Late Lessons, Jensen Huang and AI
Markdown version

F3. Risk as a way of thinking#

One facet of how Andrew Maynard thinks and works, read from his own perspective. Built from the perspective notes (B01–B32, FFTF-A, SUPP-self) and checked against the original posts, columns, papers and book passages. Posts are cited by date and slug; papers by file name and page; Films from the Future (FFTF) and Future Rising (FR) by printed page; Rethinking Risk (2017) as RR by book page; the 2023 TechTrends interview as TT by PDF page (his quoted words only). [AI-edited] marks his lecture text as smoothed by a model; [mixed] marks the 2026 orphan-risks paper, whose long-standing ideas are his but some of whose analytical apparatus came from a Fable 5 draft. Quotations are exact apart from straightened apostrophes and dropped italics.

1. The facet in his terms#

The shortest statement of this facet is in the first chapter of Films from the Future. After listing a working life spent on risk (inhalation research, nanotechnology safety, teaching risk assessment, running risk centres), he writes: “if there’s one thing I’ve learned over the years, it’s that I have less and less patience for how many people tend to think about risk” (FFTF p.22). Established approaches “work reasonably well” for conventional technologies, but “run out of steam rather fast when we’re facing technologies that can achieve things we never imagined”. We are trying to “squeeze the new wine of technological innovation into the old wineskins of conventional risk thinking”. The answer is to “realign how we think about risk with the capabilities of the innovations we’re creating” (pp.22–23).

Four features of that passage recur through his work.

His September 2026 account of himself says the same. When a technology fits no previously encountered type of risk, the whole mindset about risks, benefits and how to move between them has to change. That is why he talks of risk innovation, the risk landscape, navigating rather than managing, risk as a threat to value and orphan risks. He offers them as mental models that open possibilities, not as operating procedures. They rest on quantitative risk science rather than replacing it, and they are held with humility against the hubris of false precision.

The record supports this account closely and adds depth to it. The formative insight is older than AI. It is older than the name “risk innovation” too. It came out of nanotechnology, where he learned from inside a quantitative field that the field’s own categories did not track what mattered. And play, creativity and serendipity were part of the risk-innovation idea from the day it was named. They were not added later.

2. Where the insight came from: an apprenticeship in nanotechnology#

He came to risk as a measurer. He was an aerosol physicist. In the early 1990s his own techniques collected and imaged ambient particles “just a few nanometres in diameter” (nnano.2015.120 p.482). He led research at the UK Health and Safety Executive. In 2004 he was at NIOSH “measuring what happened when you opened a packet of carbon nanotubes” (2026-04-12 What-Thirty-Years). His later framing grew out of that work. It was not imposed on it from outside.

What nanotechnology taught him was that the quantitative tools were necessary and yet could mislead when they were pointed at the wrong thing. The lesson arrives in stages across a decade.

2006–2009: the data run out. As Chief Science Advisor to the Project on Emerging Nanotechnologies he wrote that “Relying on existing knowledge to quantify the risks of engineered nanomaterials will engender false assumptions of safety. We need to think differently” (PEN research strategy 2006, PDF p.15). Before Congress in 2007 he said “we do not yet know what are the right questions to ask regarding potential risks”. He added that such knowledge would come from “researchers who are given the freedom to explore new avenues and follow interesting leads” (House testimony 2007-10-31, PDF p.33). Already, curiosity-led exploration was his answer to not knowing which questions to ask. In 2009 he put the point in the form he still uses. “Numbers—hard data—can be comforting. But without a clear idea of their relevance, they can also be misleading.” The section heading that follows is “When the data run out – innovate!”, for a world where the science-based picture of risk “looks increasingly like a Swiss cheese” (2020science 2009-08-29 ten-things-everyone-should-know-about-nanotechnology-safety). This is the seed of risk innovation, six years before the name.

