S1. The core nanotechnology-lessons papers: what they add to the map#

Supplementary reading for 05-maynard-risk-and-ai-map.md. Share S1: eleven papers, reports and commentaries from 2005 to 2023 on nanomaterial risk, nano-safety research strategy, nano regulation and the lessons of nanotechnology. Prepared 26 September 2026. The map was read but not edited. This note covers Andrew Maynard’s own thinking only. It makes no comparison with Jensen Huang or with the content of the EEA Late Lessons reports (where his own papers discuss those reports, what they say is reported here as his reading).

How to read this note#

• Order. The items are taken chronologically (2005 to 2023), because the main value of this share is that it documents the formative period, which the map knows only through later posts (map §1 “Limits”; §7 phase 0).
• Citations. Items are cited by filename (paths relative to Resources/maynard-papers/) and page:
• journal pages where the PDF carries them (2006 Nature pp.267–269; 2008 Nat. Nanotechnol. pp.444–447; 2011 Nature p.31; 2011 Nat. Mater. pp.554–557; 2016 Nat. Nanotechnol. pp.998–1000; 2007 Ann. Occup. Hyg. pp.1–12);
• “p. X of 35” for the 2005 Particle and Fibre Toxicology paper;
• report pages for PEN 3 (report page = PDF page − 2, so both are given);
• PDF pages for the 2023 comment, which carries no journal pagination;
• section headings for the two Markdown conversions (2011 Toxicol. Sci.; 2014 Health, Risk & Society), which have no page numbers.
• Quotations are short and exact. Everything else is paraphrase.
• Map references. C1–C17 are the map’s commitments (§4), T1–T11 its threads (§6), and “tension N” its §8 tensions.
• “My reading” marks interpretation, not report.

Provenance at a glance#

None of these items is AI-written or AI-assisted. Weight follows authorship.

1. Oberdörster, Maynard et al. (2005), Principles for characterizing the potential human health effects from exposure to nanomaterials: elements of a screening strategy#

File: papers/2005_Oberdorster-Maynard-et-al_Principles-Characterizing-Health-Effects-Nanomaterials_PFT.pdf (Part. Fibre Toxicol. 2:8). Skimmed as instructed. §4.1 and the conclusion were read closely; the in vitro and in vivo sections were read for framing.

Provenance. An ILSI Research Foundation/Risk Science Institute working-group report. It was funded by EPA’s Office of Pollution Prevention and Toxics, and its authors include DuPont and Procter & Gamble scientists. Maynard chaired the sub-group on physicochemical characteristics (p.30 of 35), so §4.1 (“Physicochemical Characterization”, pp.7–13) is the part most attributable to him. The rest reflects a toxicology consensus.

The argument. The report sets out elements of a screening strategy for hazard identification, “the first step” in the risk-assessment process (p.2). It treats nanomaterials as potential hazards whose “direct risk” depends on “the probability of exposures occurring” (p.6). It argues that biological activity may depend on physicochemical parameters “not routinely considered” in toxicity testing (p.2). Its core methodological move, in his section, is to design measurement around ignorance: - measure or record every potentially significant characteristic; - characterise the material as administered, not just as supplied; - collect enough data to allow “retrospective interpretation of toxicity data in the light of new findings” (p.7); - because it is not known whether mass, surface area or particle number is the relevant dose metric, make all three “measured or derivable” in every study (p.29; recommendation 5, p.13).

It also accepts limits: characterising “every possible aspect” is “impractical” (p.7), so it sets priorities of essential, desirable and optional measurements (Tables 1–2, pp.10–11).

Key concepts. Hazard identification as step one of quantitative risk assessment; hazard versus risk via exposure probability; dose metrics (mass, surface area, number); characterisation at several points (as supplied, as administered, in situ); benchmark control particles; tiered testing; recording data for later reinterpretation; prioritising measurement by importance and feasibility.

What this adds to the map. - Deeper evidence (C4, T8). It documents the technical core of the “professional grammar” the map describes in prose (§2; C4): hazard, exposure, dose and characterisation as a formal, quantitative discipline, not a figure of speech. - New idea: measurement designed for ignorance. When the right metric is unknown, measure several, keep records that can be reinterpreted, and prioritise by feasibility. The map has nothing on this. My reading: it is the methodological ancestor of his later instinct that we lack even “the beginnings of a framework” for AI exposure (2023-11-26). In 2005 he could hedge across three candidate metrics; for AI there was not yet a candidate set to hedge across. - Bearing on Andrew’s notes. - Note 1: his quantitative foundation is real and technical. - Note 2: even at its most quantitative, his practice built humility into method. Numbers were gathered on the assumption that today’s interpretation may be wrong.

2. Maynard (2006), Nanotechnology: A Research Strategy for Addressing Risk (PEN 3)#

File: papers/2006_Nanotechnology-Research-Strategy-for-Addressing-Risk_PEN-3.pdf. Read in full.

Provenance. Sole-authored, written as Chief Science Advisor to the Project on Emerging Nanotechnologies (Wilson Center and Pew). He frames it as “one scientist’s personal perspective” meant to stimulate discussion (p.9 / PDF 11), and calls his action plan “a perspective to stimulate dialogue” (p.34 / PDF 36). The foreword (pp.3–4 / PDF 5–6) is David Rejeski’s and is excluded. This is advocacy from inside an institution, which matters for tension 12 (honest broker and advocate).

