Concept index: Andrew Maynard on risk, technology, AI and being human (2014 to September 2026)#

A consolidated glossary of the concepts and frameworks in Andrew Maynard’s public writing, built from the batch notes (notes/B01–B32 and their digests) and the chapter notes on Films from the Future (book/notes-FFTF-A–F). It is a reference map of his thinking for later use as a lens. It does not compare his work with anything else.

How to read this file#

Evidence rules. Only his own prose counts. The following are excluded as evidence, although what he chose to publish and how he framed it is sometimes noted: - AI-generated text: ChatGPT, Claude, Fable, o1-pro, Deep Research, Manus and Perplexity output, including AI-written papers and stories. - Guest posts (Allenby, Miller, Dudley on GLP-1) and quoted material. - Modem Futura podcast posts, at the user’s request. The one partial exception is 2024-11-06 ai-in-a-world-of-trump, a Modem Futura promo that contains a few paragraphs of his own written framing on governance. It is cited once and flagged. - The standing rulings: most of the PhD-site note in 2026-06-14 everything-you-wanted-to-know-about; the “Useful stuff” definitions in 2026-05-15 ai-movies-may-be-less-dystopian-than-we-think; notes 2–3 of 2026-01-31 lost-in-the-moltbook-hall-of-mirrors; and 2025-11-24 start-here.

Some sources are co-written and so count as weaker evidence: - 2019-08-13 responsible-innovation (with Elizabeth Garbee) - 2019-11-19 the-trouble-with-connectedness (with Bas Boorsma) - 2023-09-11 its-time-to-get-serious-about-ai-and-sdgs (with José Lobo) - AI and the Art of Being Human (with Jeff Abbott, and deliberately written with AI assistance), whose excerpts and foreword are treated as co-written

The following are mixed-provenance sources, and only his framing counts: - 2026-07-16 orphan-risks-frontier-ai-maynard. A Fable-assisted paper that he rewrote; he claims its “argument architecture”. - 2026-01-17 i-cracked-and-wrote-an-academic-paper. Claude-drafted. He says the concept “honest non-signals” “came from Claude”. - 2026-09-24 being-an-academic-in-an-age-of-ai. His lecture, drafted into prose by Claude and line-edited by him. The ideas are good evidence; the exact phrasing is slightly less secure.

Citations. Posts are cited as date plus slug. Once a post has been cited with its slug, later mentions may give the date alone. Medium-era slugs drop their trailing hash, and some long slugs are shortened. The Appendix maps every date cited to its full corpus filename. “FFTF p.X” means Films from the Future: The Technology and Morality of Sci-Fi Movies (2018), by page.

Dates. Before 2019 the dates are approximate. Many are migrated Conversation, Medium or 2020 Science posts, and some carry Substack dates that differ from first publication. Several later posts republish older text: - 2023-04-16 and 2025-03-02: the 2018 Ex Machina chapter - 2025-04-13: the 2018 Jurassic Park chapter - 2023-11-15: the 2018 orphan-risks article - 2020-01-01 and 2020-09-26: the 2018 Minority Report chapter - 2022-02-10: a 2020 talk

In these cases “first appearance” gives the earliest date the idea is documented, not the reposting date.

Centrality scale. - Core: spans many years, organises other ideas, and he returns to it unprompted. - Recurring: appears repeatedly across several periods, but supports rather than organises. - Occasional: a handful of appearances, or important in one period only. - One-off: a single developed appearance so far.

A weighting note. Orphan risks, the subject of his 2026 paper, is treated here as what the record shows it to be. It is a recurring tool inside his much larger and older risk-innovation and threat-to-value framework, first named in 2018. It is not the centre of his thinking. The ideas with the longest, densest record are these: - risk as a threat to value, and risk innovation - responsible innovation and could/should - hubris and the ethics of permission - complexity and irreversibility - plausibility as a discipline - who decides - the being-human lens - after 2023, AI’s reach into cognition and formation

Centrality at a glance#

Family 1. Rethinking risk#

This is the oldest and most original layer of his work. It began in nanomaterial and occupational aerosol science (13 years at HSE and NIOSH, FFTF p.118–122; PEN; the NNI environment and health committee) and was turned into a general theory of risk for emerging technologies.

1.1 Risk innovation — core#

• Definition. “parallel innovation in how we think and act on risk” (2016-01-11 thinking-innovatively-about-the-risks-of-tech-innovation). Later: “we need parallel innovation in how we think about and act on risk” (2018-12-13 tech-startups-orphan-risks). Risk tools built since the Industrial Revolution for mining, chemicals and pollution “belong to a different world”; emerging technologies are “new wine” for old wineskins (FFTF p.23). Risk should be approached the way entrepreneurs approach innovation, turning creative ideas into frameworks people want to use (2019-11-01 how-to-build-a-better-brain-machine-interface). He reframes risk from a brake on progress into support for beneficial progress.
• First appearance. 2016-01-11 thinking-innovatively, the popular version of his 2015 Nature Nanotechnology column “Why we need risk innovation”, which is outside the corpus. The institutional line runs from Michigan students (2013) to the ASU Risk Innovation Lab (2015) to the Risk Innovation Nexus (2017–2020) (2020-11-05 risk-innovation-and-the-future; 2026-07-16).
• Main later appearances.
• Book era and early applications: FFTF p.22–23 (“where much of my current work lies”); 2018-09-03 tech-companies-need-a-social-risk-reboot; 2019-11-01; 2020-07-30 life-on-mars; 2020-10-15 the-ethics-of-advanced-brain-machine-interfaces; 2021-09-07 should-we-be-worried-about-elon-musks-tesla-bot (three questions: what is important to you, what orphan risks threaten it, what small steps help).
• 2023, aimed at AI: 2023-04-04 what-are-the-alternatives-to-calling (placed in the governance lineage as “aimed at supporting agile decisions around intangible risks”); 2023-05-31 existential-risks-of-ai; 2023-07-12 regulating-frontier-ai-models; 2023-08-08 thinking-differently-about-ai-and-risk (“perhaps more relevant than ever”); 2023-09-25 building-a-better-futures-tough (an ATT research domain: “Novel theories, models, and approaches to risk”); 2023-10-25 10-million-for-ai-safety-research; 2023-11-15 navigating-orphan-risks; 2023-11-21 ai-and-risk-innovation (the Risk Innovation Planner).
• 2024 to 2026, new settings: 2024-08-25 advanced-technology-transitions-model; 2024-12-17 navigating-the-challenges-and-opportunities-of-advanced-biopreservation-technologies; 2025-05-18 exploring-ai-through-cause-and-effect; 2026-05-03 are-design-principles-for-responsible; 2026-07-04 just-how-good-is-anthropics-fable; 2026-07-16 orphan-risks-frontier-ai-maynard.
• Evolution.
• 2015–16: a critique of regulation “built around previous technologies”, plus the threat-to-value move and a risk “market” of constituents “willing to invest in protecting” what they value.
• 2018–2021: a practical tool for startups and investors (orphan risks, risk landscapes).
• 2023: pointed at AI developers, and offered as the corrective to AI’s slide back into an ethics framing.
• 2024: rebuilt as a tool for navigating technology transitions, and tested with an NSF biotechnology consortium.
• 2025: folded into AI-literacy teaching through the language of care.
• 2026: its first scholarly application to frontier-AI governance.
• Constant throughout: it is pro-innovation, and it treats conventional risk methods as necessary but insufficient.
• Centrality. Core. It is the institutional and intellectual root of most of Family 1.

1.2 Risk as a threat to value (with value vs values, existing and future value, and reciprocal threats) — core#

• Definition. Risk is “a threat to something of importance to an individual, a community, or a business organization” (2018-12-13). Value includes health, wealth and environment, but also “dignity, belonging, identity, belief, even what it means to be human” (FFTF p.23–24). The framing “extends conventional thinking rather than replacing it” (2018-12-13; 2023-11-15). Four refinements:
• Value is not values. Value means worth; values mean right and wrong. Value is “more effectively operationalized in policy and decision-making” and the framing is “agnostic to particular worldviews” (2023-11-21; 2024-08-25; 2024-12-17).
• Existing and future value. Risk balances maintaining existing value against enabling future value, so not adopting a technology can itself be a threat (2016-03-02; 2024-08-25).
• Reciprocal threats. Threatening what others value comes back to threaten you (2018-12-13; the internal/reciprocal distinction in 2024-12-17; “your risk is my risk” in 2026-07-16).
• Effect as value or care. In AI terms, an “effect” is “what we consider to be of value — or what we care for” (2025-05-18).
• First appearance. 2016-01-11 thinking-innovatively.
• Main later appearances.
• 2016–2021: 2016-03-02 how-risky-are-the-world-economic-forums-top-10; 2017-04-10 dear-elon-musk-your-dazzling-mars-plan; FFTF p.23–24 and p.174 (AI risks may “blindside us” because “we’re not thinking creatively enough about how an AI might threaten what’s important to us”); FFTF p.225 and p.262; 2018-09-03; 2019-11-01; 2020-07-30; 2020-11-05 (value includes “equity, agency, and a sense of self”); 2021-09-07.
• 2023–2026: 2023-05-31 (catastrophe as many people losing what is deeply valuable to them); 2023-07-12 (extended to “deeply held beliefs, and even self-identity”); 2024-07-13 ai-choice-engines-sunstein; 2024-08-07 are-humanoid-robots-really-the-future (implicit); 2025-03-30 reimagining-education (the value-creation model of education); 2025-06-01 vibe-coding-moral-panic; 2026-05-03; 2026-07-16.
• Evolution. The definition has hardly changed, but its reach has grown. It started as an argument to startups and funders. By 2020 identity and agency sat explicitly in the list of values. In 2023 it became his definition of AI catastrophe. In 2024 it was the axis of his transitions model. In 2025 it became the ground for his account of why education matters, and merged with care. It is the bridge that lets him treat AI’s effects on identity, cognition and belief as risk questions rather than only ethical ones.
• Centrality. Core. Probably the single most portable idea in his work.

1.3 Risk as socially defined: harm, safety and acceptability — core#

• Definition. Predicting and responding to risk depends on “who considers what a risk (and to whom)” (2016-03-31 considering-ethics-now-before-radically-new-brain-technologies). Harm is defined by social values, so risk is “ultimately a social construct” (2023-11-26 everything-youve-heard-about-ai-risk-is-wrong). “There is no such thing as absolute safety”. Zero risk “is only possible in the absence of change”, and harm is “a social construct, not a technological one” (2024-06-20 ilya-sutskevers-safe-superintelligence-rethink). The question AI labs never ask is who decides what “safe” means. Acceptable safety is set by perception and edge cases, and the bar in the home is “stratospheric” (2024-08-07).
• First appearance. 2016-03-31, with 2016-04-01 will-driving-your-own-car (the public “must have a say”).
• Main later appearances. 2019-03-05 should-we-be-treating-algorithms (no zero risk; discrimination, dignity and liberty count as harms); 2023-11-26; 2024-06-20, which sets out a ladder of safety complexity: bridges, then chemicals and biological agents, then poorly understood emerging technologies, then superintelligence; 2024-08-07; 2026-07-16 (who decides what counts as an AI risk is, in effect, a public question).
• Evolution. Early on it is an argument for public participation. By 2023–24 it is aimed at AI-safety culture, which he sees as treating safety as an engineering property.
• Centrality. Core, especially from 2023 onward.

1.4 Symmetric risk: the risks of not innovating, and of caution itself — core#

• Definition. Lost future value counts. Failing to develop autonomous vehicles, optogenetics or graphene water filters is itself a risk (2016-03-02). Speculation mistaken for fact can cause the “tragedy” of forgoing beneficial technologies (FFTF p.206). Precautionary action has victims of its own (FFTF p.241–244, the Yellowstone thought experiment). He weighs “the risks of going too fast, the risks of not going fast enough”, notes that regulations could themselves be risks, and rejects “zero exposure — as in no AI” as a default (2023-11-26). The paradox of innovation: “Too much blind speed, and you risk losing your way. But too much caution, and you risk achieving nothing” (FFTF p.163; restated in 2025-03-02).
• First appearance. 2016-03-02.
• Main later appearances. 2019-03-05 (not deploying carries risk); 2023-05-31 (the risk of not developing AI counts); 2024-08-25 (missed opportunities in discovery as a threat); 2025-03-30 (the risk of inertia: “sacrificing what could be on the altar of a blind devotion to what is”).
• Centrality. Core. It is why his risk thinking never collapses into precaution.