2011: a category that does not fit, and a public change of mind. “Five years ago, I was a proponent of a regulatory definition of engineered nanomaterials. I have changed my mind.” A size-based definition would make regulation “a ‘term of art’ rather than science”. His cautionary case was Libby vermiculite, whose asbestiform fibres “slipped through the regulatory net” because they did not fit the official definition of asbestos. His alternative was attribute-based “trigger points”, which “must be flexible, so that they can be modified as evidence grows” (Nature 475:31, 2011). The lesson is the one he later applies to AI: a category chosen for convenience can hide the harm it is meant to catch.

2014–2016: the columns. His sole-authored Nature Nanotechnology columns turn the same scrutiny on his own field. - The research programme set up by the 2004 Royal Society report had “worn a rut that is proving hard to get out of”. In it, “The speculation of possible risk has developed into an assumption of as-yet-to-be-discovered risk” (nnano.2014.43 p.160). - Novelty is “a rather unreliable indicator of potential risk”. It “favours the interesting (and possibly the headline-grabbing) over the important”, and “mundane risks are still risks” (nnano.2014.116 p.410). He cites, in that column, the nanotechnology chapter he co-wrote for the European Environment Agency’s Late Lessons from Early Warnings (2013). - New evidence on fumed silica, his own standard example of a safe nanomaterial, “cast doubt on what I thought I knew to be true” (nnano.2014.196 p.658). He weighed it and did not overreact. He asked instead “how are appropriate trigger points for action defined?” (p.659). - “seemingly novel challenges don’t always demand novel solutions” (nnano.2015.120 p.483). - For entrepreneurs, the greatest barriers to responsible innovation are “not necessarily time and cost, but imagination” (nnano.2015.35 p.200). - Established risk methods “were developed as a consequence of, the impacts of previous industrial revolutions”, so “we need to be jolted out of our existing mental and procedural risk-ruts” (nnano.2015.286 p.1006).

So the formative insight was never simply that a technology is new and needs new thinking. It came in two parts, and he kept both. Labels, categories and habits of mind can stop tracking what matters, and when they do they create both false alarms and false comfort. The old measuring tools, used with judgement, still work. The call for a new mindset and the insistence on the old rigour come from the same experience.

A smaller detail points the same way. In autumn 2011 he and the speculative designer James King worked with a small group of University of Michigan science and public health students. The students made creative work about “the tension between the catastrophic consequences often imagined to arise from human endeavors, and the mundane reality that often develops” (2020science 2012-02-04 exploring-speculated-catastrophe-and-mundane-reality). Art as a way of thinking about risk predates risk innovation.

3. 2015: risk innovation named as an act of imagination#

“Why we need risk innovation” (nnano.2015.196, September 2015) is the founding statement. Read closely, it supports his self-account more strongly than any later text.

Then comes the telling choice of examples. His first case of risk innovation “in practice” is “a book of seventeen haiku”, from a workshop with the V2_ Institute for the Unstable Media: “an unusual result from an academic meeting”, meant to open “risk navigation pathways to its readers that would otherwise remain hidden”. The other end of the spectrum is the EPA and NIH Tox21 high-throughput toxicology programme (p.731). Poetry and computational toxicology sit on one spectrum. Risk is “not only endemic, but integral to progress”, and “Without risk innovation, all we are left with is business as usual” (p.731).

His first public explanation, in The Conversation, borrows its structure from entrepreneurship: “Imagine what might happen if we approach risk the way entrepreneurs approach innovation”. Innovation creates value people will pay for, so risk becomes a threat to value, and its “market” is the people with something they are “willing to invest in protecting”. He describes the whole approach as “designed to open up new ideas and possibilities” and notes that it “opens up new ways of imagining the risk landscapes that new technologies both face and help to form”. He also makes a claim about safety that conventional risk thinking would not make: “this lack of creativity and flexibility in how potential risks are understood and addressed only increases the chances of things going wrong”. And he cannot resist a joke about CES gadgets: “(And yes, for someone out there, I’m sure smart shoes will be a life-enhancing experience.)” (2016-01-11 thinking-innovatively-about-the-risks-of-tech-innovation).

The upshot is that creativity is treated as a safety property from the start. His 2026 claim that play and creativity are “integral” to breaking out of stovepiped risk thinking is therefore documentably true of the idea’s origin, not a later gloss.