The argument. - Some nanomaterials will present risks unlike any before, but “Saying they will be different does not mean they will be more threatening” (p.8 / PDF 10). He dismisses “grey goo” in the same passage. - Existing knowledge is not enough: quantifying nano risks from it “will engender false assumptions of safety” (p.13 / PDF 15). - US risk research is small and misdirected. He estimates that highly relevant research is about 1% of the NNI budget, around $11 million a year (pp.18–20), and shows research is out of step with the market (Figure 3). His examples: 24 lung projects and none on the gut, although many products are eaten. He asks whether the emphasis reflects relative risk or “because pulmonary toxicologists are more active in this field?” (p.21 / PDF 23). - The NNI’s promotional mission makes risk “a box to be checked on the road to realizing nanotech’s considerable benefits” (p.15 / PDF 17). Its risk plan is “more a thought exercise than an actual plan” (p.17). - Industry cannot lead the research, because it has “an economic incentive to sell products”, may not publish, and its findings “might be considered suspect” (p.32 / PDF 34). - His remedies: - a government-led strategic framework with “top-down, authoritative oversight” of research (p.5 / PDF 7) that “must have teeth” (p.33); - a new interagency group with budget authority; - $100 million over two years; - joint funding with industry on the Health Effects Institute model (p.38); - international coordination; - a rolling independent National Academies assessment (p.41). - Two prioritising objectives. “Oversight” means protecting people “in the absence of complete information, while not unnecessarily stifling innovation” (p.27 / PDF 29). “New Knowledge” is the second. He also calls for “Research … into how to do research” (p.27). - A 10-year priority heat-map (Table 4), labelled a “personal evaluation” and “highly subjective” beyond 2010 (p.29 / PDF 31). - Long-term needs include “emergent behavior and convergence between different technologies” (p.28 / PDF 30). - “Safe” is defined in a footnote as “a relative term” (p.9). - Risk research serves innovation: anyone who wants nanotechnology to succeed “has a vested interest” in it (p.9). - No documented harm “could be misleading, as appropriate surveillance has not been in place” (p.14 / PDF 16).

Key concepts. Strategic, targeted risk research; the gap between research and market; the conflict between promoting a technology and overseeing its risks; the independence of research; the catch-up problem (“risk research is playing a game of catch-up”, p.28); oversight under incomplete information; research on research methods; absence of evidence is not evidence of absence; “safe” as relative; a “high stakes bet” by governments and industry (p.14).

What this adds to the map. - Earlier origins. Several ideas the map dates later appear here in 2006: - safety as relative (map: 2016-03-31, 2024-06-20); - the pacing or catch-up problem (map: 2016-04-01 “pacing gap”); - structural incentives working against risk research (map: from 2022-02-12); - the case that industry cannot govern its own risks, because of incentive and credibility (map: 2023, “Industry can’t get AI governance right on its own”); - risk research as the friend of innovation (map C9, from 2016); - the need to innovate in risk methods, “how to do research” (map C3, founded 2016-01-11); - convergence and emergent behaviour as long-term risk issues (map: 2015-01-30). - New idea: the promoter–overseer conflict. An agency that promotes a technology should not be relied on to scrutinise its risks. It is his institutional form of “who is certifying that this is responsible” (lens 8). It recurs in items 5 and 8. The map has capture (2023-10-30) but not this older, structural version. - Qualification (C12; tension 9). His first governance remedy was top-down, authoritative and government-led, run by experts, with a funded independent research institute. Public engagement barely appears; the public figures mainly as “public confidence”. So the map’s picture of a soft-law, agile, multi-stakeholder preference (C12) describes a later position. His formative instinct was for authority with “teeth”, independent expert assessment and co-funded independent science. This supports the map’s reading (tension 9) that his concrete proposals put experts at the centre, and gives it an early origin. - Qualification (tension 12). The honest-broker role he adopts in 2018 (FFTF p.246) came after years of open, institution-based policy advocacy for more funding and a new structure. My reading: advocacy is not a late drift from an honest-broker baseline. It was his starting mode, later disciplined. - Bearing on Andrew’s notes. - Note 2, directly: the sentence on quantifying from existing knowledge is the earliest explicit statement in this share of the concern that numbers can give false comfort when understanding is missing. - Note 1: he still used numbers, but bounded ones: - funding percentages; - a mass-equivalence calculation (58,000 tonnes a year of nanomaterials “might be the equivalent of between 5 million and 50 billion metric tonnes” of conventional material if number or surface area drives hazard, p.10 / PDF 12); - a heat-map he labels subjective. - Humility is present too. He labels his projections, calls the plan a perspective, and asks for rolling review.

3. Maynard et al. (2006), “Safe handling of nanotechnology”, Nature 444: 267–269#

File: papers/2006_Safe-Handling-of-Nanotechnology_Nature.pdf. Read in full; the extracted text is letter-spaced but legible.

Provenance. Fourteen research leaders (Aitken, Donaldson, Oberdörster, Philbert, Seaton, Stone, Warheit and others), with Maynard as lead author (“argue Andrew D. Maynard and his co-authors”, p.267). A consensus manifesto; weight moderate.

The argument. “Fears … may be exaggerated, but they are not necessarily unfounded” (p.267). The “spectre of possible harm — whether real or imagined” could slow nanotechnology. And “the way science is done is often ill-equipped to address novel risks” (p.267), because risk research has “a low priority in the competitive worlds of intellectual property, research funding and technology development” (p.267). The paper sets five grand challenges with 3- to 15-year timelines: 1. exposure instruments, including a cheap universal aerosol sampler logging number, surface area and mass, and “smart sensors” that indicate harm rather than just exposure; 2. validated toxicity screening, including the urgent question of whether fibre-shaped nanomaterials behave like asbestos; 3. predictive models, leading to materials that are “safe by design”; 4. life-cycle evaluation, cradle to grave; 5. strategic research programmes built on collaboration, communication and coordination.

The sampler would give “a historic record that can be interpreted in the light of new knowledge” (p.268), because “We don’t yet know which aspects of airborne nanomaterials should be measured” (p.268). End users of risk data (industry, consumers and policymakers) “must play a central role in shaping what is done and how” (p.269). Developing economies should not be “denied essential information” on safe design (p.269). Missing an asbestos-like hazard would be “potentially devastating” both to exposed people and to the industry (p.268).

Key concepts. Grand challenges; smart sensors; safe by design; predictive toxicology; life-cycle thinking; records kept for later reinterpretation; risk communication over the internet; justice for developing economies; the fibre paradigm as the model case.

What this adds to the map. - Literal transfer, clearly marked. The asbestos fibre paradigm, and quartz particles whose harm is reduced by a clay coating (p.267), are carried over as mechanisms to new materials. This is literal transfer where the mechanism (physical form plus chemistry) is expected to recur. It is the model for his later rule of behaviour, not label (C13). - Earlier origins. Structural incentives against risk work (2006; map from 2022). Justice framed globally (developing economies), where the map’s justice evidence begins with the occupational-health story in FFTF (C11). - Deeper evidence (C9). Risk research is presented as what makes a “sustainable” industry possible, and litigation and insurance fears as threats to its value (p.267). This is an early, enterprise-facing form of the later threat-to-value frame (compare tension 4). - Bearing on Andrew’s notes. The paper pairs a strongly quantitative agenda with an explicit admission of not knowing what to measure. Its answer is instruments that keep options open. Note 2 is foreshadowed as method: numbers are to be gathered without pretending to know what they mean. Item 10 shows how far this agenda was realised.