1.5 Orphan risks and the risk landscape — recurring (a tool within risk innovation)#

• Definition.
• 2018. Risks “too ill-defined, too complex, or too irrelevant” to attend to, which can still derail an enterprise. They are mostly social and rarely governed by law (2018-12-13).
• 2019. A fixed list of 18 used by the ASU Risk Innovation Nexus. The method: start from value to the enterprise, investors, customers and communities; map the landscape; look for clusters that converge into “blindsiding impacts” (2019-11-01).
• 2020–21. Risks “hard to quantify and easy to ignore” that “have a habit of coming back to bite” (2020-11-05; 2021-09-07).
• 2026. Risks with no agreed tools, standards or mitigations and no one accountable for them in practice. They are known and named, but unowned (2026-07-16).
• Precursor. 2016-02-01 we-dont-talk-much-about-nanotechnology-risks-anymore: nano risks that “slip under the radar” because “few people know what questions they should be asking”.
• First named. 2018-12-13, presented as a new term.
• Main later appearances.
• 2019–2021: 2019-01-16 responsible-innovation-entrepreneurs; 2019-03-01 personal-fitness-tracking (consumer data); 2019-11-01; 2020-10-15; 2020-11-05; 2021-09-07 (the Tesla Bot list).
• 2023–2024: 2023-11-15 (the 2018 article re-aimed at AI through Meta’s Galactica); 2023-11-21 (the Planner); 2024-03-21 elon-musks-neuralink-plays-mind-games (the 2019 landscape “still stands up reasonably well”); 2024-12-17 (ATP-Bio, where ethics, perception and regulation dominated).
• 2026: 2026-07-04; 2026-07-10 i-asked-anthropics-fable-5-to-create-a-video-game (orphan risks as things to “adopt”); 2026-07-16; 2026-08-02 what-we-can-learn-with-ai-by-not-trying-to-learn.
• Notably absent from the 2024–26 cognition essays and from the 2026-09-24 lecture.
• 2026 sub-concepts (so far one-off, all in 2026-07-16).
• Four filters. “Can we measure it? Is it big enough? Can we evidence it? And can we afford to keep it?” He traces the first three to a definition of risk as “the probability of a specified, severe harm event”.
• Safety differential. The gap between a firm’s safety and compliance frameworks.
• Severity floor and risk aperture. The severity floor is “a boundary of accountability”. The aperture “has a tendency to travel” from company frameworks into statute.
• Proposed disclosures. An orphan-risk register and an aperture log.
• Conversion channels. The routes by which stakeholder harm becomes cost to the firm “are not equally open to everyone”.
• Values drift inside an incentive field. “sincerity almost always operates inside an incentive field”, so the frameworks are “working as designed. It’s just that the design itself may be flawed.”
• De facto governance layer. Company safety frameworks have become one.
• Evolution. The concept stays the same while its audience changes. It began as an entrepreneurial self-interest tool (2018–2021), was re-aimed at AI developers (2023), and became a public-accountability and disclosure framework for frontier-AI governance (2026). At every stage it sits inside the threat-to-value frame.
• Centrality. Recurring. Most visible in 2019–2021 and mid-2026; never the organising idea.

1.6 Social risk, non-technical hurdles and social licence — recurring#

• Definition. Showing a product is “safe enough” and compliant no longer suffices. Products must not threaten livelihoods, fairness, beliefs or aspirations (2018-09-03). Several related ideas sit here:
• Technical fixes don’t reach the deeper risks. Concrete harms invite technical fixes and mask subtler social and psychological risks that technology cannot solve (2016-03-02).
• Wicked problems. They shift in response to attempts to solve them, and “numeric logic is often trumped by what we intuitively think and feel is important” (2016-03-12 itll-take-more-than-tech-for-elon-musk).
• Social licence. Success depends on whether society “grants” the freedom to proceed (2017-04-10).
• Resistance protects value. Resistance to innovation is usually value-protective, not anti-science (FFTF p.224–226). It is often the “handling”, not the technology, that causes harm (FFTF p.225–226).
• Complex, not complicated. The biggest barriers to humanoid robots are “likely to be societal rather than technological”; acceptance is complex while engineering is “merely complicated” (2024-08-07).
• Centrality. Recurring. It is the applied face of threat-to-value.

1.7 Hazard, exposure and the chemical-risk template — recurring (identity-defining)#

• Definition. His professional method:
• Hazard “taken in isolation” misleads. Assess exposure across use, manufacture and disposal, and weigh net trade-offs (2015-01-10 are-quantum-dot-tvs; 2016-02-01, where the two Vantablacks show the same material low-risk in one use and risky in another).
• Four exposure–hazard questions (can it get in; can it cause harm; by what mechanism; how much). Innovation is irresponsible by process, “irresponsible innovation on a grand scale” even if the risk proves negligible (2021-03-28 how-safe-are-graphene-based-face-masks).
• Terms of art vs terms of science. “nature doesn’t care what we call a material” (2022-02-10 are-we-asking-the-right-standards-questions), which argues for behaviour-based governance of “sophisticated materials”.
• Transfer to AI.
• 2019. Five transferable concepts: hazard vs risk, exposure, consequences, exposure–response, and weight of evidence. He coins “algorithmic exposure” and “algorithmic exposure-response relationship”, while insisting “an algorithm is not a chemical” (2019-03-05).
• 2023. Tests risk = Fn(hazard, exposure) on AI. Exposure might be “hints of ideas encountered over hours of social media use”, and no framework exists yet (2023-11-26).
• 2024. Chemicals become one rung on a ladder of safety complexity (2024-06-20).
• 2026. Evolutionary mismatch with synthetic chemicals and vaccines, used as a structural analogy for AI (2026-01-10).
• Biographical root. Occupational health (FFTF p.118–122):
• technological sophistication does not mean safety;
• harm lands on “the first tier” of workers;
• uncertainty can suit those who profit, as with the mine owners and black lung;
• carbon nanotubes may behave like asbestos.
• Evolution. Literal method (nano) → structural template for algorithms (2019) → exploratory test that found no AI framework (2023) → background analogy (2024–26), alongside growing wariness of analogy (see 5.4).
• Centrality. Recurring. The earliest and most technical layer of his risk identity.

1.8 Risk from first principles; more than probability — recurring#

• Definition. “There’s more to risk than probabilities”: the type and size of harm, who bears it, and unquantifiable losses such as opportunity, hope, agency and dignity (2020-07-30). In 2023-11-26 he breaks risk into five elements:
• cause and effect: “no cause, no risk”, so “AGI going rogue” is a hazard without a causal pathway;
• magnitude;
• harm, defined by social values;
• time;
• perception.

AI takes each element “to a whole new level”. In 2026-07-16 the probability-of-severe-harm definition is diagnosed as the source of how AI companies filter out risks. - Centrality. Recurring.

1.9 Precaution — recurring (as a qualifier and foil, never a creed)#

• Stance. Caution scaled to irreversibility and ignorance. Gene drives call for exceptional caution because “we don’t have the luxury of rebooting” (2016-01-20 three-ways-synthetic-biology). He calls precaution “often misunderstood”, yet endorses the UNESCO COMEST formulation (plausible, morally unacceptable harm; a proportionate response; a participatory process) as “a sound philosophy” (2020-07-30). “Better safe than sorry” is a practical rule under mechanistic uncertainty (2021-03-28).
• Across the AI period.
• 2023. Declines the FLI pause letter (2023-04-04 alternatives) and rejects a zero-exposure default (2023-11-26), but floats “a chance to take a breath (a pause even)” (2023-11-18 sam-altman-openai-impacts).
• 2024. The “Avoid” quadrant as a precaution-like stance (2024-08-18). A narrow, conditional pause on relationship chatbots: “pausing — or even rethinking” their development (2024-10-27).
• 2025. Calls the transatlantic history of precaution “fraught with misunderstanding, misinterpretation” (2025-03-02). Plans “just on the off chance” an edge-case scenario is right (2025-04-06).
• 2026. A small-probability, high-consequence risk justifies questions and research, not restriction (2026-01-10). “We can’t pause it” (2026-09-24).
• Book view. Precautionary action has its own victims (FFTF p.241–244).
• Centrality. Recurring. His position is stable: proportionate, participatory and evidence-weighted, never a default ban.

1.10 Risk perception, expert crowds and blindsides — recurring#

• Definition.
• Imagined publics. Manufacturers act on unevidenced assumptions about what the public thinks (2018-08-22 second-guessing-consumer-views).
• Blind spots and habituation. Humans have a blind spot to sudden catastrophic risk, and a “wow” to “meh” habituation that is adaptive but dangerous (FFTF p.255–256, 284–286).
• Expert crowds. Crowdsourced expert rankings capture a “risk perception zeitgeist” rather than over-the-horizon risk and show “regression to the mean of conventional thinking” (2024-01-14 wef-global-technology-risk-trends; 2024-01-17). By 2025 he says mainstream experts “do not grasp how disruptive the technology may turn out to be”. Disruptions will be “I told you so” moments for insiders and blindsides for everyone else (2025-01-19 wef-global-risks-2025).
• Cultural cognition. People are persuaded by “their sort of person”, a finding he carries over from Kahan’s nanotech work at PEN (2024-09-01).
• Moral panics. They are “rarely cut and dried” and signal threats to what is important to people (2025-06-01). “moral-panic fires” and techlash appear in his self-map (2026-07-10).
• Centrality. Recurring.

1.11 Risk communication: safety message first; the limits of warnings and literacy — recurring (rising 2025–26)#

• Definition. From “decades” in risk assessment, management and communication (2025-11-09 universities-chatgpt-mental-health), the usual tools fail:
• “be careful” warnings;
• bans, which need surveillance;
• guardrails that risk “neutering” models and push students to less safe tools;
• AI-literacy classes that “risk becoming performative”.

He prefers dialogue, listening, trust and safe spaces. 2026-05-10 do-not-do-this-with-ai puts “the safety message first”: benefits are self-evident and risks are not. He offers five plain-language don’ts and five dos, and argues that literacy alone will not change behaviour. - Roots. Risk Bites, launched 2012 (FFTF p.126); the 2018 ten-risk video. - Centrality. Recurring and rising.

1.12 Value-based resilience — recurring (few appearances, conceptually core)#

• Definition. Resilience is not bouncing back but “protecting and preserving what is considered to be “of value.””, including livelihoods, beliefs, identity and self-worth. It is about thriving, not surviving, and resilience talk can reflect a “privileged Western perspective” (FFTF p.261–265; 2018-10-25 environmental-resilience-and-the-day-after-tomorrow). It has the same structure as threat-to-value.
• Later echoes. Networks both build and undermine resilience, and dependence is fragility (2019-11-19, co-written). The “Adapt” quadrant (2024-08-18). Resilience rather than speed in governance (2025-04-24 us-and-china-vie-for-ai-k-12-leadership). Building everyone’s ability to thrive with advanced AI (2025-08-31).
• Centrality. Recurring.

1.13 Comparative risk against a baseline — occasional#

Driving your own car as a future public-health risk comparable to smoking (2016-04-01). Waymo’s crash rates checked against his own reading of the human baseline, and the open question of whether humans are the right metric (2023-11-09 waymo-safety-study-shows-benefits).

Family 2. Responsible innovation, ethics and governance#

2.1 Responsible innovation — core#

• Definition. The book’s core question: how to “reap the benefits of innovation without running into serious problems along the way” (2018-10-12 everything-you-wanted-to-know-about-films-from-the-future). His gloss of RRI: think about consequences early, give everyone affected “a say”, and admit it is “fiendishly difficult” in practice (FFTF p.22). He adopts Stilgoe, Owen and Macnaghten’s anticipation, reflexivity, inclusion and responsiveness as a working framework. He pairs “responsible” with “responsive” (FFTF p.24, 26; 2023-10-19). Caution is not smashing technology (2018-12-15 if-elon-musk-is-a-luddite).
• First appearance. 2015-01-30 responsible-development-of-new-technologies.
• Main later appearances.
• 2016–2019: 2016-01-11 the-fourth-industrial-revolution (“much more is needed”); 2016-04-01 (DOT’s AV policy as a model); 2019-04-15 tech-companies-need-an-ethics-reset; 2019-08-13 (co-written).
• 2023: 2023-05-05 us-white-house-embraces-responsible-innovation (“fiendishly hard to operationalize”; “if we wait until we have a problem … we’ve probably waited too long”); 2023-10-30 white-house-goes-all-in-on-responsible-ai.
• 2024–2026: 2024-03-31 we-have-a-technology-problem-and; 2024-05-21 openais-problem-with-the-movie-her; 2025-02-23 evo-2-dna-ai; 2025-04-06 responsible-innovation-and-ai-acceleration; 2025-07-23 americas-ai-action-plan; 2025-08-31 holding-on-to-our-humanity-age-of-ai; 2026-05-03.
• Evolution.
• 2015–16. Endorses the European framework, but says more is needed.
• 2019. Adapts it for US entrepreneurial culture: responsibility grown within innovator communities (codes of conduct, “creating mutual worth”). Names three structural vulnerabilities: tight coupling, latency and value mismatch (2019-08-13).
• 2023. Calls for anticipatory investment. Says the US has resisted it through “an ethos of go fast, break things” (2023-10-30 white-house-goes-all-in).
• 2024. Names responsible innovation among fields too narrow for today’s transitions (2024-03-31), and sees company “lip service” and “childish irresponsibility” (2024-05-21).
• 2025. Responsible AI is “going out of fashion at lightening speed” [sic] (2025-02-23). It may “seem futile” against AI acceleration, and would fare “not that well” in a US–China race (2025-04-06). The AI Action Plan casts it as a barrier (2025-07-23). It must be augmented by channelling innovation and building people’s capacity to thrive (2025-08-31).
• 2026. Principles work only if actually used (2026-05-03).
• Centrality. Core. A lifelong commitment whose confidence visibly declines after 2024.