4. The mental models and what each one opens up#

He uses five linked ideas. Each is best understood by what it lets people see or say that they could not before.

Risk as a threat to value#

“Risk starts with something that is worth protecting.” The probabilistic definition is “a useful starting point”, and worth also includes “security, friendships, social acceptance, and our sense of personal and cultural identity” (nnano.2016.28 p.211).

The clearest account of why he needs this frame is personal, and he tells it against himself. Facing a one-in-a-million chance of serious harm from contrast dye before a CAT scan, he writes: “As a physicist, I’m expected to be good with numbers.” He could not make “rational sense” of the trade-off. He signed “not because I’d done the math and it made sense, but because that was what I was expected to do”. The value that decided it was not being embarrassed in front of the staff (RR p.195). The numbers were right and did not help, because they did not capture what mattered to the person deciding. He concludes that “Risk calculations are also highly dependent on what is considered important, as well as who decides what’s important” (RR p.194).

What the frame opens up is explicit in his own words. - It makes resistance intelligible. “I’m not sure I buy the idea of ‘risk aversion’”, because it hides “the things that people find too important to risk losing” (RR p.193). In The Man in the White Suit “everyone is shrewd enough to see how change supports or threatens what they value” (FFTF p.225). Fifteen years later he builds a moral-panic timeline and refuses to mock the panics, reading them through “how threats to what’s important to people can lead to responses that may seem irrational on the surface” (2025-06-01 vibe-coding-moral-panic). - It turns yes-or-no decisions into design questions. Conversations “can be elevated from simplistic ‘go/no-go’ options” to how gains and losses are balanced, which “opens the door to creative and innovative approaches to protecting existing and future value” (RR pp.197–198). His most striking sentence goes further: “Risk in this instance is not a danger to be avoided, but an inevitability that reveals what the primary value is within a complex landscape” (RR p.197). - It puts benefits in the same account. “Beyond existing value, future value is also important”, and the “greatest risk” is losing sight of technologies that make the world “a better place, not just a different one” (2016-03-02 how-risky-are-the-world-economic-forums-top-10). In 2023 he used the same move against the extinction framing: seeing catastrophe as loss of value at scale lets us “consider the potential loss of solutions to pressing challenges” (2023-05-31 existential-risks-of-ai). - It turns back on the actor. It includes “the reciprocal dangers of threatening what is important to others through what they do” (2018-12-13 tech-startups-orphan-risks). He later calls this “your risk is my risk” (2026-07-16 [mixed]).

He is candid about the costs. It is “not necessarily a comfortable reconceptualization” (2016-03-02). It is “admittedly, a somewhat subjective way of thinking about risk”, valued because it has “the advantage of opening up conversations” (2018-12-13). And he jokes, “imagine a regulator including interpersonal relationships in risk assessments—it’s hardly likely to make the process any easier” (RR p.197). He separates “value” from “values” (2016-01-11), and later calls the frame “agnostic to particular worldviews” (2023-11-21 ai-and-risk-innovation).

The risk landscape#

The landscape is “the risk landscape that lies between new ideas and their successful implementation” (2018-12-13). It has “shifting hills and valleys” (nnano.2016.28 p.211). Technologies “both face and help to form” it (2016-01-11), and it is full of people, their agendas and their power (2023-07-25 oppenheimer-and-ai).

The picture has a physics behind it. From chaos theory he takes two lessons together: “we cannot wield perfect control over complex technologies within a complex world”, and yet there are boundaries that help in “separating out plausible futures from sheer fantasy”, and “points of stability”, with futures “that can be squandered if we don’t think ahead” (FFTF p.41). Unpredictable within bounds, with some leverage: that is the ground for a map rather than a forecast, and for steering rather than control.

“Navigate” is his working verb across the whole record. Titles such as “Navigating the risk landscape” (nnano.2016.28) and “Navigating the fourth industrial revolution” (nnano.2015.286) come early. His 2026 retrospective describes risk innovation as “a reframing of risk from something to be minimized to something to be navigated creatively in pursuit of value” (2026-04-12 What-Thirty-Years).