4. Maynard (2007), “Nanotechnology: the next big thing, or much ado about nothing?”, Ann. Occup. Hyg. 51(1): 1–12#

File: papers/2007_Nanotechnology-Next-Big-Thing-or-Much-Ado-About-Nothing_Ann-Occup-Hyg.pdf. Read in full.

Provenance. Sole author; the published form of his 2006 Warner Lecture to the British Occupational Hygiene Society. High weight. This is his clearest statement of the move from occupational-hygiene risk assessment to an emerging technology.

The argument. - Framing: - hype runs both ways, on promise and on fear; - the “power of people to decide which technologies succeed”, on real or perceived risk, is now a significant factor (p.1); - occupational hygienists protect “the first line of people to face possible risks” (p.2). - Nanotechnology is best seen as “a way of thinking or doing things, than a discrete technology”, which makes general talk about its risks hard (p.3). Grey goo fears “would seem to be unfounded” (p.3). - He states a hypothesis, that hazard depends on physical and chemical structure, and insists it be tested (“Without validation, it is little more than an interesting diversion”, p.5). He then weighs the evidence for and against, including a contradictory study (p.7). - He extends the conventional paradigm. Risk is a function of hazard and exposure, plus “a third component … Characterization” (p.7). - Under metric uncertainty he proposes, quantitatively, an instrument whose response “reflects current uncertainty over what should be measured” (p.8). Hazard potential is modelled as proportional to particle diameter to the power α, and α is averaged across the candidate metrics. Dropping number concentration as implausible gives α = 2, a surface-area response (pp.8–9, Fig. 6). - He then asks how to set levels of control when there is “insufficient information available for a quantitative risk assessment” (p.9). His answer: the range of responses runs between precaution (hazardous until proven otherwise) and inaction (negligible until proven otherwise) (pp.9–10), and he proposes control banding. Control banding is not “a substitute for conventional risk assessment and control” but a pragmatic tool for “decision-making based on incomplete information” (p.10). The pharmaceutical and COSHH scheme “is not directly applicable to engineered nanomaterials. But the concept is.” (p.10) - The summary: “push existing knowledge as far as it will go”, and where it fails, do targeted research (p.11).

Key concepts. Hazard + exposure + characterisation; hypothesis and weight of evidence; an instrument response hedged against ignorance; control banding (an impact index against an exposure index); decisions under incomplete information; precaution and inaction as the two ends of a spectrum; “a way of thinking” rather than a discrete technology.

What this adds to the map. - New idea: the explicit literal–conceptual distinction. “Not directly applicable … But the concept is” (p.10) is the clearest statement in his record of how he transfers tools between domains: the mechanism of a tool may not carry over, but its logic can. This speaks directly to tension 1, where the map says of his transfer rule that “He never states the rule”. For tools, he does state it here, in 2007. (For AI he does not.) - New idea: quantification that admits ignorance. The α-averaging instrument is quantitative method that keeps the unknown in plain view. It neither drops numbers nor claims precision. - Qualification (C6). His position on precaution was already set in 2007, in sole-authored prose: neither pole. The middle ground “will require a shift in perspective on how risk is evaluated and managed” (p.10). That is an early form of risk innovation (C3), arising from a quantitative problem. - Earlier origins. Plausibility applied to grey goo (2007; map FFTF 2018). The “first tier” or “first line” of exposed workers as a justice concern (p.2; map FFTF pp.118–122). Public power over technology’s success (p.1; map T1 on perception). - Bearing on Andrew’s notes. - Note 1: strong. Conventional risk assessment is the base, extended by characterisation and supplemented by banding: explicitly not a substitute. - Note 2: strong. When the data for quantitative risk assessment do not exist, he neither waits nor invents precision. He builds a decision tool for the gap, and he calls the uncertainty “overwhelming” (p.10). - The other direction: even here he reached for a formula (α). His instinct under uncertainty was still to structure the unknown quantitatively where a physical model allowed it. My reading: the lack of AI numbers reflects the lack of any agreed physical or causal model to quantify, not a distaste for quantification.

5. Hansen, Maynard, Baun & Tickner (2008), “Late lessons from early warnings for nanotechnology”, Nat. Nanotechnol. 3: 444–447#

File: Maynard supplied/nnano.2008.198.pdf. Read in full.

Provenance. Four authors. The lead is Steffen Foss Hansen (DTU); Joel Tickner (UMass Lowell) is a leading advocate of precaution. Maynard is second. Shared positions, and probably the most precaution-leaning text in his record, so weight moderate. Box 1 reproduces the EEA’s twelve lessons; those words are the EEA’s, not his. A companion blog post of his exists (web/2008_2020science_Late-Lessons-from-Early-Warnings.md, outside this share) and could confirm which emphases are his.

The argument. The paper tests nanotechnology against the twelve lessons of the 2001 EEA report and gives a mixed verdict. - Positives: - early risk discussion is unusually prominent; - critical questions are being asked early; - collaborations cross disciplinary boundaries; - research is beginning to be targeted; - stakeholders are being engaged; - existing oversight is being questioned (p.447). - Failures: - the “global response to these warning signs has been patchy” (p.444); - disciplinary blinkers, such as EPA’s chemistry-bound world view, which defines “new” by molecular identity while size and shape change behaviour (p.445); - reliance on idealised assumptions of sealed, small-scale processes, against the grimy reality of nano workplaces (p.445); - the NNI both promotes and oversees risk. “Perhaps more insidiously”, R&D decisions follow what promotes the technology rather than what protects people (p.446); - “many governments still call for more information as a substitute for action” (p.446). - Specific claims: - lay and worker knowledge matters: those who make and use a product often “have some of the clearest ideas about what is important” (p.445); - “citizens around the world are as much stakeholders” as governments and industry, but their engagement “has been very limited” (p.446); - on alternatives: nanotechnology could be used, but “it will be questionable whether it should be used” (p.446); - on benefit claims: if promised benefits fail to materialise, or if feared harms are not investigated and later prove real, trust and decision-making suffer (p.446); - act on what is known now, with “review procedures for course corrections” (p.446); - build safety in at the design stage, “because economic interests are not fully entrenched at that point” (p.447). - Conclusion: the question is not whether the lessons were learned “but whether we are applying them effectively enough” (p.447).