2.2 Could vs should — core (the rhetorical touchstone)#

• Definition. Ian Malcolm’s challenge in Jurassic Park. The more complex the science, the more pressing the gap between what can be done and what should be (FFTF p.36–39). Early computing built cyberspace on what “we could do rather than what we should” (FFTF p.144–146). “Should we?” must sit alongside the question of whether we can, for seductive technologies (FFTF p.68–72).
• Appearances. 2018-08-26 jurassic-park-the-rise-of-resurrection-biology; 2019-03-07 sci-fi-movies-are-the-secret-weapon; 2019-04-26 avengers-endgame (technological reach vs collective grasp); 2019-07-23 neuralinks-technology-is-impressive; 2021-09-07; 2022-06-13 jurassic-park-dominion; 2023-08-28 dear-mr-musk-can-i-have-my-book-back; 2025-04-13 de-extinction-conservation-futures (republished, reaffirmed); 2025-05-11 designing-responsible-technological-futures (“What are the technologies we want in our lives?”).
• Centrality. Core, but more slogan than analysis.

2.3 Beyond ethics: operationalising responsibility; ethics vs risk framing — recurring#

• Definition.
• Ethics boards are not enough. Principles and advisory boards are necessary but insufficient, and at worst “smoke-and-mirrors”. What works is standards, measurable expectations, enforceable checks and balances, culture, stakeholder buy-in and long-term training (2019-04-15, drawing on an entrepreneurial-ethics course he has taught since 2013).
• AI ethics has crowded out AI risk. AI ethics is booming while research on AI’s own risks is thin (2021-08-03 we-need-to-get-more-innovative).
• Reframe from ethics to risk. Governance should move from an ethics framing to risk and socially responsible innovation. Ethics sorts right from wrong but gives no practical framework. AI pulled the field back to ethics (the 2017 Asilomar AI Principles) (2023-04-04 alternatives).
• Good intentions are no guarantee. The belief that good intentions will deliver beneficial AI is “sheer fantasy” (2023-12-15 pope-francis-artificial-intelligence).
• Principles-based ethics has limits for advanced assistants (2024-04-24).
• Legal is not the same as good practice (2026-05-03).
• Centrality. Recurring.

2.4 Governance toolbox and lineage: agile, anticipatory, soft law, pacing — core#

• Definition. Regulation is built around previous technologies, and new ones get shoehorned in. His model is flexible, anticipatory, participatory policy that evolves with the technology: “a refreshing change from attempting to retrofit existing regulations” (2016-04-01; 2016-01-11 thinking-innovatively).
• Lineage he claims.
• A 2008 WEF note, and a 2010 Davos proposal with Tim Harper for a Global Centre for Emerging Technology Intelligence (2023-12-22 un-governing-ai-for-humanity).
• WEF’s 2018 definition of agile governance, “adaptive, human-centred, inclusive and sustainable policy-making”, from a council he sat on.
• The genealogy recombinant DNA (bioethics) → Human Genome Project (ELSI) → nano and synthetic biology (agile and anticipatory governance, soft law), driven by the “pacing gap” (2023-04-04 alternatives).
• Positions.
• Hard regulation is “a blunt tool” (2023-04-10 as-ai-goes-to-washington). Rather than an AI-only agency, a cross-agency initiative for advanced technology transitions (2023-05-17 ai-senate-hearing-may-2023).
• A governance spectrum runs from public engagement through soft law and agile governance to hard law (2023-09-25). The pacing problem, after Gary Marchant (2024-12-17).
• Agile post-launch correction for generative-AI failures (2024-02-25). Agile, technology-agnostic education policy (2024-12-13).
• Formulate the governance problem before regulating, and fit oversight to agent capabilities (2025-05-04 an-important-new-model-for-guiding-agentic-ai-oversight).
• Early-2000s nanotechnology governance as a literal model: “balanced, proactive, and above all collaborative” (2025-07-23).
• Hard law where it fits: deepfake criminalisation and developer liability (2024-02-25), and apps designed to exploit cognitive biases “can and should be regulated far more” (2025-08-31).
• Disclosure of how risks are selected (2026-07-16).
• Evolution. A consistent preference for adaptive, multi-stakeholder soft law, with hard law kept for specific harms. Confidence in government declines over time: democratic governments lack “imagination, vision, or agility” (2025-01-07 universities-need-to-step-up); under Trump he expects “a tapestry of soft governance mechanisms” (2024-11-06, written framing in a Modem Futura promo); and by 2026 he sees a gap that universities must help fill (2026-09-24).
• Centrality. Core.

2.5 Technology versus use: doubts about use-based regulation — occasional (2023, significant)#

• Definition. The nanotech-era mantra, regulate what people do with a technology “not the technology itself”, may fail for general-purpose AI, because capability erodes the distinction. Predictive policing is a moral hazard that “cannot be addressed by naively separating the technology from its use”. A “foundational general purpose technology” might be dangerous in itself, and open-sourcing it irreversible. His “current thinking lies between” licensing and open source.
• Sources. 2023-05-17; 2023-05-22 can-large-language-models-be-used; 2023-07-12. Related: technology is not neutral, and risk belongs to sociotechnical systems (2023-12-15; 2024-01-14).
• Centrality. Occasional, but a real departure from his own field’s past consensus.

2.6 Precedents of scientific and community self-governance — recurring#

• Definition. Bottom-up models: Asilomar 1975, the Responsible Nano Code, DIY-bio norms, Debian’s constitution (2019-08-13). Always qualified by Sarewitz’s point that leaving risks and ethics to experts is wrong-headed. The 2015 AI open letter and Asilomar 2017 are read as shoring up against permissionless innovation (2018-10-04 superintelligence). Asilomar 1975 is “a landmark in establishing the foundations of responsible and beneficial genetic manipulation”, and he praises the Evo 2 team’s self-restraint (2025-02-23). He adds that self-governance will not scale as capabilities spread (2022-06-13).
• Centrality. Recurring.

2.7 Care: from duty of care to the “hard” concept of care — recurring (rising since 2025)#

• Definition.
• Duty of care, 2018. Suppliers of implants owe “a continuing duty of care” (FFTF p.150).
• Lifetime responsibility, 2024. Implants carry “a lifetime responsibility to patients” (2024-09-18 neuralink-blindsight). Tech entrants lack the medical sector’s ethos of lifelong patient care (2024-10-02 23andme-in-trouble).

• “Hard” care, 2025 (2025-03-09 the-hard-concept-of-care-in-technology-innovation):

  • care as “something of substance” that can ground policy, governance and decision-making, and “hard” in the sense of difficult;
  • care against control, which reduces agility and responsiveness;
  • care against aggregate thinking, criticising those with “little time for the individuals who make up that aggregate”;
  • Tronto’s “privileged irresponsibility”;
  • technologies as imperfect creations needing stewardship (Frankenstein).

  It draws on Emma Frow and Alison Gopnik. The detailed closing definition in that post is Deep Research’s, not his. - Care in teaching, 2025. “a culture and a mindset of care” (2025-05-18). - Duty of care for institutions, 2025. Universities that supply ChatGPT and encourage its use take on “a social, moral and (I would assume) legal duty of care” (2025-11-09). - Later traces. “Care” is one of the book’s four postures (co-written, 2025-10-14). It appears in the students’ skills list (2026-04-14) and in formation, “care and effort in its formation” (2026-07-04). - Centrality. Recurring. A named 2025 development with roots in 2018.

Family 3. Innovators, hubris and the ethics of permission#

3.1 Myopically benevolent science; a non-demonising account of developers — core#

• Definition. Scientists and technologists pursue their work sincerely, justified by a vague, untested idea of social good, and never ask the people they claim to serve. Stratton “never bothered to ask anyone else” (FFTF p.21). He applies this to himself, “even myself at times” (FFTF p.220). He admits to coopting a social-good narrative in his own 2000 aerosol paper (FFTF p.219). There is a general benevolence that gives way to self-interest in specific situations (FFTF p.220–221). Developers “care deeply”, “they’re not fools”, and they fail through absorption, “their version of ‘responsible’”, and what they “don’t care to find out” (FFTF p.17, 36–38). The remedy: responsible science is good intentions plus humility plus other people’s expertise (FFTF p.39).
• Appearances. 2018-10-11 a-nanoscale-perspective-on-the-man-in-the-white-suit; 2023-09-08 living-in-a-material-world (taught every year); 2023-10-02 responsible-ai-lessons-from-nanotechnology (AI “driven by a small group of experts and companies” who believe they understand enough); 2025-04-13 (reaffirmed).
• Centrality. Core. His basic model of developer psychology: epistemic and structural, not villainous.

3.2 Hubris, technologies of hubris, and humility — core#

• Definition. Hubris is excessive self-confidence that lets one ignore obstacles (FFTF p.166). It is double-edged: it drives visionary leaps and carries “a massive risk of unintended consequences”. The type case is Nathan, “tech-savvy, but socially ignorant” (FFTF p.163). Humility is the antidote, “But humility alone isn’t enough” without plausibility and other people (FFTF p.167–168). Hubris can also be collective and civilisational, as in the Anthropocene (FFTF p.260).
• Appearances.
• Republications: 2023-04-16 ai-and-the-art-of-manipulation; 2025-03-02 the-lure-of-permissionless-innovation.
• 2023–2024: 2023-11-26, where he calls his own title “a little hubristic”; 2024-06-30 seth-is-conscious-ai-possible, the “hubristic” go-fast culture of OpenAI and Anthropic.
• 2025: 2025-03-09 (too arrogant to see what sustains well-being); 2025-04-13 (Colossal’s hubris is “palpable” and AI companies are “full of themselves”).
• 2026: 2026-03-29 can-ai-create-an-undergraduate-degree-plan (academic hubris); 2026-04-11 ten-questions-about-ai-and-higher. Here hubris runs both ways: naive adoption, and “trusting that traditional mastery alone is enough”.
• Evolution. It starts as a critique of innovators and by 2026 applies symmetrically to institutions that resist AI.
• Centrality. Core.

3.3 Permissionless innovation (critiqued) and the reversibility test — core#

• Definition. “innovation that is conducted in the absence of permission from anyone it might impact” (FFTF p.159), after Thierer’s default-to-permit doctrine. His distinctive refinement: “permissionless innovation isn’t necessarily reckless innovation”. It is responsibility judged by the innovator alone, and “a single innovator cannot see the broader context” (FFTF p.162). So checks and balances are needed on “who gets to do what” where impacts are widespread or irreversible. He admits its lure: his own PhD rule-bending (FFTF p.161) and exhilaration at Musk’s Mars plans.
• First appearance. 2018-10-04 (FFTF Ch. 8).
• Main later appearances.
• Scare-quoted as innovation driven by “power, wealth and a lack of accountability”: 2018-12-15; 2019-01-16.
• 2023: 2023-04-16 (republished, “probably more so now”); 2023-05-12 unraveling-the-luddite-narrative (the “Luddite” slur reflects a move-fast worldview); 2023-08-25 ai-platos-cave; 2023-10-19 marc-andreessen-ditch-sustainability.
• 2025: 2025-02-23 (it is displacing responsible AI, in “go fast and break things in the hope that someone else will clean up the mess”); 2025-03-02 (“more relevant now than it was then”); 2025-07-23 (the Action Plan’s “try-first” culture).
• Evolution.
• A pejorative label (2018–19).
• Then a nuanced critique that concedes its benefits (the book).
• Then an operational test (2025-03-02 footnotes): experimenting in low-risk, reversible, linear systems is fine; breaking complex systems, “people, governance, society, and the planet”, is not. This also reconciles his support for student “permission to play” (2025-03-15).
• Then a diagnosis that it is winning politically (2025).
• Public indifference hands de facto permission to “those that do care” (FFTF p.286).
• Centrality. Core.

3.4 Mega-entrepreneurs, moral certitude and the Silicon Valley playbook — recurring#

• Definition.
• The book. “the sometimes oversized roles mega-entrepreneurs play in dictating how new tech is used, and possibly abused”, and “the folly of entrepreneurial arrogance” (FFTF p.26–31). “a new type of entrepreneur” with outsized resources (FFTF p.161). Powerful people with moral certitude and the means to act: “Where do they get the right to act unilaterally on issues that ultimately impact us all?” (FFTF p.249).
• Posts.
  • Musk and Schwab are visionary but socially naive (2016-01-11 fourth-industrial-revolution; 2016-03-12; 2017-04-10). Neuralink hired with “No neuroscience experience is required” (2024-01-30, citing FFTF p.128).
  • Neuralink’s “cult of personality” (2024-01-30). An even-handed reading of the “reality distortion field” and “fail fast” playbook, which critics discount at their peril (2024-03-21).
  • Musk matters for “not how wild the ideas are, but how much influence they have” (2024-08-04 7-key-takeaways-from-elon-musk-and-lex-fridman). The Sci Fi feedback loop and a “savior complex” (2024-09-18). Altman’s “childish irresponsibility” (2024-05-21).
  • Andreessen’s technological foreshortening (2023-10-19). Altman on climate criticised, Amodei read generously (2024-10-06; 2024-10-13).
• Changes of mind. Musk revised from humble AI-risk voice to naive permissionless innovator (DOGE) (2025-03-02). Neuralink treated more generously in 2024-03-21 than in 2019–2020.
• Centrality. Recurring.