The contrast with managing is not a rejection of management. He keeps “manage” for the operational layer: safety “operationalized as assessing and managing risk” (2024-06-20 ilya-sutskevers-safe-superintelligence-rethink). He also still calls some orphan risks “manageable” (2021-09-07 should-we-be-worried-about-elon-musks-tesla-bot). Navigation names the stance within which management tools are used. Several later lines say what the stance involves: - “traditional ‘set it and forget it’ management doesn’t work”; success needs “rapid course correction” (2025-05-18 exploring-ai-through-cause-and-effect); - AI innovation can be channelled “much as a flood can’t be halted, but it can be directed” (2025-08-31 holding-on-to-our-humanity-age-of-ai); - “you can only begin to realize the benefits of a technology if you understand what can possibly go wrong, so that you can navigate around it”, so as to “avoid it or flip it, and so get to the good” (2026-09-24 being-an-academic-in-an-age-of-ai [AI-edited]).

“Flip it” is the entrepreneur’s instinct in the risk-innovation idea: a threat can sometimes be turned into an opening, not only reduced.

A small marker of the vocabulary shift: his 2008 World Economic Forum notes called for policies “to foster and manage the effective and sustainable development of emerging technologies” (quoted in 2023-12-22 un-governing-ai-for-humanity). By the mid-2010s the verb is navigate.

Orphan risks#

This is the concept that changed most. It began as a reframing of a familiar taxonomy. Rumsfeld’s knowns and unknowns leave out risks that are “‘known knowns’ if you’re looking in the right place, but aren’t taken as seriously as they should be”. They are dismissed as “too ill-defined, too complex, or too irrelevant to be worth paying attention to” (2018-12-13). He places its seeds in 2013, “while I was teaching entrepreneurship students at the University of Michigan who faced a bewildering landscape of hard-to-quantify social and political risks that none of their business tools addressed” (2026-07-16 [mixed]).

By 2020 the orphan risks are “hard to quantify threats to value that often slip between the cracks of conventional risk approaches” (2020-10-15 the-ethics-of-advanced-brain-machine-interfaces). In April 2026 he uses the term for AI’s human-side risks, to “dignity, belonging, identity, autonomy, democratic participation, what it means to be human”, which “no existing institution owns” (2026-04-12 What-Nanotechnology). In July 2026 the question becomes institutional: not which risks are orphaned but “by what process does a known risk come to be nobody’s responsibility?” His closing line is that the risks most likely to blindside frontier AI “are the ones its institutions have organized themselves not to see” (2026-07-16 [mixed]).

Tools as catalysts#

The tools he built were deliberately small. The Risk Innovation Planner’s design brief was to help a startup founder “develop a risk innovation mindset” in about 30 minutes, and “What the Planner does not do is provide answers to problems”. It is “a catalyst” (2023-11-21). He also offered a quadrant model “I’ve been playing with for some time now” as “definitely not rocket science!”, under the heading “Not Quite a Tool Yet” (2024-08-25 advanced-technology-transitions-model). And he described the early risk-innovation work, looking back, as “providing organizations with simple tools for identifying and navigating potential threats to value” (2026-05-03 are-design-principles-for-responsible, n.4). The tools exist to shift thinking, which is the order of priority in his self-account.

5. Built on, not discarded#

He describes each new layer as an addition. Probability is “a powerful way of making trade-offs” (RR p.193), and the value frame is “an evolution of the old black-and-white mathematics of risk” (RR p.200). Threat to value “extends conventional thinking rather than replacing it” (2018-12-13). In 2026: “This does not abandon the idea of risk as involving the probability of harm”, and the approach is offered “not as an alternative, but as an augmentation” (2026-07-16 [mixed]).

