite happy to coopt a narrative of social good” for curiosity-driven work (FFTF p.219). He was fooled by Claude (2026-02-08), and he worries his public writing may be “an ego trip” (2026-05-17).
• Charity toward opponents. The CAIS statement deserves “pausing to consider the statement rather than dismissing it as merely fear mongering or naive speculation” (2023-05-31). Amodei writes with “a vision that’s tempered with humility and reason” (2024-10-13). Google gets sympathy after the Gemini failure (2024-02-25). Of a film he criticises: “there is something rather churlish about reviews that overlook what has been achieved” (2026-03-22). Of Musk: “I get where he’s coming from”, and what matters is “not how wild the ideas are, but how much influence they have” (2024-08-04).
• Directness when warranted. He is angry at the Waymo mob and the “lazy” narratives around it (2024-02-18). He calls tech leaders’ good intentions “sheer fantasy” as a basis for beneficial AI (2023-12-15). He criticises his own university’s AI boosterism (2026-05-03; 2026-09-24).
• Invitation, humour and self-deprecation. Posts end by asking for readers’ views. He is “a risk & emerging tech professor who can’t draw” (2026-05-10), made Risk Bites with “little time and no talent” (2024-09-04), and apologises “to the attention economy gods” (2026-05-10).
• Stated values. In a prompt describing himself to ChatGPT he wrote: “I do not write pieces that are polarizing, that are preachy, that push an ideology or an agenda” and “I do place human wellbeing and flourishing at the heart of my work” (2025-01-30, his prompt text).
• Norms for disagreement. “Collaborate with humility, and be willing to change your perspective in the light of new information” (2024-05-12 chatgpt-shaming-is-a-thing).

How firmly. This is his most stable trait. The register darkens slightly in 2025–26, with more open criticism of ASU and US policy and more open fear. The ethic of respect, candour and invitation does not change.

3. How it developed over time#

Formation, to 2014 (from his later accounts). Physics training (FFTF pp.219–220; 2024-03-17), occupational aerosol science at HSE and NIOSH (FFTF pp.118–122), nanotechnology risk at NNI and PEN (FFTF pp.214–218), the 2020 Science blog from 2007 with its 2009 civil-society series (FFTF p.191), and Risk Bites from 2012 (2024-09-04). Two roots of the method are set here: an evidence conscience, and the conviction that knowledge must travel.

2014–2017: the evidence voice with a public conscience. He writes on press-release exaggeration (2014-12-14), hazard in context (2015-01-10), citizen science (2016-01-12) and public universities (2016-01-31). He reposts Winston’s “Scientist’s Manifesto” because it “still resonate[s] deeply with me” (2014-11-23; the manifesto is Winston’s), and admits Bill Joy’s vision “still haunts me” (2016-01-11 the-fourth-industrial-revolution). Humility here is mostly institutional: experts must engage and serve.

2018 to mid-2019: the method stated in full. Films from the Future is his epistemic manifesto: films as lens, plausibility and Occam’s Razor, the honest broker, the self-implicating account of scientists, “It’s good to talk”, and “Don’t Panic”. He adds that the book is no “infallible” guide (2019-03-06). The chemicals-to-algorithms essay (2019-03-05) shows the technocratic evidence voice at full strength.

2019–2022: teaching and consolidation. The film course makes the method pedagogical (“has to be felt”, 2021-01-15), and Future Rising (2020) extends it to stories. His first generative-AI experiments (2022-09-16; 2022-12-08) are strikingly positive, hailing “a leap in machine-augmented communication” while listing attachment, gullibility and disinformation.

2023: thinking in public becomes the medium. The Substack brings fast, provisional posts (“a little rough … worth posting”, 2023-04-04): refusals to sign letters grounded in “no silver bullets”, the quiet-voices argument (2023-04-10), the humility essay and addendum (2023-11-26), classroom-based educational optimism hedged by “I may still be suffering from an overdose of ChatGPT” (2023-08-14), a personal rule against writing with AI (2023-09-20), and a candid account of the book’s “failure” (2023-10-08).

2024: open revision. He changes his mind on deepfakes and “technology apologetics” (2024-02-25; 2024-03-31), on AI fiction (2024-09-28), and on agentic influence within a week (2024-10-27). He critiques FHI through thermodynamics while calling for futures work “grounded in humility” (2024-04-28), names plausibility as his technique (2024-11-17), and consults a colleague (Hazel Kwon) before judging a persuasion study (2024-09-01).

2025: AI inside the method. He breaks his writing rule (2025-01-30), floats the “artisanal intellectual” (2025-02-09; “a bit of a throwaway”, 2025-02-16), uses reasoning models for synthesis and reverses his 2018 view of exponentials (2025-04-06), sets out and then publicly corrects his parasocial theory of engagement (2025-05-25; 2025-11-19), builds tools to think with (2025-06-01; 2025-07-13), refuses AI for interpretive policy analysis (2025-07-23), insists even objectors “need to know what you’re talking about” (2025-08-10 the-scared-witless-educators-guide-to-gpt5), and moves into fiction (2025-11-23 to 2025-11-30).