3.5 Immoral logic, speculative fear and the right to act unilaterally — recurring (book-rooted)#

• Definition. “Good intentions lead to seemingly logical, but not necessarily moral, actions” (FFTF p.231–240). The “conceits” behind action on technological prediction: exponential trends, amplified biases, “artificial certainty”, ideology over responsibility, and the “enlightened” acting without consent (FFTF p.240–241). Speculative expert fear (Joy, gray goo) fed the ITS bombings of nanotechnologists (2018-11-15 even-bad-sci-fi-movies). Musk’s “truth-seeking” AGI: ultimate social truth is usually “an excuse to impose an ideology” (2023-07-19 elon-musk-maximally-curious-agi).
• Appearances. 2018-10-18 the-honest-broker-meets-dan-browns-inferno; 2018-11-15; 2023-07-19.
• Centrality. Recurring in the book, occasional in posts. It is portable: his critique applies equally to utopian and to catastrophic AI futures claims.

3.6 Industry influence, lip service and the “good intentions” fantasy — recurring (2023 onward)#

• Definition.
• Capture and influence. Lawmakers are captured by executives and VCs (2023-04-10). Industry leaders asking for regulation are an “atomic technologies moment”, with a possibly cynical reading of an exclusionary playing field (2023-05-17). Firms have shaped regulation in ways that seemingly favour leaders, and Altman was policymakers’ “shiny person” (2023-11-18; 2023-10-30 white-house-goes-all-in).
• Intentions and gaps. Good intentions are “sheer fantasy” (2023-12-15). There is a gap between responsible-innovation talk and conduct (2024-05-21). “sincerity almost always operates inside an incentive field” (2026-07-16).
• Developers’ self-presentation. They act “as if they’re the first people to notice” these risks (2026-09-15 will-ai-really-kill-us-all). They lack the breadth “to decide for humanity” (2026-09-24).
• Balance. Enthusiastic partnership sits alongside the criticism: ASU–OpenAI (2024-01-18); praise for Waymo, the Evo 2 team and OpenAI’s system cards (2024-09-01); respectful interest in Anthropic’s constitution (2026-01-22).
• Centrality. Recurring.

Family 4. Complexity, transitions and the shape of the future#

4.1 Advanced technology transitions (ATT) — core (his umbrella frame since 2023)#

• Definition.
• A transdisciplinary research focus, a domain of thought leadership, an area of education and a platform for mobilising expertise (2023-04-12 navigating-advanced-technology-transitions).
• Technology transitions are the “theories, frameworks, and practices” that support beneficial development, adoption and use. “Advanced” marks transformative, converging, non-linear technologies (2023-09-25).
• Such transitions are “not navigable through conventional thinking” and need “pilots” (2024-08-11 school-of-advanced-technology-transitions; 2024-03-31).
• In his own prompts: “new models, frameworks etc” for “transformative, convergent and synergistic technologies” (2025-02-04 openai-deep-research-ai-scholarship).
• In his self-map: “human flourishing and navigating complex advanced technology transitions” (2026-07-10).
• Precursors. 2015-01-30 (converging technologies as a “socio-techno-environmental system”); 2016-01-11 fourth-industrial-revolution (the capability–responsibility gap, which “has continued to widen”); FFTF convergence; 2021-04-09 bounded-infinities (the solution problem).
• Main appearances.
• 2023: 2023-04-04 alternatives; 2023-04-10; 2023-04-18 universities-need-to-be-investing; 2023-05-17 (a cross-agency initiative); 2023-09-25 (the NSF research agenda); 2023-10-30 white-house-goes-all-in; 2023-11-29 the-year-that-generative-ai-changed-the-world.
• 2024: 2024-03-10 harari-teaching-future; 2024-03-31; 2024-08-11; 2024-08-18 (the Avoid/Adapt/Extend/Embrace model); 2024-08-25 (the threat/opportunity model); 2024-10-13 amodei-machines-of-loving-grace.
• 2025–2026: 2025-01-07 universities-need-to-step-up-their-agi-game (three foci); 2025-03-27; 2026-05-21 magnifica-humanitas-and-being-human; 2026-08-30 do-universities-have-a-place-in-bill; 2026-09-24.
• Unresolved tension. Past lessons are neglected; each wave tends to “re-invent the wheel”, and AI advocates are “blissfully unaware” of earlier transitions (2023-04-12). Yet the present is “unlike anything we’ve had to grapple with before” and “night and day different” (2024-08-11). See 5.4.
• Evolution. Label (April 2023) → policy research agenda (September 2023) → a field needing its own school, pilots and tools (2024) → the frame for universities’ role and for the “who we are” domain (2025–26).
• Centrality. Core from 2023.

4.2 Convergence and base code (bits, bases, atoms; cross-coding and transcoding) — core to 2021, then background#

• Definition. Convergence is “what happens when different strands of innovation intertwine together” (FFTF p.18). It is where transformative power and the greatest pitfalls both lie, and capability outruns understanding (FFTF p.20–21). It brings “multiplicative dangers” (FFTF p.281). Growing mastery of three base codes (bits, DNA bases, atoms) and “cross-coding” between them raises the chance of “serious and irreversible mistakes” (FFTF p.183–189). The risk is “bricking” the world with no re-install option (2021-02-25 how-our-mastery-of-biological-physical-and-cyber-base-code). Intelligent machines merge all three codes (2021-04-09).
• Appearances.
• 2015–2020: 2015-01-30; 2018-12-15 (“for the first time in human history”); 2020-10-15 (the “synergistic scaling of ability, accessibility, and use” as the real turning point).
• 2021: 2021-02-25 (extended to a social base code of norms, behaviours and even ideas); 2021-04-09.
• 2023–2025: 2023-01-31 (convergence and emergent properties in his self-description); 2024-01-01 the-future-of-being-human-in-2024 (the base code of being human, including language); 2024-08-04 (Musk’s “synergistic tech integration”); 2025-02-23 (DNA as language); 2025-04-13 (AI plus gene editing).
• Evolution. Engine of transformation (2015–18) → base-code framework (2018–21) → social base code (2021) → the base code of being human (2024) → language as the common substrate (2025). The through-line is the path from material convergence to cognitive and identity concerns.
• Centrality. Core in 2015–2021, recurring after.

4.3 Complexity, bounded unpredictability and normal accidents — core#

• Definition. “Complex systems behave in unpredictable ways” and perfect control is impossible. But chaos is bounded, which separates “plausible futures from sheer fantasy”, and points of stability make good futures reachable, if they are not squandered (FFTF p.39–42). Perrow’s normal accidents, with the Arkema cascade in Hurricane Harvey as his real chemical-industry illustration (FFTF p.42–43; 2021-10-05 butterflies-chaos-theory).
• Related ideas.
• The grid blackout: systems look stable “until, suddenly, it isn’t” (2015-01-30).
• Tight coupling, latency and value mismatch (2019-08-13).
• Non-linear catastrophic failure from overlooked events (2020-11-05).
• Six cause–effect models for anticipating AI’s consequences: linear, S-curve, exponential, hysteresis, jagged, chaotic (2025-05-18).
• A fictional cascade: “a seemingly insignificant chain of bad decisions by AI agents cascaded into global systemic failure”, the Great AI Reset (2025-11-24 part-1-letters, his story; the views belong to characters).
• Centrality. Core.

4.4 Tipping points, broken symmetries and early warnings — core#

• Definition. Pippard’s ladder: complex, tightly coupled systems change suddenly and irreversibly at unpredictable points. The past “may not adequately predict what will happen in the future”, so the argument that past technologies turned out fine is not safe. LLMs might be “a turn of Pippard’s ladder”, but “It’s hard to tell” (2023-05-04 tipping-points-and-broken-symmetries). Future Rising (2020): “spot early warnings and stay clear of critical tipping points” (2024-09-08).
• Appearances.
• Book era: a machine-learning and NLP tipping point (FFTF p.170).
• 2023: a “tipping point in human history”, with disruption compressed “from years to months” (2023-09-25). Growing non-linearity between technological cause and societal effect (2023-10-19). “suck it and see” experimentation passes hard-to-spot tipping points, and ChatGPT’s launch did (2023-11-29).
• 2024: a tipping point where technologies change “who we are” (2024-01-01). The four-quadrant model (2024-08-18).
• 2025: tipping points in AI 2027 (2025-04-06). Hysteresis and chaotic effects, with emotional dependency on companion bots as his paradigm case (2025-05-18).
• Change of mind. His original use of Pippard’s ladder was only a warning. In 2024-08-18 he entertains an “Embrace” quadrant, going through tipping points deliberately, while admitting he cannot yet picture it.
• Centrality. Core.

4.5 Speed, irreversibility and the timescale mismatch — core#

• Definition.
• Rising irreversibility. Across eras (pre-industrial, industrial, then “the nuclear and digital age”, globalisation and warming), consequences spread faster than they can be contained: “playing with fire in a world made of kindling” (FFTF p.166–167; 2023-04-16; 2025-03-02).
• The solution problem. Every solution seeds the next problem, and consequences now outpace fixes (2021-04-09). Latency: harms surface more slowly than innovation cycles (2019-08-13).
• Gaps. The pacing gap (2023-04-04 alternatives). The widening capability–responsibility gap (2016-01-11 fourth-industrial-revolution). Early neglect locks technologies into failure-prone trajectories (his 2015 Nature Nanotechnology claim, restated in 2019-08-13).
• Irreversible missteps. “small missteps can have outsized and irreversible consequences” (2023-10-08 a-guide-to-responsible-innovation).
• The timescale mismatch. Responsible innovation processes are “constrained by human timescales” (years) while AI may move in months, so current efforts may “seem futile” (2025-04-06).
• Later forms. The validation gap, in which AI generates knowledge faster than we can validate it (2026-06-12). The co-written book foreword’s closing window before lock-in (2025-10-19, weaker evidence).
• Centrality. Core.

4.6 Inevitability: guide and steer, not stop — core (strengthens over time)#

• Definition. Innovation cannot be switched off: “manage and channel the inevitability of technological change” (2024-08-18). “Technology is not deterministic”, yet we cannot stop innovating, only guide and steer (2025-03-30). Later images: “The AI genie is out of the bottle”, and innovation is like a flood that can be channelled, not halted (2025-08-31); “the boat has already left the harbor” (2026-05-21); “We can’t pause it” (2026-09-24, flagged as a possibly flawed working assumption, paired with criticism of “if we don’t go fast, somebody else will” race logic).
• Roots. The obligation to innovate (FFTF p.288). Embryo screening as a natural experiment: “we cannot put the genie back in the bottle” (2024-04-14). Irreversible integration of AI (2024-07-21 artificial-intelligence-dune-villeneuve).
• Centrality. Core.

4.7 The “fix” frame and AI solutionism — recurring#

• Definition. Framing the world as problems for technology to fix. If AI can fix anything, that includes “fixing” people (2024-10-06 the-double-or-nothing-bet-on-ai-fixing-the-climate; 2024-09-18). Who decides what is “normal” and what needs fixing (2024-10-13)? Other forms:
• the energy “double or nothing” bet;
• the pencil-and-paper test of whether hype solves problems unsolvable otherwise, or whether we “manufacture problems” (2024-12-29 fantasy-top-ten-lists-2025);
• most current problems have roots in “previous good ideas” (2023-10-19).
• Counterweight. He also rejects ideological vetoes of technical fixes. On geoengineering he takes a “yes and” stance: treat the patient and change behaviour (FFTF p.267–269).
• Centrality. Recurring.

4.8 Technological foreshortening and the granularity of impacts — one-off (sharp)#

Andreessen’s collapse of history into smooth upward trends, which hides “the pain and suffering in the detail” and raises “who decides who will suffer and who will thrive” (2023-10-19). It is echoed in the 2025 care-versus-aggregate argument (2.7).

Family 5. Epistemics of the future: plausibility, prediction and analogy#

5.1 Plausible vs imaginable (disciplined imagination) — core#

• Definition. “what is plausible, rather than simply imaginable, is vitally important” (FFTF p.171). Bostrom’s superintelligence is “currently scientifically implausible” (FFTF p.171). Speculation causes harm “when make-believe is treated as plausible reality” (FFTF p.205), through violence but also through policy, investment and forgone benefits. Let imagination sketch futures, then rein it in by what is feasible and ethical: “grounded in plausibility rather than hyperbole”, which he calls a technique of his own (2024-11-17 navigating-the-ethical-dilemmas-of-brain-computer-interfaces).
• Appearances.
• 2018–2019: 2018-10-04; 2018-11-15; 2019-07-23 (informed thinking about plausible issues, not “paralysis by analysis”).
• 2021–2024: 2021-10-05 (bounds separate plausible from fantasy); 2023-04-16; 2023-04-24 ai-risks-primer; 2024-04-28 beyond-the-future-of-humanity-institute (physical plausibility vs “philosophical elegance”).
• 2025–2026: 2025-04-06 (edge-case stress tests); 2026-09-24 (informed speculation about “possible futures rather than real futures”, with humility).
• Centrality. Core. His most consistent epistemic discipline, applied equally to hype and to doom.