Key concepts. Early warnings; lessons learned versus lessons applied; institutional ignorance and disciplinary blinkers; real-world conditions against idealised assumptions; lay knowledge; independence of regulators; alternatives to the favoured technology; paralysis by analysis; design-stage intervention before interests entrench; symmetric scrutiny of benefit and risk claims.

What this adds to the map. - Earlier origins. - Everyone a stakeholder, and lay expertise (map: FFTF p.222, 288; 2023-05-15), here in 2008. - Could vs should (map: FFTF pp.36–39), here in 2008 applied to technology choice. - Plausibility applied to benefit claims as well as risk claims (C7 “hype and doom alike”), here as scrutiny of claimed “pros”. - New idea: the timing logic. Act before economic interests entrench (p.447); act on what is known with built-in course correction. This is a Collingridge-style argument (my label†, not his). The map has lock-in only in education (2024-05-05) and “early engagement” as process (C13). - Deeper evidence (C13, T2). This is the first sustained instance in his record of reasoning from a set of historical cases to an emerging technology. It says openly that some lessons “are not directly applicable to emerging technologies” while “many … are directly relevant” (p.447). That is selective, conceptual transfer (see §12). - Qualification (for the next stage). The 2008 verdict on nanotechnology was contemporaneous and fairly critical: “distracted”, with promoters overseeing risk, research not answering critical questions, stakeholders not fully engaged (p.447). His 2023 retrospective (item 11) calls the nano transition “reasonably successful”. When he offers nano as a model for AI, the contemporaneous record says the model included real failures. My reading: the 2023 account is a smoothed retrospective, not a contradiction, but the map should not read “nano lessons” as a success story alone. - Bearing on Andrew’s notes. On note 2 from the action side: humility about knowledge does not justify waiting. “More information as a substitute for action” is named as a failure. This matches his note that humility goes together with the need “to grapple with emerging issues”.

6. Maynard, Warheit & Philbert (2011), “The new toxicology of sophisticated materials: nanotoxicology and beyond”, Toxicol. Sci. 120(S1): S109–S129#

File: papers/2011_Maynard-Warheit-Philbert_New-Toxicology-of-Sophisticated-Materials_PMC.md. Read in full. Cited by section, since the file has no page numbers.

Provenance. Lead author, with David Warheit (DuPont) and Martin Philbert (Michigan). A review with shared positions. The problem-formulation and “Looking to the future” sections align closely with his sole-authored 2011 Nature Comment (item 7) and so can be weighted as his.

The argument. Nanotoxicology has “something of an identity crisis” (introduction): - its field-defining size range rests on “definitions of convenience, not of science”; - many nanomaterial toxicities are scalable, and so predictable, from larger materials; - “History suggests that not every new technology leads to new hazards and not every new hazard is associated with a new technology.”

What is needed is “a differential approach to toxicology”, focused on where new materials deviate from established ones. Risk questions should be decoupled from the technology label: decoupling the materials from the technology “is helpful in formulating science-based questions”.

The real challenge is “sophisticated materials”, whose form and chemistry interact synergistically and whose behaviour may be active, self-assembling, context-dependent or signal-driven. Future materials will “resemble the complexity of human-scale engineered devices” rather than simple chemicals, which demands a “systems-based approach” (Looking to the future).

Because “Effective problem formulation is a cornerstone of contemporary risk assessment”, the authors propose three “technology independent” principles for deciding what to study (Identifying Relevant Materials): - emergent risk: harm “not apparent, assessable, or manageable based on current approaches”; - plausibility: a qualitative, science-informed likelihood, “a crude but effective filter to distinguish between speculative risks—which are legion—and credible risks—which are not”; grey goo is emergent but not plausible; - impact: a “qualitative reality check” on whether research would reduce harm.

On dose, the paper states that although “the dose makes the poison” still holds, “there is considerable uncertainty over what is meant by dose”. If hazard scales with a parameter other than the one measured, “the hazard will remain ill quantified”.

The closing section: - “the risk assessment paradigm remains relevant”, and decisions must link dose to response; - but “Quantitative toxicology and risk assessment are unlikely to keep pace” with sophisticated materials, so a knowledge gap will grow; - closing it needs new ways of deciding under incomplete information, life-cycle framing, predictive models, pushing risk assessment “upstream in the innovation process”, and “a new science of risk” drawing on the physical, biological and social sciences.

Key concepts. Sophisticated materials; the differential approach; definitions of convenience; decoupling hazard from technology branding; problem formulation; emergent risk, plausibility and impact; scalable versus non-scalable effects; “ill quantified” hazard; active and self-assembling materials; systems-based assessment; a new science of risk; assessment moved upstream; a “fleeting opportunity” to share pre-competitive hazard data (Biointeractions).

What this adds to the map. This is the most important item in the share for the map. - Earlier origin (C7): plausibility as a named, formal principle, in 2011. The map dates the plausibility discipline as “formalised in FFTF (2018)” (§3 table) and treats it as core (C7). Here it is already a defined filter separating speculative from credible risks, with grey goo as the example, seven years before the book. It is also openly qualitative. - Earlier origin (orphan risks, C3). “Emergent risk”, defined as harm not apparent, assessable or manageable with current approaches, is the conceptual precursor of orphan risks (2018) and of “outmoded ideas about risk are a risk” (2020). It names the class of risk that conventional tools miss. - Partial correction to tension 1. The paper offers an explicit transfer rule, although for materials rather than AI: - use technology-independent principles; - decouple the risk question from the technology’s name; - take a differential approach that concentrates effort where the new departs from the known; - expect neither that new technology means new hazard nor the reverse.