The foundation is not a formality. He keeps using it. - Benchmarking “helps tether speculative ideas to plausible realities” (nnano.2016.28 p.211). - Algorithms read through chemical risk (2019). He carried hazard, exposure, dose-response and weight of evidence over to algorithms (“Of course, an algorithm is not a chemical”). He did so to stop algorithmic risk being built “on an evidentiary stack of cards” (2019-03-05 should-we-be-treating-algorithms-the-same-way-we-treat-hazardous-chemicals). - Graphene masks (2021). He reasoned from deposition physics to a respirable range of about 5–10 µm. Missing release data was “in itself a red flag”, and “hope alone is not good enough” (2021-03-28 how-safe-are-graphene-based-face-masks). - Waymo (2023). He re-ran a safety study’s baseline on the back of an envelope (2023-11-09 waymo-safety-study-shows-benefits). - First principles for AI (2023). Faced with a year of AI-risk noise, he went back to “first principles”: “no cause, no risk”, and “bleach is hazardous, so is a piano”. A day later he added a note on why he had left out “risk = hazard x exposure”: he did not want to imply “that zero exposure — as in no AI — is a default risk management strategy” (2023-11-26 everything-youve-heard-about-ai-risk-is-wrong). - Safety as social (2024). Using the risk-science toolkit itself (acceptable risk, “a one in a million risk of harm is considered OK”), he argued that safety is “ultimately a social construct” and that “zero risk — the corollary of absolute safety, is only possible in the absence of change” (2024-06-20).

The direction of travel runs both ways. Where old tools cannot see social risks, he asks for new thinking. Where a new field is naive, as with algorithmic bias in 2019, he asks for the old rigour.

6. Humility against false precision, without paralysis#

His September 2026 account names humility as the guide for approaching something as poorly understood as AI. The concern is as old as his quantitative work: - “false assumptions of safety” (2006); - numbers “can be comforting” but “misleading” (2009); - a statistical parameter “may not adequately reflect a risk parameter of relevance”, and “How you think about nanotechnology risk is probably incomplete” (nnano.2016.28 pp.211–212); - “The more precise we try to be with our predictions of the future, the less likely they are to be accurate”, with the danger that we “act as if the future is something we can fully control” (FR p.148).

The mathematics of change helps us see “as change threatens to take away what we value”, but only “when tempered with humility and guided by our humanity” (FR pp.166–167).

For AI the humility is sharper. “The more I study artificial intelligence, the less certain I am that we even know how to formulate the problems we face around AI, never mind manage the risks” (2023-11-26). The landscape is one “that only the foolish would claim to understand with certainty” (same post). He tells a research agent writing about his own field, “No recommendations at this point – remember the humility bit” (2025-02-04 openai-deep-research-ai-scholarship, his prompt). He treats the claim of absolute safety as hubris, not ambition (2024-06-20). His rule for speculation is “don’t disallow speculation, but do it within a context of humility”: know it is speculation, look at possible rather than real futures, accept that data must follow, and bring in different voices (2026-09-24 n.4 [AI-edited]).

This humility is not an excuse for doing nothing. In 2016 he stated the balance as a rule. Researchers need “the freedom to ask the ‘what if?’ questions”, and premature action on immature science is itself risky. Hence “quick to question, and slow to respond”, together with the ability to act “where early warnings of potential harm do begin to emerge — even before the science is mature” (nnano.2016.28 p.212). Later he refuses the idea that unease should be ignored: “it’s pretty much impossible to manage risks if you don’t talk about them”, and existential risks should not be dismissed, but they can be approached “without running around like headless chickens” (2026-09-15 will-ai-really-kill-us-all).

The same stance explains his sparing use of numbers for AI. He uses them where they bite: the Waymo baseline, or the scale arithmetic he foregrounds when Altman’s “fewer than 1%” of users in distress is set against more than 800 million weekly users, with visible cases “the very small tip of a very large metaphorical iceberg” (2025-11-09 universities-chatgpt-mental-health, text and n.1, the arithmetic reported from Bloomberg). He avoids them where they would be theatre. His diagnosis of the World Economic Forum risk survey makes the same point about expert aggregation. Opinions “regress to the mean”, which guards against speculation but tends “to devalue risks that are poorly understood by a broad base of mainstream experts” (2025-01-19 wef-global-risks-2025).