5.2 Occam’s Razor for futures and risk priorities — occasional (the 2018 baseline)#

Scenarios that need more untested assumptions are less likely. Gray goo and superintelligence rest on “a house-of-cards stack of assumptions”, and backing them over the evidence-based harms of new materials is “more an act of faith than of reason”. But the probability is “not a zero probability”, and the Razor is only an aid to decision-making (2018-11-01 contact-occams-razor; FFTF p.279–281).

5.3 Exponential extrapolation, later exponential blindness — recurring (with a reversal)#

• Definition, 2018. Exponential predictions are beguiling but fragile. Moore’s Law is a self-fulfilling roadmap. A petri-dish extrapolation can be “mathematically reasonable, but it’s practically nonsensical” (FFTF p.199–202; 2018-11-15).
• Reversal, 2025. He keeps the physical caveat that resource limits halt exponentials. But the exponential becomes a diagnosis of human blindness: “we are really bad at wrapping our heads around rapid exponential growth”; with Bartlett’s beaker, “one minute to midnight”; and “it always will feel like an intellectual exercise until it’s too late” (2025-04-06).
• Trajectory thinking. Three s-curves: capability, utilisation and perception (2024-12-13 are-educators-falling-behind-the-ai-curve). Three trajectories: an AI winter, an S-curve (his most likely case) or an exponential. Even the tail “demands new ways of thinking” (2025-03-30).
• Centrality. Recurring.

5.4 Analogy and its limits: learning from past transitions vs AI’s novelty — core (a key tension)#

• Using the past. Past technologies are used mainly structurally, as lessons about process, engagement and governance, not hazard:
• GMOs are “a masterclass in how naivety, hubris, greed” and non-engagement block progress, and nano’s early engagement dodged a bullet (2023-05-15 erik-schmidt-ai-regulation; 2023-10-02);
• rDNA and Asilomar;
• nuclear: AI is “hidden, dispersed, readily accessible”, unlike contained nuclear programmes (2023-07-25 oppenheimer-and-ai);
• chemicals (2019-03-05; 2023-11-26);
• the phrenology–eugenics–ML lineage (FFTF Ch. 4).

He says to look back “at what’s already known” (2024-12-01 geoengineering-aerosol-monitoring-john-aitken), and complains that each wave tends to “re-invent the wheel” (2023-04-12). - Doubting analogies. - The calculator, internet, printing press and industrial revolution comparisons fail to capture AI’s “sheer uniqueness and profundity” (2024-05-05 blackberry-or-iphone-educational-ai). - Frontier AI “defies analogy”. It is not “calculators on steroids”, search engines, stochastic parrots or simulacra: “Rather, they are different”, and governance must “move beyond easy analogy” (2026-01-22 think-you-know-ai-think-again). - Past frameworks yield “categorical errors” (2026-09-24). - His working resolution. Analogies are mindsets and structural lessons, not templates. He keeps using them (the Trojan horse; the chemical and vaccine mismatch in 2026-01-10) while denying that any one captures AI. - Centrality. Core. Essential for any later use of his work as a lens.

5.5 Epistemic humility, uncertainty and weight of evidence — core (method)#

• Definition. The weight of evidence matters, with no cherry-picking. “we cannot predict the future”, and new technologies add to uncertainty (FFTF p.241–249). He warns against knee-jerk reactions to single studies (2019-03-05). The novelty of AI risk creates an “understanding-vacuum” filled with “dogmatic overconfidence”. His remedy is to drop egos and collaborate across disciplines (2023-11-26). A related “insights vacuum” is filled by “loud (but not necessarily informed) voices” (2023-04-10). His recurring hedge: “I freely admit that I may be wrong” (FFTF p.170).
• Centrality. Core.

5.6 Fallible, slow-correcting and non-neutral science — recurring#

• Definition. Science self-corrects, but it can take “decades or centuries”, and until then the choice of data and interpretation is shaped by human bias, especially in small, noisy fields amplified by hype (FFTF p.68, 75–76; 2020-01-01 the-science-of-predicting-bad-behavior; 2020-09-26). Scientists are believers too. Belief grows from evolved heuristics, and science is distinctive for seeking evidence for belief. Critical thinking alone is “almost inhuman”, and creativity alone leads to fantasy (FFTF p.276–282).
• Centrality. Recurring.

Family 6. Power, justice and who decides#

6.1 Who decides: no abdication to experts; everyone a stakeholder — core#

• Definition. Leaving technology questions to scientists, innovators and politicians is “an abdication of responsibility” (FFTF p.288; 2018-11-08 films-from-the-future-the-last-chapter; 2019-03-06). When science touches everyone, “we all become stakeholders” (FFTF p.282). Publics have “a pretty high level of expertise in what’s important to them” (FFTF p.222). Scientists and technologists “don’t have a monopoly on expertise” (FFTF p.222).
• Appearances.
• 2016–2019: 2016-01-11 fourth-industrial-revolution (ordinary people); 2016-03-31 (the public must have a say in defining risk); 2019-04-26 (technologies “tacitly sanctioned by millions”).
• 2023: 2023-05-15 (the “leave it to the technical experts” mentality “never plays out well”; everyone has the right to play some role; publics need not understand AI’s workings to judge what it threatens); 2023-05-25 leading-ai-expert-says-we-should (alignment as a question of whose values and who decides); 2023-09-04 why-public-engagement-is-so-important (insiders’ “too complicated” claim as hubris in a democracy); 2023-11-29 (“cannot be left solely to AI experts”); 2023-12-22.
• 2024: who decides what “safe” means (2024-06-20); what is good for society (2024-09-01); what is “normal” (2024-10-13).
• 2025–2026: 2025-05-18 (everyone affected has a stake); 2025-08-31 (responsibility is not just for companies or policymakers); 2026-07-16 (what counts as an AI risk is a public question).
• Evolution. The commitment is constant. Over 2025–26 his answer to who acts shifts toward institutions. Companies, governments, civil society and publics each fall short, and universities should act as “an accelerator and a catalyst” (2026-09-24; 2025-01-07 universities-need-to-step-up).
• Centrality. Core.

6.2 Public engagement: two-way, relational, not deficit-model — core#

• Definition. “It’s good to talk”: non-engagement is “an extremely high-risk strategy”, unethical and epistemically poor, and there is “a moral imperative to engage” when impacts are significant, which limits developer autonomy (FFTF p.226–227). Engagement mechanisms include Danish consensus conferences, ECAST, museums, science cafés and citizen science (FFTF p.228–229).
• In posts.
• Citizen-led science that empowers communities, against the elitist view that nonscientists should “revere, but not interfere” (2016-01-12 can-citizen-science-empower). Trust earned through trustworthiness (2020-12-15 why-trustworthiness-matters).
• “Quiet voices”: governance and justice scholars, developers buried in companies, affected communities, ordinary people with extraordinary insights (2023-04-10). “intentional two-way engagement with key communities” through Participatory Technology Assessment, Public Interest Technology and NISE Net (2023-09-04).
• Engagement attached to every pathway in his transitions model (2024-08-25).
• The deficit model, “debunked decades ago” (2024-10-13) and “repeatedly shown not to be effective”. Relational and parasocial communication as a way for experts to combine reach with impact (2025-05-25 why-parasocial-communication-is-important). The parasocial self-correction (2025-11-19).
• Centrality. Core.

6.3 Technology, inequity and justice — core#

• Definition.
• Power and privilege. Powerful technologies in few hands amplify power, and “powerful technologies come with great social responsibility” (FFTF p.110–118). Privileged technology and the trickle-down fallacy (FFTF p.100–103).
• Workers. The disposable workforce and first-tier exposure (FFTF p.118–122).
• Luddites reclaimed. “Neo-Luddite” is “usually a term of derision and censorship”. The Luddites were “not fighting against technology, but against its socially discriminatory and unjust use”, and responsible innovators share their spirit (FFTF p.191–193; 2018-10-08 enter-the-neo-luddites-transcendence; 2023-05-12).
• Appearances.
• 2016–2020: 2016-01-11 thinking-innovatively; 2018-09-19; 2019-03-31 design-principles-for-de-marginalizing-the-future (“The future is designed by the powerful”); 2020-11-12 (benefits for everyone, “not just a privileged few”).
• 2023: 2023-07-27 chatgpt-and-college-applications (AI as translator and equaliser, conditional on closing the digital divide); 2023-10-19; 2023-12-22 (the “AI divide”).
• 2025–2026: 2025-03-09 (privileged irresponsibility); 2025-03-30 (“intelligence is free” as democratising but “deeply contentious”); 2026-03-08 ai-linkedinification (people at the edge of convention most exposed); 2026-07-16 (unequal conversion channels); 2026-09-24 (“Somebody is paying somewhere”).
• Centrality. Core.

6.4 Whose future? The power to imagine and build futures — recurring#

• Definition. Humans as architects of the future whose abilities can “rob others of the futures they aspire to” (2020-12-10 we-need-to-rethink-our-relationship-with-the-future). Future Rising’s stolen futures (2024-09-08). The danger of tech visionaries imposing futures most people do not want (2021-09-07; 2022-02-12 scarlett-johanssons-amazon-alexa; 2022-06-13). Aim for the future we want, not one someone else decides (FFTF p.288). Who will write “the stories that determine our beliefs” (2024-09-22)? The task is “not to abdicate power” but to use it responsibly (FFTF p.44; 2025-04-13).
• Centrality. Recurring.

6.5 Dependency and “who owns you” — recurring#

• Definition.
• Implants. Implant users depend on suppliers for upgrades and security, so “they own you, or at least your health and security” (FFTF p.150). He calls this the “technological equivalent of indentured servitude” (2020-08-28 navigating-the-complex-world-of-advanced-brain-machine-interfaces; 2020-10-15; 2024-03-21).
• Data and inference. Super-consumers under competitive pressure (2022-02-12). Intimate data outliving companies, guarantees and regulations (2024-10-02). ChatGPT memory as an informant that infers by “joining the dots” (2025-10-05 when-chatgpt-turns-snitch).
• Concentration and systemic dependence. Unelected billionaires and concentration (2024-01-17 ai-global-risks-2024-wef-davos). Dependence on AI and the risk of a market crash (2025-11-12 is-planetary-health-without-ai-moribund).
• Centrality. Recurring.

Family 7. What AI is#

7.1 From one converging strand to a categorically different technology — core (his principal AI arc)#

• 2008–2021. In 2008 “AI wasn’t even on my radar” (FFTF p.168); his AI thinking grew out of nanotech debates. AI is one member of a converging set (2015-01-30; 2016-03-02; FFTF p.19–20), though he flags an NLP “tipping point” (FFTF p.170). Intelligent machines merge base codes (2021-04-09).
• 2023. AI sits among technologies at a tipping point, “capable of modulating the very foundations of our understanding of self, sentience, and consciousness” (2023-04-04 welcome-to-the-future-of-being-human).
• 2025.
• Models simulate “the ability to think, to reason, and to solve problems with agency” and “stand apart from pretty much any previous technology”. Treating them as learning aids is “a categorical error” (2025-03-15 ai-playgrounds-in-higher-education).
• Conditional: if AI is a general accelerant and identity-changer, it belongs in “a fundamentally different category to every previous technology” (2025-03-30).
• “arguably the first technology” able to emulate uniquely human attributes (2025-11-30 postscript-letters).
• 2026.
• “Rather, they are different” (2026-01-22).
• “the first technology of it’s kind” [sic] able to slip into the mind (2026-05-10).
• The one domain, “who we are”, where no other technology “has come close to” AI (2026-05-21).
• Unlike “any other technology in human history” (2026-09-24).
• Centrality. Core.

7.2 Intelligence as a “term of convenience”; bounded optimality; the brain is not a computer — recurring#

• Definition. “there is no absolute definition of intelligence. It’s a term of convenience” (FFTF p.171). He prefers Russell’s bounded optimality, the best average decisions within constraints (FFTF p.173). Our obsession with intelligence treats it as shorthand for what makes us different. Human exceptionalism is “an evolutionary illusion”, and “Being smart doesn’t make you good”. Warped definitions will yield warped AI (FFTF p.104–108; 2018-09-13 limitless-pharmaceutically-enhanced-intelligence). The brain-as-computer analogy is rejected (FFTF p.78–79, 95).
• Later. 2023-04-16 (republished). AGI and superintelligence are “ill-defined” (2026-04-11). Intelligence scarcity as a social institution (11.5).
• Centrality. Recurring.