2026: self-implication, disclosure and a darker register. “I cracked” with full process disclosure and a named credit problem (2026-01-17); the beeswax confession (2026-02-08); rules of thumb framed as personal (2026-05-10); an open film corpus (2026-05-15); an anonymised writing contest in which he entertains being wrong (2026-07-19); sabbatical doubt about his public writing (2026-05-17); and, at King’s, “I was wrong”, fear stated openly, and “informed speculation” with humility (2026-09-24).

Constants: evidence plus humility plus imagination; symmetric scepticism toward hype and doom; public scholarship as duty; plural expertise and distributed responsibility; a non-demonising, self-implicating tone; films and stories as tools for thinking.

Shifts. 1. Plausibility moves from mainly deflating speculative risk (2018) toward preparing for tails (2023–26). 2. He moves from studying AI at a distance to using it as a core instrument, and becomes his own case study of its risks. 3. Confidence falls in engagement, literacy and responsible innovation as remedies: literacy classes “risk becoming performative” (2025-11-09), and the safety message comes first (2026-05-10). 4. Emotional candour grows, from “Don’t Panic” (2018) to “scares the life out of me” (2026). 5. His role moves from honest broker toward more open advocacy: a signed letter (2024), rules of thumb (2026), universities told to “step up” (2025–26).

4. Key concepts (his, unless marked)#

5. Connections to his other threads#

• Risk as a threat to value, and risk innovation. His epistemics explain why he widens risk beyond measurable harm. If harm is “ultimately a social construct” (2023-11-26), and publics are expert in what they value (FFTF p.222), then defining risk cannot be left to technical experts. Plural expertise is the epistemic basis for threat-to-value.
• Responsible innovation and could/should. His definition of responsible science is epistemic: “the humility to recognize your limitations, and the willingness to listen to and work with others” (FFTF p.39). Myopic benevolence and self-certified responsibility (FFTF p.162) are failures of knowledge before they are failures of ethics.
• Hubris and permissionless innovation. Hubris is the vice his method is built against. The reversibility test (2025-03-02) squares his love of experiment and play with his critique of “move fast and break things” and of “naive experimentation pushing us past hard-to-spot tipping points” (2023-11-29).
• Manipulation, cognition and formation. His account of human reasoning, that we build reality from “shadows” (FFTF p.176) and are “wonderfully adept at feeling like the decisions we make have a rational basis” (FFTF p.277), underlies both his humility and his AI-risk thesis. Being fooled by Claude and wondering whether his AI-assisted paper escaped his own vigilance make him evidence for the cognitive Trojan horse (2026-01-10). The thesis that AI bypasses critical reasoning is also a threat to his own method.
• Being human; writing as self. His refusals of AI in writing (“diminish myself”, 2023-09-20; “the very human piece of me”, 2025-01-30) and his embodiment argument about reading (2026-07-19) tie method to identity.
• Education and the university. Playgrounds, play and films in class are his pedagogy; his public-scholarship ethic drives his critique of promotion-and-tenure incentives (2016-01-31; 2026-09-24).
• Technology transitions and analogy. He uses past technologies as structural lessons and mindsets, while insisting AI “defies analogy” (2026-01-22). This is the epistemic form of his transitions thinking.
• Governance. “No silver bullets” (2023-04-04) and a preference for agile, soft-law approaches are the governance form of his epistemic humility.

6. Tensions, ambiguities and gaps#

1. Humility and authority. He declares that “everything you’ve ever heard about AI risk is wrong” and that he is less certain the more he learns. In the same register he invokes “over thirty years”, “40 years” and “I know my stuff”. He also dismisses other people’s “mantle of expertise” (2024-06-20). These are compatible: humility about the problem, confidence about method. But readers may find the boundary unclear, and he rarely says which of his own claims are low-confidence.

2. Plausibility versus preparing for tails. The 2018 Occam’s-Razor ranking (superintelligence as “an act of faith”) and the 2025–26 willingness to plan “on the off chance” rest on different thresholds. He signals the shift on exponentials (2025-04-06), but never restates the plausibility rule itself. He republished the 2018 superintelligence scepticism unchanged in 2023 and 2025, while his views on machine consciousness and agentic capability moved (see timeline.md). Interpretation: the discipline is intact; the evidence base is in flux, and he has not said how he updates it.

3. Anecdote, n-of-one and self-selected evidence. Much of his AI knowledge comes from his own experiments and his own classes: a course he designed (2023-08-14), a single novella, one paper, one hat. He labels this candidly. But strong claims rest on it, for example ChatGPT as “a profoundly effective catalyst” or human-only scholarship as soon “quaint”. The same person is designer, subject and judge, which is the epistemic-vigilance problem he himself names (2026-01-17).