The map’s “behaviour, not labels” (2022-02-10) and its “analogy as mindset” reading (C13) have a clear, formal 2011 statement. The map should say he stated the rule for materials in 2011, and did not restate it for AI. - New idea: hazard misread through the wrong metric. Quantifying against the conventional measure (mass) when harm tracks another (surface area) leaves hazard “ill quantified”. My reading: this is the materials-science form of Andrew’s note 2. A number can be precise and still measure the wrong thing. - New idea: technologies of dynamic behaviour. Active, signal-responsive, self-assembling materials whose behaviour depends on context, which he expected to strain existing paradigms. My reading: this is the category his later AI concerns fit (stochastic, context-dependent, emergent behaviour; map 5.8). The link is not his. - Bearing on Andrew’s notes. - Note 1, directly and strongly: “the risk assessment paradigm remains relevant.” The call for “a new science of risk” is framed as bridging a pace gap on top of the paradigm, not replacing it. - Note 2: this is the earliest text in the share to say quantitative methods structurally cannot keep up with innovation of this kind. It prefers qualitative filters (plausibility, impact) for problem formulation when data are missing.

7. Maynard (2011), “Don’t define nanomaterials”, Nature 475: 31#

File: Maynard supplied/475031a.pdf. Read in full.

Provenance. Sole author, Nature Comment, written as director of Michigan’s Risk Science Center. High weight.

The argument. “Five years ago, I was a proponent of a regulatory definition of engineered nanomaterials. I have changed my mind.” Risk depends on many parameters (size, shape, porosity, surface area, chemistry) in ways that vary by material and context, so a “‘one size fits all’ definition” fails. A policy-driven definition would make regulation “a ‘term of art’ rather than science”. His precedent: Libby vermiculite, which contained asbestiform fibres yet escaped regulation because it did not fit the official definition of asbestos. Instead he proposes: - nine or ten attributes with material-specific “trigger points” for action, based on current science; - triggers that allow for change in a material over time; - the unresolved question of how much change should trigger concern (“a 1% change … or … a 50% change?”); - “enough is known today” for an expert panel to begin; - trigger points that “must be flexible, so that they can be modified as evidence grows”, even though US law makes adaptive regulation hard.

The deeper error is “assuming that nanomaterials are a unique class”.

Key concepts. Behaviour over definitions; the danger of a “term of art”; regulatory trigger points; adaptive, evidence-responsive regulation; categories that let hazards slip through (Libby vermiculite).

What this adds to the map. - Correction to §7 “Changes of mind”. A signalled reversal in 2011, in his own words. The map’s table of reversals he names begins in 2020. Adding this shows that open reversal is a habit of long standing (map “His method”: humility tested in public). - Earlier origin (C13). “Nature doesn’t care what we call a material” (2022-02-10) is the later, popular form of this 2011 argument. It also anticipates his distrust of category labels in AI debates (for example AGI as “rather ill-defined”, 2026-04-11). That link is my reading. - New idea: adaptive regulatory triggers. Adaptive regulation is proposed here as hard law with evidence-tuned thresholds, not as soft law. This qualifies C12 (a preference for soft law). In 2011 he wanted adaptive precision inside regulation. - Bearing on Andrew’s notes. - Note 1: the triggers are quantitative thresholds grounded in current science, a direct continuation of quantitative risk assessment. - Note 2: the thresholds are provisional by design, to be revised as evidence grows. The unanswered “1% or 50%” question is left open rather than faked.

8. Maynard, Bowman & Hodge (2011), “The problem of regulating sophisticated materials”, Nat. Mater. 10: 554–557#

File: Maynard supplied/nmat3085.pdf. Read in full.

Provenance. Lead author, with Diana Bowman and Graeme Hodge (regulation and law scholars, Michigan and Monash). Shared positions; moderate to high weight.

The argument. Nanomaterial regulation is a “‘wicked’ public policy problem”: stakeholders cannot even agree what the problem is. Some define it as worker and consumer toxicity, some as long-term environmental release, some as corporate power, intellectual property and the lack of citizen engagement (p.554). Hype and speculation cloud it. Regulation “cannot afford to be based on imagined futures that are only weakly connected to reality”, yet the nanotechnology “brand” was “not necessarily … detrimental”, since it raised profile and funding (p.554). The dialogue should be decoupled from speculative visions, informed by “plausible emergent risks”, and “grounded in established approaches to identifying, assessing and managing risks” (pp.554–555).

The paper sets out seven challenges: - moving past the “language game”, in which proponents and opponents bring in ideologies “under the cover of slippery language” (p.555); - keeping oversight language separate from promotional language; - filling knowledge gaps; - setting standards that are fit for regulation rather than for development; - identifying regulatory gaps and triggers through foresight aimed at “plausible emerging risks”, not “ill-defined imagined futures” (p.555); - balancing innovation and safety, which are “inextricably intertwined” (pp.555–556), where safety is often “merely an add-on” (p.556) and governments hold dual roles of promotion and oversight; - moving forward with caution: regulatory evolution is essential, but “we would be remiss in throwing out the old and embracing the new, simply because we can” (p.556).

It adds transparency and trust. Regulation has “typically been built on quantitative risk assessment — the purview of experts — and quietly modulated by political and economic interests”. It has been competent but “has tended to deal retrospectively with well-established risks”, while citizens now expect to challenge decisions and need “the tools to engage effectively” (p.556). Success depends on a “legitimate social licence” (pp.556–557).

Key concepts. Wicked problems; the language game; plausible emergent risks against imagined futures; oversight separated from promotion; standards fit for regulatory purpose; regulatory gaps and triggers; safety and innovation intertwined; dual roles; caution about new regulatory forms; quantitative risk assessment as expert and retrospective; social licence; citizen tools.

What this adds to the map. - Earlier origins. - Social licence (map: 2016-03-12, 2017-04-10, 2018-09-03), here in 2011. - “Can we even formulate the problem?” (map open question, 2023-11-26 and 2025-05-04), here as a named wicked problem in 2011. - Plausible against imagined futures (C7). - Distributed, “more responsive” regulation (C12), qualified by caution. - New idea: the language game. Slippery definitions and imagined futures are used by both sides to import ideology. The map has “stories as risks” and hype, but not this governance-level point that contested terms are instruments. My reading: it is directly usable on contested AI vocabulary (“safe”, “aligned”, “AGI”). - Qualification (note 1; C3). His critique of quantitative risk assessment is precise. It is not that the method is wrong. It is that it has been expert-owned, “quietly” shaped by interests, and backward-looking. The remedy is to add legitimacy and foresight, not to remove the quantitative base: “grounded in established approaches” and “remiss in throwing out the old”. This is perhaps the clearest single text for Andrew’s first note. - Deeper evidence (C5). Who decides: citizens need tools to engage with government and industry decision-makers, and quantitative risk assessment as the “purview of experts” is part of the legitimacy problem.