7.3 Emulation without understanding — recurring (2024–26), in tension with rising capability#

• Definition.
• “counterfeit” minds that give the illusion of thought “in a very human way” (2024-03-03 dune-part-two-artificial-intelligence).
• AI “has no intrinsic understanding of the world” and “doesn’t know what it doesn’t know” (2024-09-22 five-ai-generated-podcast-episodes).
• “a generator of ideas, not an understander and implementer of ideas” (2024-10-08 ai-captures-this-years-nobel-prize).
• Mimicry with “no comprehension of what it means to be a person” (his prompt in 2024-09-28).
• A “simulated understanding of the world” (2025-01-26). Evo 2 as “a DNA-based stochastic parrot” (2025-02-23). “blind emulation” (2025-03-30).
• Embodiment and qualia (2025-09-21 a-kindergarteners-guide-to-ai-qualia).
• AI emulates “the processes through which new knowledge is constructed and tested” (2026-06-12). A disembodied AI cannot know what reading feels like to a “flesh and blood human” (2026-07-04; 2026-07-19). Displayed reasoning as performance (2026-09-20 reasoning-llms-just-want-to-have-fun).
• Tension. Capability is a real step change: “most definitely not the generative AI of 2022” (2025-02-09). By 2026 he rejects the “stochastic parrots” label (2026-01-22).
• Centrality. Recurring.

7.4 AI consciousness, personhood and moral status — recurring#

• Definition. We should “transcend the notion of ‘human’” and focus on rights and worth. What natural authority do we have over our creations, including AI (FFTF p.58–61)? He is sceptical of moral machines. The way forward is not “making moral machines” but partnership, with the hope of “artificial emissaries” (FFTF p.177–178).
• 2023. Against “dehumanizing” AIs to justify control (2023-08-18 being-human-in-an-augmented-future). He endorses computational functionalism (“stands up to scrutiny”). He names the “economic expediency of denying consciousness” and calls for norms and policy, “more than just research” (2023-08-23 could-we-build-conscious-ais-in-the-future). Personhood beyond biological “wetware” (2023-09-28 the-creator-and-being-human).
• 2024–2026.
• Personhood beyond human exclusivity (2024-06-16).
• With Seth, real machine consciousness is probably unreachable, and his “thermodynamics argument” returns. The pressing problem is AI that seems conscious (2024-06-30).
• Human-likeness as a questionable gold standard (2024-09-29).
• “human-adjacent” values (2025-11-25, fiction).
• A moral character “at once deeply human and deeply alien” (2026-01-22).
• Evolution. Open to AI rights (2018) → near-term plausibility within a week (August 2023) → back toward biological naturalism, with attention shifting to seeming consciousness (2024) → alienness (2026).
• Centrality. Recurring.

7.5 Superintelligence and AGI: agnosticism, plausibility and the thermodynamic doubt — core (as a stance)#

• Definition. “something of an agnostic”. He attended Asilomar 2017 and says “I was at a scientific meeting, not a religious convention”. He rejects “a very human idea” that narrow intelligence plus power means domination (FFTF p.168–171; 2018-10-04; 2018-11-25).
• Appearances.
• Machines “may one day surpass human intelligence” (2020-12-10).
• “Not a fan” of the superintelligence hypothesis, but he accepts “an exceptionally powerful autonomous artificial entity” as plausible and worth red-teaming (2023-05-25).
• A physicist’s critique of Bostrom and FHI ideology (2024-04-28). Thermodynamics used to question whether AGI is “feasible — or even advisable” (2024-06-30).
• Takes agentic and AGI-adjacent trajectories seriously after o3 (2025-01-07 universities-need-to-step-up), while staying sceptical of near-term AGI (2025-04-24). Plans around AI 2027 as an edge case, “speculation — no more” (2025-04-06).
• “ill-defined” (2026-04-11). “irrelevant to this conversation” (2026-09-24).
• Centrality. Core as a stance. Remarkably stable: agnostic and implausibility-focused, while increasingly willing to plan for tails.

7.6 Agentic AI — recurring (rising 2024–26)#

• Definition. His working definition: an AI able to determine how to achieve goals, and to adjust them, “by manipulating the environment around it”, including behavioural and social environments. “we’ve never had the ability to create machines that can decide on their own how to solve problems” (2025-05-04). He endorses Kasirzadeh and Gabriel’s autonomy, efficacy, goal complexity and generality, but adds two points: the framework misses AI’s “direct causal effects on the beliefs, understanding, and behaviors of individuals and groups”, and risk scales “closer to exponential than linear”.
• Appearances.
• Seed: open AI ecosystems that “independently decide what’s best for you” (2016-03-02).
• 2024: assistants (2024-04-24); agentic social AI (9.6); the human amanuensis and three-actor model, “AI-directed and human-executed implementation” (2024-11-24).
• 2025: “the keys to the digital kingdom”, agents that act “without asking for permission first” (2025-03-22 when-agentic-ai-takes-charge-manus). “Opportunity” in motive/means/opportunity (2025-07-06).
• 2026: agentic “organoids” needing BSL-4-grade containment (2026-01-31, his own notes only). Humans as “just another cog in the works” (2026-09-24).
• Centrality. Recurring.

7.7 Relational technology, bidirectionality and “not just a tool” — core (2026; roots from 2023)#

• Definition. The “harness” metaphor assumes “capability, but not understanding” and a user left unchanged by use. He argues for “transformation as intrinsic to capability” and “working in relationship” rather than command and control (2026-02-22 what-we-miss-when-we-talk-about-ai-harnesses). LLMs are “a relational technology” even for coding, so companies owe “character constancy” (2026-04-26 why-im-falling-out-of-love-with-claude). Treating AI “as just a tool, is potentially dangerous” (2026-05-21). “This is not just a tool — unless you consider a tool as something that changes who you are” (2026-09-24).
• Roots. 2023-01-31, where working with ChatGPT felt like “a kind and able colleague” and this “slightly worries” him. 2024-09-28, where his default “tool, not a collaborator or partner” gives way to “Now I’m not so sure”. The human–AI “we” (2025-01-12).
• Tension he holds. “Do remember that you’re working with a machine, not talking to a person” (2026-05-10). The relationship is real and consequential, but personhood is an illusion that design invites.
• Centrality. Core in 2026.

7.8 Augmentation, not replacement; catalyst, not substitute — recurring#

Art will always have “humans in the loop somewhere” (2022-09-16). He endorses “augmented intelligence” (2024-09-25). AI amplifies expertise (2025-01-12). AI as “a catalyst to human-initiated thinking” (2025-03-09). AI as a “barrier-thinner” in a spiky, fractal model of discovery (2025-07-27 spiky-surfaces-and-jagged-edges). “This is not AI acting as an independent researcher” but “massively-augmented research” (2026-07-04).

Family 8. The AI risk landscape#

8.1 Plural, mundane-but-serious AI risk — core#

• Definition. His ten risks from Risk Bites (2018-05-12 10-potential-risks-of-artificial-intelligence): technological dependency, job replacement, algorithmic bias, non-transparent decisions, value misalignment, lethal autonomous weapons, re-writable goals, unintended consequences, existential risk from superintelligence, and heuristic manipulation. It is a “whole landscape”, not a ranking around extinction.
• Appearances.
• AI’s risks are “far more mundane” but “no less serious for this”: autonomy, justice, equity and control over our lives. AI is on a “knife edge” (2020-11-12 is-artificial-intelligence-going-to-kill-us-all).
• The list is reaffirmed (2023-04-24 ai-risks-primer).
• Mundane-but-disruptive risks: misinformation as a tool of power, concentration, chokepoints, the AI divide, weapons (2024-01-17). He praises the UN report for not getting “tangled up in speculative long term AI risks” (2023-12-22).
• Against the Action Plan’s narrow list, “a large portfolio” running from dignity and health to social cohesion and democracy (2025-07-23).
• Neglected risks: weakened infrastructure, premature adoption, effects on behaviour and wellbeing (2026-03-22 are-you-an-ai-apocaloptimist).
• The 2018 list “still holds”, plus risks that have risen since: cyber, water and energy, privacy, deepfakes, systemic disruption, governance, children’s development, “psychological/cognitive disruption” (2026-09-15).
• Centrality. Core.

8.2 Existential and catastrophic risk: low probability, taken seriously, reframed as loss of value — core#

• Definition.
• 2018. Occam’s Razor gives superintelligence a low but “not a zero probability”.
• 2023. A risk “of potentially existential proportions” is emerging, yet he did not sign the pause letter (2023-04-04 alternatives). He did not sign the CAIS extinction statement either: extinction is “too narrow and absolute a framing, and too human-centric”. Catastrophe means large numbers losing what they deeply value, and conventional risks are “shavings off the tip of the AI iceberg” (2023-05-31).
• 2024. He criticises the existential-risk ideology of FHI, longtermism and EA (2024-04-28), yet ignoring catastrophe “is in itself a risky strategy” (2024-06-23 existential-risk-jay-baruchel).
• 2026. Treat existential risk as low-probability, high-impact, “without running around like headless chickens” (2026-09-15). AGI is set aside as “irrelevant” to the real risk map (2026-09-24).
• Centrality. Core.

8.3 Bias, prediction and pre-justice — recurring (prominent 2018–2020, then background)#

• Definition. Algorithmic bias: AI apps and machines “that reflect the unconscious (and sometimes conscious) biases of those who develop them” (FFTF p.77). A lineage runs from phrenology through Lombroso and eugenics to fMRI, ML “criminality” classifiers and predictive policing. Pre-emption denies agency and the presumption of innocence (FFTF p.70–83). Opacity: “we no longer know how they’re thinking” (FFTF p.78).
• Appearances. 2018-09-10 minority-report; 2019-03-05 (predictive inequity); 2020-01-01; 2020-09-26 the-seductive-slippery-slope; 2023-05-22 (harm through being “authoritative rather than accurate”); 2024-12-20 sora-has-a-bias-problem.
• Centrality. Recurring.

8.4 Dependency, relinquished decisions and the drain of human agency — recurring#

Technological dependency, the first of the 2018 ten risks. Dependence as fragility (2019-11-19). Decision-making shifting from people to machines, with “we are already irreversibly integrating AI into every aspect of our lives” (2024-07-21). Agency draining through AI-to-AI networks (2024-11-24). Cognitive surrender (9.10). A fictional society’s “de facto AI parents” (2025-11-24 part-1-letters).

8.5 AI as a socio-technical system — recurring#

ChatGPT’s launch was “a social and commercial as much as a technological step” (2023-11-29). Technology is not neutral (2023-12-15). Risk should be analysed at the level of sociotechnical systems (2024-01-14). Technology innovation and society are intertwined (2024-10-06). Scenario thinking wrongly treats AI as happening “to” society (2025-11-30).

8.6 Everyday and relational AI risks — recurring (2025)#

• Email. AI-written email erodes the “relational connective tissue” of organisations, rising to “catastrophic” at organisational scale. The scores were generated by GPT-5 Pro, so they are his framing only (2025-09-07 the-hidden-risks-of-using-ai-for-email).
• Memory. Memory as informant (2025-10-05).
• Power and dignity. An advisor who “holds your academic career in their hands” running a student’s work through ChatGPT (2025-10-26).
• Mental health. Visible harms are “the very small tip of a very large metaphorical iceberg”, and small percentages of hundreds of millions of users add up (2025-11-09).
• Seed. 2023-04-05 can-chatgpt-adversely-impact-mental.

8.7 Deepfakes and a bedrock of reality — occasional#

Deepfakes threaten “a bedrock of reality” (2019-09-04 how-to-ensure-our-digital-legacy). In 2024 he changes his view: he had trusted common sense to spot fakes, is now “far less sure”, and signs a letter supporting criminalisation and developer liability (2024-02-25 ai-rollercoaster-of-a-week).

8.8 Other recurring named harms#

• Jobs. Automation takes away “choice” (FFTF p.123). A Luddite-like backlash that “actually succeeds” (2024-08-07).
• Arms race. 2023-07-25; “inevitable, no matter how bad an idea anyone thinks it is” (2025-04-06).
• Energy and water. 2023-09-11 (co-written); 2024-10-06; 2026-09-15.
• Inference-based privacy. 2022-02-12; 2025-10-05.

Family 9. Manipulation, cognition and formation#

This is the thread that grows most. It runs from neurotechnology acting on the brain (2016), to AI exploiting cognitive biases (2018), to language as the medium of influence (2023), to relational design and emergent influence (2024), to the bypassing of epistemic vigilance and AI’s part in how people are formed (2026).

9.1 Technology acting directly on the mind and identity — recurring (the seed)#

• Definition. Neurotechnologies could “alter how someone thinks, feels, behaves and even perceives themselves and others” without control or consent. He raises “cyber substance abuse”, tDCS in classrooms, and TMS altering moral judgement (2016-03-31).
• Book. Ghost hacking: identity is malleable, “persuasive influences” can change us, and the fear is of becoming “someone else’s puppet” (FFTF p.129–145).
• Later. Who may write to the brain; emotional manipulation by wired-in ads (2019-07-23). “Mental monoculture” and manipulated group think by producers, governments “or even AI systems” (2024-11-17, extending Gordon and Seth).
• Centrality. Recurring.