7. Why AI makes the change of mindset non-negotiable#

The record has two layers. The general layer, argued from 2014–15, is that converging technologies outrun frames built for earlier industrial revolutions. The AI layer, argued from 2018 and much more strongly from 2023, is that AI acts on the faculties we would use to navigate it.

He holds continuity and novelty together, and this matters. The nanotechnology rules carry over: “nature doesn’t care what we call a material, it just cares about how it behaves”, and “there is no reason to assume that new materials, by default, present new risks” (2022-02-10 are-we-asking-the-right-standards-questions). Of the whole career he writes, “The specifics have changed enormously. The pattern hasn’t” (2026-04-12 What-Thirty-Years). The reading that fits all his statements is this. Lessons about process, humility and how societies meet new technologies transfer. Categories, labels, thresholds and track records may not. His habit is to ask which is which, not to assume that everything is new or that nothing is.

8. Nuances and tensions#

His account is well supported, but the record has edges that a faithful reading should keep.

Tool and mind. In 2017–2020 the concepts were also sold as tools and as a business case. The Risk Innovation Accelerator was “building tools” for startups and investors (2018-12-13), and the Planner aimed at “competitive advantage” (2023-11-21). He chose to meet entrepreneurs in their own language rather than preach at them. His own later lesson explains the choice: “if you want a fast-moving organization to attend to a risk, you do not hand it a compliance duty; you show it a threat to something it values” (2026-07-16, his own lesson from work with Elizabeth Garbee). The tools were always framed as catalysts. Still, the mental-model reading is partly a retrospective emphasis on what was, at the time, also a service.

Whose value? The frame’s early uses face the enterprise. That risks recasting ethics as enlightened self-interest, where harms to people without leverage count only if they come back to the firm. He saw this (“Quite apart from the ethical inappropriateness”, 2018-12-13). He insisted that decisions should protect what is valuable “not just to corporations and governments, but also to individuals and the communities they are a part of” (RR p.200). He noticed that AI can leave people as “engines of value creation rather than the primary recipients of created value” (2024-07-13 ai-choice-engines-sunstein). But “agnostic to particular worldviews” does not settle conflicts between values, and he offers no procedure for doing so.

An evidentiary bar. Value can be “named, mapped and watched — even where it cannot be measured” (2026-07-16 [mixed]). He does not say when watching should become acting for harms that cannot be quantified. His own resources for this are strong but unused for AI: trigger points (2011) and “quick to question, and slow to respond” (2016).

Plausibility and tails. In 2018 he put evidence-based harms ahead of superintelligence (FFTF p.281). By 2025–26 he takes edge cases seriously “on the off chance” (2025-04-06 responsible-innovation-and-ai-acceleration) and argues for research “even if there’s only a small chance” (2026-01-10). The two fit if one separates free speculative research from evidence-gated action, which is the 2016 rule. He rarely says this in so many words.

His own precision. In the email-risk study he let GPT-5 Pro generate 0–10 scores, partly to counter his own bias, and those scores moved him from “trivial” to “extremely seriously” (2025-09-07 the-hidden-risks-of-using-ai-for-email). The method was open and the purpose exploratory. But it leans on numbers in the way his warnings about false precision caution against.

Does it work? He is honest that the risk-innovation analysis “has yet to be shown to be useful in practice”, and he proposes three tests that could count against it (2026-07-16 [mixed]). No evaluation of the tools’ effect on outcomes exists in the record.

Navigation and irreversibility. Navigation assumes you can course-correct. He knows where that fails: “Context is everything here.” Experimenting where it is “easy to turn the clock back” is one thing; breaking “people, governance, society, and the planet” is another (2025-03-02 the-lure-of-permissionless-innovation, n.2). His hysteresis curve says the same: reducing a cause does not always reverse its effect (2025-05-18). The limits of navigation are in his own record. They are not yet part of the frame’s core statement.