9. Scherer, Maynard, Dolinoy, Fagerlin & Zikmund-Fisher (2014), “The psychology of ‘regrettable substitutions’”, Health, Risk & Society 16(7–8): 649–666#

File: papers/2014_Scherer-et-al_Psychology-of-Regrettable-Substitutions_PMC-manuscript.md. Read in full. Cited by section.

Provenance. An empirical decision-psychology study led by Laura Scherer, with Maynard second of five. Weight low to moderate; it shows what he co-produced and endorsed, not his distinctive voice.

The argument. Two online experiments (n = 1,738 and 777) on how people trade a contested known risk (BPA) against an unstudied substitute (PET): - There was no ambiguity aversion. People rated whichever chemical they read about first as less risky; order shifted choices for almost a quarter of participants (Study 1 findings). - Labelling the substitute “BPA-Free” reversed preferences even when people knew it contained an unstudied chemical (Study 2). - Contested evidence of harm was treated “similarly to a situation in which there is no scientific evidence at all” (Discussion of Study 1). - Risk perceptions are “less a hard, permanent truth and more a momentary perception” shaped by framing (Implications). This may also explain why policy responses diverge on the same evidence. - Labels should state exactly what was replaced.

Key concepts. Regrettable substitution; risk against ambiguity (second-order uncertainty); order effects; “free-of” labels; contested evidence treated as no evidence; malleable perception.

What this adds to the map. - New idea: regrettable substitution. Swapping a known, contested risk for an unknown one because the swap is framed as removal. This is absent from the map. My reading: it is a usable lens for AI (replacing human judgement, relationships or institutions with systems whose risks are unstudied, sold as removing the old problem’s flaws). He does not make this application. - Qualification (C9; T1 “perception”). The map says public concern is “a signal, not noise”. This co-authored study shows specific risk judgements can be artefacts of order and wording. The two are compatible: concern signals threatened value, but the particular form it takes is malleable. My reading: this is also early empirical ground for the map’s C14 premise that human cognition is exploitable, from the risk-communication side rather than the AI side. - Bearing on Andrew’s notes. - Note 2: the study shows controversy dissolving evidence in lay judgement. My reading: this is a reason why precise-looking but contested numbers do not settle risk questions for the public. - Note 1: the work itself is quantitative behavioural science. It is further evidence that quantitative methods remained part of his toolkit in 2014, applied to perception rather than to toxicity.

10. Maynard & Aitken (2016), “‘Safe handling of nanotechnology’ ten years on”, Nat. Nanotechnol. 11: 998–1000#

File: Maynard supplied/nnano.2016.270.pdf. Read in full.

Provenance. Lead author with Robert Aitken, two of the 2006 authors. High weight. It is a public audit of his own 2006 agenda.

The argument. The paper reviews the 2006 grand challenges using Scopus publication counts, described as “indicative of trends rather than absolute” (Fig. 1, p.998), and a table titled “A personal assessment of progress” (p.999). - Progress: - strategic programmes and funding (€261M in FP7, US$830M in US EHS research, 2006–15); - high-throughput toxicity screening; - an understanding of fibre-like nanotubes; - environmental fate models. - Little or none: - exposure instruments; - “smart sensors” (no real progress); - predictive models of behaviour in the body; - safe-by-design. - Lessons: - global coordination and cross-sector collaboration were learned; - conventional controls such as PPE and exhaust ventilation “also work effectively for nanomaterials” (p.999); - the portfolio became unbalanced, with toxicology “far outstripping exposure-based research”, and material choice did not track likely potency (p.999). - Calibration: some anticipated risks “may not be as high as was originally thought”, which shows “the process of science is working” (p.999). - A warning: as “careers and funding pathways are built around assumptions of substantial nanomaterial-specific risk”, pragmatic evidence-based decisions become harder (pp.999–1000). - The biggest takeaway: nanomaterials are “just one component in an increasingly complex matrix” of designed materials. The 2006 challenges still apply if “nanomaterials” is replaced with a general term (p.1000).

Key concepts. A public self-audit; balance in research portfolios; the gap between new and useful knowledge (“a disconnect between new knowledge and useful knowledge”, Table 1); risk estimates revised downward; vested interests in risk; the nano category dissolving into materials in general.

What this adds to the map. - Deeper evidence (method: humility tested in public). The map’s evidence for this habit is mostly 2020–26. This is a formal, quantified audit of his own earlier predictions and agenda. - New idea: hubris and lock-in can form around risk, not only around innovation. The warning about careers built on assumed risk is sharp. The map notes hubris applied to AI refusers by 2026 (5.2 “Hubris”). Here, in 2016, he applies the logic to his own risk-research community. - Qualification (tension 2). This is documented calibration in both directions. Some feared nano risks shrank. The fibre concern proved specific (“which materials present fibre-like risks, and under what circumstances”, Table 1). Existing controls sufficed. This is the empirical backing for his later reuse of “grey goo to real risks” (FFTF p.281; 2024-06-23) as a template, and for his wariness of both alarm and complacency. - Literal transfer vindicated in part. Occupational controls carried over directly, and the fibre paradigm held for a subset of materials. This is the nearest case in his record where the literal carry-over worked, and it is a limited one. - Bearing on Andrew’s notes. - Note 2: the most revealing text. His own 2006 quantitative agenda under-delivered where he most wanted it (measuring exposure, sensors that indicate harm, predictive models). Meanwhile toxicology output grew without always yielding “useful knowledge”. My reading: this is experience of the limits of confident research programmes, a likely root of the humility Andrew describes. He does not draw that link here. - Note 1: his conclusion still calls for “pragmatic, evidence-based decisions” and “evidence-based knowledge” (p.1000). The base is kept.

11. Maynard & Dudley (2023), “Navigating advanced technology transitions: using lessons from nanotechnology”, Nat. Nanotechnol. (Comment)#

File: Maynard supplied/s41565-023-01481-5.pdf. Read in full; PDF pp.1–3.

Provenance. Lead author with Sean Dudley (ASU Knowledge Enterprise). High weight. The map already uses its Conversation companion (2023-10-02). The paper adds the quantum and gene-editing material and some specific wording.