9.2 Artificial manipulation: AI exploiting cognitive vulnerabilities (Plato’s Cave; the “human club”) — core#

• Definition. The more plausible and “far scarier” AI risk than superintelligence: machines that learn our biases and vulnerabilities and use them “dispassionately”, with “the ability of future machines to bend us to their own will” (FFTF p.159, 174). We each live in a personal Plato’s Cave, and “anyone—or anything—that has the capability of manipulating these shadows has the power to control us” (FFTF p.176). Human manipulators share our foibles; a machine outside the “human club” does not. So we need tests for “when we are being played by machines” (FFTF p.177).
• First appearance. The 2018-05-12 ten risks (as “heuristic manipulation”); 2018-10-12, “Scary stuff, and not that implausible” of Ex Machina.
• Main later appearances. 2018-11-25; 2019-01-16; 2019-03-07; 2019-06-12; 2023-04-16 (republished as “more important today”); 2023-08-25 (machines outside human morality may leave humans “manipulated, used, and enslaved”); 2024-06-16; 2024-06-30 (conscious-seeming AI: “instinctively incapable of acting on this knowledge”).
• Evolution.
• 2018: embodied and fictional (Ava).
• 2023: language.
• 2024: deliberate design (9.5), political economy (9.7) and emergence (9.6).
• 2025: a structured risk (9.8).
• 2026: a bypass of vigilance that needs no manipulation at all (9.9).
• Centrality. Core. The strongest continuous AI-specific thread in his work.

9.3 Engines of persuasion, prediction and nudging — recurring#

Big data plus machine learning as “a subtler and more Machiavellian approach” to controlling people, restrained only by scruples and privacy law (FFTF p.80–82). Predictive inference of inner states, and dependent super-consumers (2022-02-12). Machines learning about humans faster than we learn about ourselves, making prediction and nudging “liberating, or deeply chilling” (2024-04-21 can-ai-be-used-to-automate-social). Next-token prediction of human decisions as a means of manipulation (2025-07-06).

9.4 The language turn: language as the medium of trust, influence and formation — core (2023–26)#

• Definition. “Language plays a large part in how we develop these relationships”; chatbots offer “only the illusion of a reciprocal relationship”, and guardrails fail “as soon as things become personal” (2023-04-05).
• Development.
• Self-replicating ideas and ideologies spread by machines “adroit at manipulating language” that can “slip under the checks and balances of our ability to reason and critique” (2023-04-26 in-bill-joys-why-the-future-doesnt).
• LLMs’ “seductive mastery of language” as a driver of catastrophic value loss (2023-05-31).
• AI is now “part of the mix of influences that inform and modulate our thinking” (2023-08-21 the-messiness-of-the-provenance-of-ideas).
• Language is part of the base code of self-identity (2024-01-01).
• DNA as “just another type of language” (2025-02-23).
• AI’s “mastery of language” can slip past critical reasoning (2026-06-12).
• “Language is formative”: AI is “actively taking part in the formation process” and gives AI “the ability to use language as a lever” (2026-09-24).
• Seeds. The Turing Test’s conversational form and the 2018 NLP tipping point (FFTF p.170).
• Centrality. Core from 2023. It links manipulation, being human and education.

9.5 Hyper-anthropomorphism and relational persuasion — recurring (2024)#

• Definition. His coinage: “a concerted effort to create AI’s that are intentionally designed to engage our anthropomorphizing cognitive biases” (2024-05-15 anthropomorphizing-gpt-4o). Emotional trust is “relational rather than transactional”, and users may “give away more of ourselves to them and their creators”.
• Development.
• Persuasion is social, repeated and relational. A companion in a “one-way relationship” can play our biases “whether this is malicious in intent, or simply an emergent property of the technology” (2024-09-01 is-chatgpts-new-voice-mode-dangerously-persuasive).
• Persona alignment through cultural cognition.
• Benevolent persuasion at scale is the deeper worry: “who decides what is good for society” and “where does democracy fit”.
• Hyper-humanizing and “anthropomorphic heavy lifting” produce “stories that are hard not to trust, and yet are not trustworthy” (2024-09-22).
• Centrality. Recurring.

9.6 Agentic social AI, stochastic agency and hidden influence — recurring (late 2024)#

• Agentic social AI. “AI that gains agency through its ability to make use of human agency”, needing no AGI or consciousness (2024-10-20 learning-to-live-with-agental-social-ai). People must adapt in “a few short years”, so learning to live with it must be “strategic and intentional”.
• Stochastic agency. An explicit revision a week later: “random and unpredictable” influence from emergent “micro goals” of model, user and context, which can harm even when companies act responsibly (2024-10-27 personal-ai-chatbots-and-stochastic-agency).
• Hidden influence. Covert AI alteration of the social signals people exchange (2024-11-03 can-ai-alter-how-you-feel-about-someone).
• Personal evidence. He felt the “affective pull” himself, even knowing it was a bot.

9.7 The economic gradient toward manipulation; dual use; paternalism — occasional (2024, analytically important)#

“the capabilities that make socially beneficial AI Choice Engines viable are the same as those that make AI-driven persuasion and manipulation possible”. A quadrant of value to the individual against value to the deploying agent: Self-Determination, Empowerment, Manipulation, Transformation. Power pulls deployments toward Manipulation, including by governments, and individuals become “engines of value creation rather than the primary recipients” (2024-07-13). The same doubt about benevolent nudging runs through his fiction “Soul Update” (2026-02-11).

9.8 Motive, means and opportunity; emergent vs designed manipulation; universal vulnerability — recurring (2025)#

A crime-solving triad as a risk frame (2025-07-06 ai-risk-motive-means-and-opportunity): - motive: “internal motives” in cornered models; - means: predicting human cognitive behaviour; - opportunity: agents with access to the world.

The premise is human: believing our decisions are rational is “one of our great weaknesses as a species”. Risk is defined by trajectory, “what might be possible given current trends”. His XENOPS scenario probes behaviour, not performance (2025-07-13 whats-grok-4s-moral-character).

AI now triggers responses “previously exclusively the domain of human relationships”; “we all have some degree of vulnerability”. Emergent manipulation can be managed but not eliminated, while designed exploitation “can and should be regulated far more” (2025-08-31).

9.9 The cognitive Trojan horse, epistemic vigilance and honest non-signals — core (2026)#

• Definition. His coinage (2026-01-10 is-ai-a-cognitive-trojan-horse; arXiv 2601.07085). AI is so appealing that we invite it into our cognitive lives without knowing its influence. It may slip past evolved “epistemic vigilance” (Sperber), which trusts by default. Four mechanisms:
• processing fluency (LLMs are “optimized for processing fluency”);
• multidimensional attractiveness;
• speed and volume;
• the Intelligent User Trap.

This is a second-order evolutionary mismatch: AI may impair “the very cognitive abilities we rely on” to compensate. - Development. - “Honest non-signals”: genuine AI traits that lack the tacit meaning they carry in humans. The concept “came from Claude” and he adopted it. AI safety should include “designing systems that present more calibrated trust-cues”, and he proposes “a collective form of epistemic vigilance” (2026-01-17). - The “reasoned hallucination”: “The reasoning was impeccable. The advice unfounded.” (2026-02-08 beeswax-hallucinations-and-ai-inventions). - Frictionless, invisible risk: AI can “slip unawares into our mind and change how we think” (2026-05-10). - Set inside the “who we are” domain (2026-05-21). Erosion of epistemic agency as a coming blindside (2026-07-16). Slipping “beyond our cognitive defenses” (2026-09-24). - Framing. A small-probability, high-consequence risk that justifies research, not restriction. - Centrality. Core in 2026. The culmination of 9.2 and 9.4, but the mechanism no longer needs manipulative intent.

9.10 Cognitive surrender, the “easy button” and the illusion of understanding — recurring (2024–26)#

Technology that “makes you think, rather than doing the cognitive heavy lifting for you” (2024-01-07 the-future-of-being-human-is-analog). Diminished critical thinking as the long-term threat of AI in learning, with “slippage” from personalised learning (2024-08-25). NotebookLM’s voices “bypassed my critical thinking” (2024-09-22). The illusion of understanding (2026-04-11; 2026-05-10). He adopts Shaw and Nave’s “cognitive surrender” (2026-05-21), and the “easy button” “fools you” into feeling productive (2026-09-24).

Contrast with his 2023 view. ChatGPT was then “a profoundly effective catalyst for engaged and creative thinking”, sometimes because of its flaws (2023-08-14 chatgpt-stimulates-creativity-critical-thinking).

9.11 Constitutive resonance, reverse formation and LinkedInification — recurring (2026)#

• Constitutive resonance. His term (SSRN 6343880): a two-way coupling that changes both human and AI (2026-02-22; 2026-05-21).
• LinkedInification. LLMs flatten people into professional stereotypes, which “degrades people”. It reveals “a largely-hidden AI hand promoting specific social norms and expectations”, toward “a nebulous gray goo of conventionality” (2026-03-08).
• Reverse formation. “the AIs we have trained to ‘think’ like us are now beginning to train us to think like them” (2026-07-19 publish-or-perish-ai-vs-human-vs-human).

9.12 Formation — core as a word and concept from 2024#

The process by which people become who they are, now shared with AI: - the formation of social skill for living with AI must become “strategic and intentional” (2024-10-20); - AI extends from emulating learning outputs to “the formation of those outputs” (2026-06-10 is-anthropics-new-ai-model-poised); - a paper’s worth lies partly in the “care and effort in its formation” (2026-07-04); - play contributes to the formation of a mindset (2026-08-02); - “Language is formative” (2026-09-24).

Family 10. Being human#

10.1 The future of being human (a personal lens) — core#

• Definition. How technologies affect “each of us personally” rather than abstract humanity, with the value of remaining “masters of our own destiny” (2023-04-04 welcome-to-the-future-of-being-human).
• Origin. He traces the framing to his 2017–18 chapter title “Being Human in an Augmented Future” (FFTF Ch. 7; 2023-08-21), and it named a 2021 CIFAR call he helped shape.
• Later. 2023-07-31 ai-and-the-future-of-being-human; 2024-01-01; 2025-10-14 ai-and-the-art-of-being-human; 2026-05-21.
• Centrality. Core.

10.2 Worth, dignity and the “convenient lie” — core (book), recurring after#

• Definition. Technologies are most dangerous when they lead a society to “forget the worth of others” and “rob us of our souls, even as it sustains our bodies” (FFTF p.62). The “convenient lie” that some beings are less than human justifies exploitation (FFTF p.57–59). Seduction by benefit: a programme “too valuable to end” (FFTF p.61–62).
• Later. 2018-09-06 never-let-me-go; 2019-10-10 the-science-and-ethics-of-cloning; 2023-08-18 (against dehumanising AI); 2024-04-14 welcome-to-the-age-of-swipe-and-select-embryos (“painful knowledge”; who decides what counts as suffering); 2024-05-21 (disdain for dignity); 2025-10-26 (dignity in advisor misuse).
• Centrality. Core in the book, recurring later.

10.3 “Normal” vs “human”: othering difference and augmentation norms — recurring#

• Definition. Societies accept difference until it threatens what they value, then slide from abnormal to “not human”. Instinctive assimilation of augmentation is “socially lazy”, and values-guided frameworks are needed (FFTF p.136–141). ““different” is no longer simply something we’re born with, but something we have the means to create” (FFTF p.55).
• Normalisation and coercion. “would you be OK” vs “would you do this” once enhancement becomes the norm (FFTF p.98–100).
• Later. Augmentation norms and subscription-locked capability (2024-01-30 first-in-human-trial-of-neuralink-bci); who decides normal (2024-10-13); othering technology (2024-02-18 setting-fire-to-self-driving-cars-is-bad).

10.4 Extrinsic vs intrinsic technologies; the base code of being human — recurring (2024 is the key statement)#

Most past technologies were extrinsic to the self. Emerging ones may be intrinsic, changing us from molecules to beliefs, possibly “without our agreement or permission”. This is a tipping point from changing who we are to “what we are”. Language is part of the base code of identity (2024-01-01). It is the precursor of the “who we are” axis.

10.5 What we do / where we live / who we are — core (2025–26)#

• Definition. A model he uses “increasingly frequently”. Transformative technologies move us along these axes, and AI may push us “far further” (2025-03-30). The three foci are also where universities should do integrated work (2025-01-07 universities-need-to-step-up).
• Later. Reused for planetary health (2025-11-12). Mapped onto Rerum Novarum, Laudato Si’ and Magnifica Humanitas: “who we are” is the domain where AI is unprecedented (2026-05-21).
• Centrality. Core.

10.6 Technology as constitutive of being human — recurring#

We are “already a technologically augmented and enhanced species” (FFTF p.140). Asking whether he is a technology optimist is like asking whether he is an “oxygen pessimist” or optimist. Treating technology as “something we do and not something we are” is misleading and dangerous. He also begins to question his own “technology apologetics” (2024-03-31). AI is intertwined with who we are (2025-11-30).