9. How this facet connects to the others#

10. What is distinctive, and what these framings are worth for AI#

What is distinctive#

  1. An insider’s reframing. He is a measurer who knows the numbers well enough to know where they stop helping. His critique of conventional risk thinking comes from inside quantitative risk science, and he keeps the science. Most critics of AI-risk framing come from ethics or STS; most defenders of quantification do not reframe.
  2. Creativity as a risk competence. From “risk entrepreneurship” and seventeen haiku (2015) to “not thinking creatively enough” about AI (FFTF p.174), he treats failure of imagination as a cause of harm.
  3. Risks and benefits in one frame. Future value and forgone benefit sit inside the risk account. That lets him refuse both “zero exposure” and “go fast and break things” on the same grounds.
  4. A stance between accelerate and prevent. Navigation is his answer to “safety absolutists and the move-fast-and-break-things crowd” (2026-04-12 What-Thirty-Years), and “There will be no single silver bullet” (2026-04-12 What-Nanotechnology).
  5. Risk explained by structure, not villains. “sincerity almost always operates inside an incentive field” (2026-07-16 [mixed]). He explains orphaning without accusing anyone, which keeps builders in the conversation.
  6. Humility that still acts. He is uncertain about formulating the problem, yet firm that risks must be talked about and navigated.

An independent assessment of the framings for AI#

Setting aside how much weight he himself gives each idea, here is how they fare when applied to AI.

Risk as a threat to value is well suited to AI. Many of AI’s most discussed harms have no clean dose–response: dependence on companion systems, erosion of trust inside organisations, drift in epistemic agency, the flattening of identity he calls “LinkedInification” (2026-03-08 ai-linkedinification). A probability-of-harm frame has “little or nothing to run on” for these (2026-07-16 [mixed]). A value frame can at least name them, say who holds the value, and track threats over time. It also explains public backlash to AI as information, not ignorance. It counts forgone benefits, which the precaution debate usually leaves out. Its weaknesses matter more for AI than for startups. The actors whose value is most at stake (users, data workers, communities near data centres) have the least leverage to make their losses register. There is no rule for adding up many small, dispersed harms. And a regulator needs thresholds the frame does not supply. It is most useful as a lens for seeing and as a language for talking with builders, less so as a basis for binding decisions.

The landscape and navigation fit a technology that changes faster than evidence can be gathered. When capabilities shift within months, prediction-then-control methods are always behind, and a stance of mapping, steering and course correction is more honest. Chaos with bounds is a better model of AI’s trajectory than either the exponential or the plateau story. The limit is irreversibility. Navigation works where mistakes can be undone. For AI’s lock-in effects (defaults, dependence, institutional adoption) it needs fixed points: pre-committed triggers and some things one does not try. His own reversibility test and hysteresis curve supply these, and they deserve a place at the centre of the frame.

Orphan risks may be the most valuable of the framings for AI at present, because it describes an institutional blind spot rather than a list of hazards. AI safety practice selects for what is measurable, catastrophic, auditable and competitively affordable. That is his 2026 analysis [mixed], and the removal and later return of persuasion in one company’s framework illustrates it. The concept directs attention to ownership and accountability, which is exactly where AI governance is thinnest. It is also testable: his three proposed tests could show it to be wrong. Its risks are that “orphan” becomes a catch-all label; that a register becomes a new box to tick; and that it says little about true unknowns, since orphans are by definition known to someone. It complements catastrophic-risk frameworks. It does not replace them, and he does not claim it does.

Humility against false precision is timely. AI discussion is full of confident numbers: benchmarks as proxies for capability, probabilities of doom, headline adoption rates. His insistence on asking what a number measures, and whether it matters, is a corrective. It needs pairing with a decision rule for acting under uncertainty. The pieces are in his own record (“quick to question, and slow to respond”; flexible trigger points; informed speculation “within a context of humility”), but he has not yet assembled them for AI.

Overall. For AI, these framings work best as a second lens alongside capability-based safety and legal compliance. Their distinctive contribution is to widen what can be seen, to put benefit and harm in one conversation, and to keep that conversation open with the people building the technology. That is what he says they are for: ideas “designed to open up new ideas and possibilities” (2016-01-11), and a framing “that opens up new possibilities rather than closing down conversations” (2023-05-31).