The argument. - ChatGPT revealed an advanced technology transition “few were prepared for” (p.1). - Nano offers “deeply relevant” insights. Nano itself moved from Drexler’s speculative vision to “a more prosaic and plausible set of ambitions” (p.1). Its programmes built on recombinant DNA, the Human Genome Project and “poorly handled” GMO commercialisation (p.2). The transition drew on arts, humanities and social sciences, public engagement and multi-stakeholder partnerships, so regulators could “place early guardrails” (p.2). The verdict: “reasonably successful” (p.2). - The transfer has largely stopped. It is limited for gene editing and “even more profound” for AI, where “many conversations are still being driven by technological experts with little understanding of the complexity” of the social, economic and political landscape (p.2). He concedes there is more engagement “behind the scenes” than public discourse suggests (p.2). - Quantum technologies raise surveillance concerns, including “mind-penetrating interrogation” via quantum-enabled magnetoencephalography, and access concentrated among the few (p.2). - Transformative technologies “always come with unintended and often hard-to-anticipate uses and consequences, and we often ignore or trivialize these at our peril” (p.3). This cites both EEA Late Lessons reports, refs 14–15. - There is a risk of missteps where progress is “driven by the need for speed, market success, and little else” (p.3). - Success “depends on human systems”: misjudge values, community dynamics or governance politics, and technologies can be “dead on arrival” (p.3). - Remedy: investment in understanding and navigating transitions, across disciplines and sectors. “[E]ach advanced technology transition will demand new ways” of working, and there remains “a tremendous amount to be learned” from nano (p.3).

Key concepts. Advanced technology transitions; transdisciplinary, multi-sector, multi-stakeholder partnership; early guardrails; plausibility as what turned the nano vision into a programme; success depends on human systems; speed and market as the drivers of missteps; access and justice (quantum); the mind as a target (quantum MEG).

What this adds to the map. - Deeper evidence (T2, T6; ATT from 2023). The paper presents ATT as the generalisation of nano’s lesson. Its reference to “even the possibility of artificial general intelligence” (p.1), listed among coming technologies, is consistent with the map’s calibration: possible, not dismissed, not centred. - Qualification. Set beside item 5, the 2023 account of nano as “reasonably successful” is kinder than his 2008 contemporaneous account. The lesson he draws for AI is about process (transdisciplinary, engaged, early), not about nano’s institutions, which he had criticised at the time for mixing promotion and oversight. - Deeper evidence (C14). Mind-penetrating quantum sensing sits with his neurotechnology and AI concerns about technologies that act on the mind (map 5.8). - Bearing on Andrew’s notes. The humility is explicit: each transition “will demand new ways”. The engagement is equally explicit, with “early and serious attention” called “vital” (p.2). No quantitative apparatus is offered for AI; the lessons offered are institutional.

12. Cross-cutting findings#

12.1 How he reasons from toxicology and material risk to new technologies: literal and conceptual transfer#

Across 2005–2023 the share shows three distinct modes, and he keeps them apart more explicitly than the map allows.

1. Literal transfer of a mechanism, where the physical behaviour recurs. - The asbestos fibre paradigm is applied to nanotubes (items 3, 6). - Quartz, TiO₂ and surface area are applied to nanoparticles (items 1, 4, 6). - Occupational controls (PPE, local exhaust ventilation) are applied to nanomaterials (items 4, 10).

He treats this as a hypothesis to validate (item 4, p.5), and the 2016 audit shows it held only for subsets. This is the only mode that depends on shared mechanism. 2. Conceptual transfer of a tool’s logic, where the tool itself does not fit. Control banding “is not directly applicable … But the concept is” (item 4, p.10). Late Lessons: some lessons “are not directly applicable to emerging technologies”, many “are directly relevant” (item 5, p.447). 3. Technology-independent principles for problem formulation. Emergent risk, plausibility and impact; decoupling risk questions from technology labels; a differential focus on where the new departs from the known; neither new technology nor new hazard implies the other (item 6; item 7).

This is the mode he carries to AI (item 11): process and governance lessons, not hazard analogies.

Implication for the map (tension 1; C13). The map says he “never states the rule” for what transfers. For materials he stated it clearly in 2007 and 2011, and labelled the principles “technology independent”. What he has not done is restate it for AI. My reading: “defies analogy” (2026-01-22) is consistent with the 2011 framework. On that framework AI is an extreme case of emergent risk: harm not assessable with current approaches, where mode 1 fails, mode 2 is strained, and only mode 3 survives. On this reading his use of nano for AI is not in tension with discontinuity. It is the 2011 method applied to a case where the differential is very large.

12.2 His quantitative risk-assessment foundations and what he built on them (Andrew’s two notes)#

Note 1: quantitative risk assessment is foundation, built on. The share supports this strongly, from his sole-authored and lead-authored texts: - 2005: dose metrics, characterisation and tiered testing, with hazard identification as “the first step” of risk assessment (item 1). - 2007: risk as hazard × exposure, extended with characterisation. Control banding is “not a substitute” for conventional assessment (item 4, p.10). - 2011: regulation to be “grounded in established approaches”; “remiss in throwing out the old” (item 8, pp.555–556); “the risk assessment paradigm remains relevant” (item 6). - 2011: science-based quantitative trigger points (item 7). - 2016: “pragmatic, evidence-based decisions” (item 10).

The move toward “a new science of risk” (2011) and, later, risk innovation is presented as bridging a gap the paradigm cannot close alone, especially its pace. It is not presented as abandoning the paradigm.

Suggested map change: where the map describes a risk scientist who “lost patience” (§2) and C3’s “necessary but no longer sufficient”, add that the impatience was earned over a decade of trying to make quantitative assessment work for nanomaterials. The founding logic of risk innovation (item 6’s closing section; item 4’s “shift in perspective”) arose inside that quantitative work, from its own limits.

Note 2: humility about numbers, joined to the need to grapple. The share gives this an origin and a history: - quantifying from existing knowledge breeds “false assumptions of safety” (item 2, p.13); - hazard measured against the wrong metric stays “ill quantified” (item 6); - quantitative assessment is “unlikely to keep pace” (item 6); - expert-owned quantitative assessment has dealt “retrospectively” with established risks (item 8, p.556); - qualitative filters (plausibility, impact) are used deliberately where data are lacking (item 6); - measurement is designed around ignorance: all three metrics, archived records, the α-averaged instrument (items 1, 3, 4); - projections are labelled “highly subjective” (item 2); - the 2016 audit shows a confident research agenda under-delivering on exposure measurement and prediction (item 10).