10.7 What’s worth protecting: joy, wonder, play and the soul of science — recurring#

• Worth protecting. What must be protected “lest we discard it without realizing what we’re losing” (2023-12-15).
• Joy and wonder. “the soul of science lies in the delight and wonder of exploring the unknown”, against “a utilitarian future” (2024-11-10 is-ai-poised-to-suck-the-soul-out-of-science). Advances have “a nasty habit of sucking the joy out of what we do” (2024-12-29). Awe and wonder as value (2025-03-30).
• Play. “play without purpose” as a critical skill (2026-08-02). Universities’ “joy of discovery” (2026-09-24).
• Roots. The humanity of science: humility, curiosity, “awe and wonder” (FFTF p.278–282).

10.8 Pro-human, not pro- or anti-AI; learning to be human with AI — recurring (2025–26)#

• From defending distinctiveness to growing with AI. He moves from defending human distinctiveness to working with AI “to make us more than we are” and learning “how to be human” with it (2025-03-30).
• The book. It is “not for or agains AI, but is pro human” [sic] (2025-10-14). Its central question: “What makes me me when AI can finish my sentences, replicate my style, and predict my choices?”
• AI as a way back to ourselves. AI as a way to rediscover who we are (2026-05-21).
• Provenance. The four postures (Curiosity, Intentionality, Clarity, Care) and the 21 tools come from the co-written, AI-assisted book, so they are weaker evidence.

10.9 AI as mirror — occasional#

Machines that mimic language and reasoning hold up “a mirror … (albeit a rather imperfect one)”. The suggestion is that what defines us may be computation machines can match (2024-01-01; 2023-12-15).

Family 11. Learning, education and the university#

11.1 Universities’ public responsibility, later universities as navigators of the AI transition — core#

• Definition. Public universities’ incentives crowd out public good. He proposes community service as a fourth leg of faculty work: “the institution gets in the way” (2016-01-31 public-universities-must-do-more). Boundary organisations such as PEN are needed (2016-02-01).
• Development.
• Universities must practise responsible innovation (2023-04-18).
• Futures research belongs in accountable public universities (2024-04-28). A school of advanced technology transitions (2024-08-11).
• Universities are “mired in tradition, convention, and self preservation” (2025-01-07 universities-need-to-step-up). Serendipity and speed (2025-12-07 universities-genesis-mission).
• “we owe it to” students (2026-03-29). Universities have been “followers and users of the technology” and “guardians of the past more than leaders toward the future” (2026-08-30).
• The governance gap, the academic crab bucket, and a new “paradise” of service that helps people “retain their humanity” (2026-09-24).
• Centrality. Core.

11.2 AI literacy: from universal skill, to ways of thinking, to its limits — recurring (with change)#

• 2023. AI literacy for all majors (2023-05-09 not-your-traditional-prompt-engineering). “every high school” (2023-07-27). The real hazard is “not smart” use (2023-08-14). Literacy beyond compliance (2023-07-10 eu-ai-act-and-education).
• 2024. Literacy as ways of thinking about persistent failure modes (2024-12-13).
• 2025. Educators must know AI first-hand (2025-08-10). He dropped his prompt-engineering course as outdated (2025-08-17). Literacy classes “risk becoming performative” (2025-11-09).
• 2026. Demonstrable “I can” skills (2026-04-14). Literacy alone will not change risk behaviour (2026-05-10).
• Centrality. Recurring. His confidence in literacy as a remedy declines.

11.3 Playgrounds, not playpens; the lowest level of tech necessary — recurring#

From Bers and Resnick: open, curated spaces with norms against harm, versus controlled paths (2024-03-17 undergraduate-playgrounds-not-playpens; 2025-03-15, “permission to play”, with educators as “fellow-travelers”). His teaching rule is “the lowest level of tech necessary” (2024-02-11). The idea extends to children’s imagination with AI video (2024-03-24) and to living books (2026-04-02).

11.4 Intelligence scarcity and abundance — recurring (2025–26)#

Daley’s “intelligence is free” challenges a scarcity model in which elite establishments hoard access. He finds this democratising but “controversial, and not entirely accurate” (2025-03-30). A crisis of abundance for academia (2025-11-30). Universities “provide a service in a world of intelligence scarcity”, and AI “claims to give everybody intelligence for free”, which is an existential threat to identity, not survival (2026-09-24; 2026-06-12).

11.5 Education as a lever against inequity; democratised learning — recurring#

Automation deprives people of “choice”, and education must keep up (FFTF p.123–125). YouTube-era learning is a missed opportunity for experts (FFTF p.125–127). AI as translator and equaliser (2023-07-27). Flattening the learning distribution curve (2023-10-24). Agent-built courses as “a revolution in how knowledge and training flow through society” (2025-03-27).

11.6 Play, serendipity and learning by not trying to learn — recurring#

Bounded infinities and metaphorical quantum tunnelling: the arts, humanities and serendipity help us escape conventional thinking (2021-04-09). Tunnelling between discovery spikes (2025-07-27). The serendipity–speed matrix (2025-12-07). Learning with “no learning expectations”, and joy as an under-appreciated metric (2026-08-02).

11.7 Smaller education concepts — occasional#

• Value-creation model and crisis of purpose. Education matters because it amplifies the human capacity to imagine and create value, and AI unsettles this: an “existential crisis” of purpose (2025-03-30).
• Adoption. Naive adoption and lock-in, “investing in concepts, not products” (2024-05-05).
• Conversation, not prompt. 2025-08-17.
• Assessment. Learning assessment vs threshold assessment (2025-08-24).
• Institutional agility. A big “if” (2026-06-10).

Family 12. AI in his own scholarship and writing#

12.1 Writing as self; authorship and provenance — recurring#

• Writing as self. “I typically don’t use ChatGPT myself when I write”. Handing writing to a machine “would be to diminish myself” (2023-09-20 what-do-college-students-think-about-chatgpt).
• Authorship. Who or what was the “creative force”? (2024-09-28).
• Craft and identity. AI “can write better than me, faster than me”, but the value lies in “the very human piece of me” (2025-01-30, his framing of a ChatGPT-written piece). “so much of my professional and academic identity is embedded in my writing” (2025-03-16).
• Later positions. AI prose is “superficially profound yet substantively hollow”, and a “cargo-cult” paper has the form of scholarship without a grasp of human readers (2026-07-19). AI named as sole author with a CRediT-AI statement (2026-09-04).

12.2 The changing practice of AI-assisted research — recurring (an arc, not one concept)#

• 2022–2023. Enthusiasm for co-creation and “machine-augmented communication” (2022-12-08). A “kind and able colleague” (2023-01-31).
• 2024. Co-creation (2024-09-28).
• 2025.
• Non-AI-augmented scholarship may look “intellectually limited and somewhat quaint” (2025-02-04).
• Reasoning models as research partners (2025-04-06). “Developed by AI, checked by humans” (2025-04-24 us-and-china-vie-for-ai-k-12-leadership; 2025-05-25).
• AI used deliberately to counter his own biases in risk scoring (2025-09-07).
• 2026.
• “I cracked”: a Claude-drafted paper, with a moral line between “slop” or profile-padding and AI-assisted discovery as a “public good” (2026-01-17).
• The validation gap (12.4). Augmentation, not autonomy (2026-07-04).
• Disenchantment with AI prose (2026-07-19). “AI used well doesn’t necessarily make things faster” (2026-09-24).

12.3 The artisanal intellectual and intellectual craft — recurring (2025–26)#

• Coinage. In a footnote, AI will “relegate human-only research to a class of artisanal intellectualism where the primary purpose is the provenance and process, not the product” (2025-02-09 can-ai-write-your-phd-dissertation).
• Self-application. A week later he applies it to himself, as the human craftsperson judging and building on AI output (2025-02-16, his framing only; the article body is Deep Research’s). He calls it “a bit of a throwaway” that grew on him.
• Fiction. Developed in the story Letters from the Department of Intellectual Craft, after C. Wright Mills (2025-11-23 to 2025-11-30). There, intellectual craft moves from affectation to value, and AI threatens identity most where intellect is “the very things that AI sets out to excel at”.
• Exclusion. A Modem Futura show-note definition is excluded.
• Status. The meaning is still unsettled.

12.4 The validation gap and the epistemic precipice — one-off (2026), linked to Family 9#

Agentic AI produces work that needs rare multi-field expertise to evaluate; even he is “reaching my own limits”. We may be at a precipice where AI generates knowledge faster than we can validate or understand it. The counter-risk is the illusion of capability through mastery of language (2026-06-12 a-quick-update-on-using-claude-fable-5).

12.5 AI-first publishing and AI as intermediary reader — occasional#

AIs as “the predominant consumers of the written word” (2026-04-02 spoiler-alert-wtf, his framing). llms.txt-indexed, AI-legible “living book” formats and AI-legible scholarship (2026-06-14, his framing only, per the ruling).

Family 13. Stories, science fiction and imagination#

13.1 Science fiction as a lens, not a forecast — core (method)#

• Definition. Films are poor guides to technology but reveal technology–society dynamics. They are useful precisely because they are untethered from accuracy, and they break institutionalised thinking. Art is “a common point of focus” across divides (FFTF p.15–25, 288–289).
• Appearances.
• 2018–2021: 2018-10-12; 2019-01-16; 2019-03-07 (films as “social and educational levelers”); 2021-01-15 (transformative learning “has to be felt”).
• 2023–2024: 2023-10-08 (the catalyst, not the destination; the book “was a failure” commercially); 2024-03-03 (sci-fi relevant but unreliable); 2024-05-26 (entrepreneurs take the cool tech and miss the social message); 2024-07-30 (the Moviegoer’s Guide class); 2024-09-18 (Virk’s Sci Fi feedback loop, adopted).
• 2026: 2026-04-02 (FFTF “far more relevant now”); 2026-05-15 (most AI films are not purely dystopian; only the category names are his).
• Danger. Dystopian fatalism.
• Centrality. Core.

13.2 Stories as the pivot between imagining and building futures — recurring#

Stories are “the pivot point” between imagining a future and building it, and open minds where preaching closes them (2024-01-21 how-can-stories-unlock-pathways-to). The flip side: who writes the stories that govern our futures (2024-09-22). Fiction’s affordances for nuance (2025-10-14; 2025-11-23). His own fiction as method (the Letters; “Soul Update”, 2026-02-11). The Wyndham-style social-consequences brief (2024-09-28).

Family 14. Temperament and method#

14.1 Obligation to innovate; “Don’t Panic” — core#

Renouncing technology “from a position of privilege” denies others choices, and “we have an obligation to explore” how technology can improve lives (FFTF p.288; 2018-11-08; 2019-03-06). “Don’t Panic”: the twin errors are panic and enchantment with the tech (FFTF p.289–290). The late form is compressed: “you can only begin to realize the benefits of a technology if you understand what can possibly go wrong” (2026-09-24).

14.2 Nuance against polarisation — core#

• Against spectacle and sloganeering. Anti-Hollywood from the start: real AI failures look like “teens troll supercomputer”, not apocalypse (2016-03-02). No “bumper sticker” stances (2023-07-25).
• Against both camps. Neither doomer nor booster: the oxygen analogy (2024-03-31). Opinions from both camps are “only loosely tethered to reality” (2026-03-22). Both AGI speculation and “nothing new” dismissal are “blinkered and naive” (2026-09-24).
• Self-description. “human wellbeing and flourishing at the heart of my work” (2025-01-30, in his own prompt).

14.3 Honest broker and public scholarship — recurring#

He takes Pielke’s Honest Broker as “the role I try to carve out for myself” (FFTF p.244–247), qualified by a turn to collective, institutional advocacy when silence becomes tacit support for inaction. He recommends the role again in the gain-of-function debate (2023-09-15 weaponizing-the-genome). Public writing is integral to his scholarship, justified only by public good (2026-05-17 the-nonsense-i-write). Other channels: Risk Bites, and parasocial reach for experts (2025-05-25).

14.4 Self-implication and candour — recurring (method)#

He confesses his own rule-bending (FFTF p.161) and his coopting of a social-good narrative (FFTF p.219). He admits he may be wrong (FFTF p.170). He was “suckered by Claude” while writing about being suckered (2026-02-08). “I was wrong” about early GPT APIs being “a toy” (2026-09-24). AI is “one of the scariest things I’ve ever seen” (2026-09-24).

14.5 Building to think — recurring (2023–26)#

• Hands-on experiments. ChatGPT as author (2024-09-28); XENOPS (2025-07-13); agent tests (2025-03-22); AI-built prototypes and apps (2025-05-18; 2025-07-27).
• Tools. The Risk Innovation Planner, the ATT models and the Trajectories Tool.
• Refusing AI when meaning matters. He wrote the Action Plan analysis deliberately without AI, because summaries miss “meaning, implications, subtexts” (2025-07-23).

Appendix: date-to-slug index for every post cited#

Every post cited in this file, by date. Look up the full filename as corpus/<full slug>.md. - In the body, Medium-era slugs appear without their trailing hash, and some long Substack slugs are shortened. - Where two posts share a date, the body adds a short slug (for example “2016-01-11 thinking-innovatively” vs “2016-01-11 fourth-industrial-revolution”). - 2025-11-24 start-here is cited only as an exclusion.