4. Instrument and object. He uses AI to study and write about AI: a Claude-drafted lecture, a Fable-assisted paper, a concept “from Claude”. He discloses this, but it makes his late thinking harder to attribute, and his warning that AI is “beginning to train us to think like them” (2026-07-19) applies to his own workflow. Beyond line-editing and source-checking, the corpus does not show how he guards against it.

5. Honest broker or advocate? He adopts the honest-broker role (FFTF p.246), but by 2026 he is issuing rules of thumb, signing a letter, asking for “safety message first”, and telling universities to “step up”. His own 2018 qualifier covers some of this: advocacy is fine when silence would be complicity, preferably through institutions. But he does not say when he crosses that line, and his advocacy is mostly individual.

6. Publics: everyone a stakeholder, but who acts? “everyone has the right to play some role” (2023-05-15) sits beside “you cannot hand a problem of this magnitude over to everyday people” (2026-09-24). He names engagement mechanisms (consensus conferences, ECAST, PTA) but rarely specifies how publics’ “expertise in what’s important to them” would shape AI decisions in practice. The move to universities as the acting institution is reasoned but only partly argued.

7. Relational persuasion cuts both ways. He champions stories and relational or parasocial communication because they get past the defences that “preaching” triggers (2024-01-21; 2025-05-25). He fears AI because it does exactly that: stories “hard not to trust, and yet are not trustworthy” (2024-09-22). He notes that the same mechanism can serve “widespread social manipulation and control” (2025-05-25) and later asks what healthy parasociality looks like (2025-11-19). But he offers no criterion for when persuasion through story or relationship respects the audience’s agency and when it bypasses it.

8. Balance as method, and its costs. His refusal of polarisation is principled, and he says it costs him attention (“certainty rather than subtlety sells”, 2023-10-08; “the temerity to be balanced”, 2026-05-10). A reader might see some positions as hedges, for example a conditional pause on emotion-exploiting companion bots (2024-10-27) set against “We can’t pause it” for AI in general (2026-09-24). He does not always say which considerations would move him from one side to the other.

9. Gaps. He rarely engages critics of his own work by name (the Wolbring correction is an exception). There is no sustained account of evidence standards for societal-scale AI risk when data lag the technology; the 2026 “informed speculation” note is the nearest. His podcast practice is excluded at the user’s request, so his engagement practice is probably under-represented here.

7. Most important sources for this thread#

1. Films from the Future (2018), especially Ch. 1 (pp.14–26: films as lens, everyone qualified), Ch. 8 (pp.159–178: self-implicating permissionless innovation, plausibility), Ch. 10 (pp.218–229: myopic benevolence, “It’s good to talk”), Ch. 11 (pp.237–249: honest broker, collective advocacy), Ch. 13 (pp.276–286: belief, Occam’s Razor) and Ch. 14 (pp.287–291: Don’t Panic, no abdication to experts).
2. 2023-11-26 everything-youve-heard-about-ai-risk-is-wrong. His fullest statement of epistemic humility about AI risk, with its own self-correction and addendum.
3. 2026-05-17 the-nonsense-i-write. Public writing as scholarship; its costs and his doubts.
4. 2026-09-24 being-an-academic-in-an-age-of-ai (lecture; Claude-drafted, his ideas). “I was wrong”, “neither an AI optimist nor an AI pessimist”, informed speculation with humility, the crab bucket.
5. 2026-02-08 beeswax-hallucinations with 2026-01-17 i-cracked-and-wrote-an-academic-paper. Self-implication and disclosure in AI experiments.
6. 2024-03-31 we-have-a-technology-problem-and. The oxygen analogy, “technology apologetics” questioned, thinking as work in progress.
7. 2024-04-28 beyond-the-future-of-humanity-institute. The physicist’s reality check on speculative ideology, and futures work “grounded in humility”.
8. 2023-05-31 existential-risks-of-ai with 2026-09-15 will-ai-really-kill-us-all. Neither hype nor doom on catastrophic risk.
9. 2019-03-05 should-we-be-treating-algorithms. The evidence voice: weight of evidence, no single-study alarm.
10. 2023-04-10 as-ai-goes-to-washington with 2023-05-15 erik-schmidt-ai-regulation. Experts and publics, quiet voices, “it’s complicated” as no excuse.
11. 2025-05-25 why-parasocial-communication-is-important with 2025-11-19 parasocial-relationships-problematic. His theory of expert–public communication and a public self-correction.
12. 2024-01-21 how-can-stories-unlock-pathways with 2025-11-23 letters-…-prelude. Stories and fiction as ways of knowing.
13. 2026-07-19 publish-or-perish-ai-vs-human-vs-human. Testing his own judgement against AI in public, and entertaining that he is wrong.
14. 2023-10-08 a-guide-to-responsible-innovation. The film method, why nuance does not sell, and the academic cost of popular writing.
15. 2016-01-12 can-citizen-science-empower with 2014-12-14 researchers-should-take-more-responsibility. Early roots: against “revere, but not interfere”, and researchers’ responsibility for hype.