The “grapple” half is equally consistent: - act “in the absence of complete information” (item 2, p.27); - control banding for decisions on incomplete information (item 4); - no “more information as a substitute for action” (item 5, p.446); - act before interests entrench (item 5, p.447).

The other direction (my reading). In nano, humility did not mean giving up numbers. Under deep uncertainty he still built bounded, heuristic quantification: the α exponent, the mass-equivalence range, funding shares, trigger thresholds. That suggests the absence of quantitative AI risk methods reflects a judgement that no defensible hazard model or dose metric yet exists (compare “the lack of even the beginnings of a framework”, 2023-11-26), not a principled rejection of quantification. It also suggests something he might endorse for AI but has not proposed: bounded, sensitivity-style quantification that makes ignorance visible (the α approach), rather than point estimates that hide it. The map’s §8 “Missing methods” entry could therefore be reframed from a gap to a considered restraint with a known alternative in his own past practice.

12.3 Earlier origins: ideas the map dates later#

12.4 Qualifications the map should consider#

1. Governance instincts (C12; tension 9). His formative remedies were top-down, authoritative, expert-led and government-funded, with “teeth” (2006), plus adaptive hard-law triggers (2011). Soft law and agile governance are a later layer, not the original disposition. Public engagement enters strongly from 2008 in co-authored work and 2011 (citizens given “tools”), and is thin in 2006.
2. Advocacy before honest brokerage (tension 12). PEN 3 is open policy advocacy from an institutional post. The 2018 honest-broker choice disciplined an existing advocate. It did not constrain a neutral scholar who later drifted.
3. The nano story is mixed (T2). Contemporaneous texts (2006, 2008, 2016) record real failures: 1% risk funding, promoters overseeing risk, paralysis by analysis, unbalanced portfolios, stalled exposure science. The 2023 “reasonably successful” is a retrospective judgement about process. Lessons drawn from nano should include its failures.
4. Precaution (C6). Already set in 2007, sole-authored: neither “hazardous until proven otherwise” nor “negligible until proven otherwise”. The 2008 co-authored paper is his most precaution-friendly text (with Tickner), and its emphasis is acting on what is known, with course correction.
5. Perception (C9). Concern signals value, but specific judgements are malleable and label-sensitive (2014, co-authored).

12.5 Lenses from this share for the next stage (stated neutrally, not applied)#

• Learned or applied? Are known lessons being applied, or only cited (item 5, p.447)?
• Promoter and overseer. Is the body that promotes a technology also the one judging its risks (items 2, 5, 8)?
• What is the dose? When the right exposure metric is unknown, what is being measured, what is being missed, and is anyone keeping records that could be reinterpreted later (items 1, 4, 6)?
• Differential focus. Where does the new technology truly depart from what is known, and where does existing knowledge still apply (item 6)?
• Label or behaviour. Is a category (“nano”; by extension any technology label) doing work that behaviour should do (items 7, 8)?
• Regrettable substitution. Is a known, contested risk being swapped for an unstudied one under a “free of” framing (item 9)?
• Timing. Is intervention happening before economic interests entrench (item 5, p.447)?
• Humility in both directions. Could the risk community’s own careers and funding be locking in assumptions (item 10)?

Digest: the most important additions from S1#

This share covers his formative period, 2005–2016, which the map knew only through later posts. Risk innovation grew inside quantitative risk science, built on it rather than set against it.

1. Quantitative risk assessment as a foundation that he built on (Andrew’s first note). His sole- and lead-authored texts keep conventional assessment at the base: - 2007: control banding is presented as a supplement for when data are thin, explicitly not a substitute for conventional assessment. - 2011: regulation should stay grounded in established approaches and not discard the old merely because it can. - 2011: the toxicology review states that “the risk assessment paradigm remains relevant”. - 2011: his call for a new science of risk is framed as a way across the paradigm’s pace problem.

The impatience the map dates to 2018 was earned over a decade of making the paradigm work for nanomaterials; risk innovation extends it from those limits.

2. Where the humility comes from (Andrew’s second note). - 2006: quantifying nano risk from existing knowledge would breed false assurance of safety. - 2011: hazard measured against the wrong metric stays poorly quantified, and quantitative toxicology cannot keep pace with sophisticated materials. - Throughout: he designed measurement around ignorance (all three dose metrics, records kept for later reinterpretation, an instrument response averaged over unknown exponents), and he labelled his own projections subjective. - 2016: his public audit found that his confident 2006 agenda under-delivered where he most wanted it (exposure measurement, sensors that indicate harm, predictive models), while some feared risks proved smaller.

Humility never meant waiting. He argued for oversight under incomplete information, for control banding, and against using calls for more data as a substitute for action.

In the other direction (my reading): in nanotechnology he still built bounded, heuristic numbers under deep uncertainty. So the absence of numbers for AI looks like a judgement that no defensible hazard model or dose metric yet exists, not a rejection of quantification. His own past practice offers something he has not proposed for AI: quantification that makes ignorance visible.

3. A stated rule for what transfers, earlier than the map allows. - 2007: he distinguishes a tool that does not carry over directly from its concept, which does. - 2011: technology-independent principles (emergent risk, plausibility, impact); decoupling risk questions from technology labels; a differential focus on where the new departs from the known; and the point that new technologies need not bring new hazards, nor new hazards come from new technologies. - Literal transfer is kept for recurring mechanisms (the asbestos fibre paradigm, occupational controls) and treated as a hypothesis to test.

This partly corrects the map’s claim that he never states the rule: he stated it for materials. On this framework (my reading), “defies analogy” marks AI as an extreme case of emergent risk. Only principles and process lessons carry over, which fits his record.

4. Earlier origins than the map gives. - Plausibility as a named, openly qualitative filter: 2011, not the 2018 book. - Behaviour over labels: 2011, not 2022. - Emergent risk as the precursor of orphan risks: 2011. - Social licence: 2011. - “Safe” as a relative term, the pacing problem, and structural incentives against risk research: 2006. - The conflict between promoting a technology and overseeing it, and industry’s unfitness to lead risk research: 2006–2011. - Lay knowledge, citizens as stakeholders, could-versus-should: 2008. - Problem formulation as a wicked problem: 2011. - A publicly signalled change of mind (on defining nanomaterials): 2011.

5. Qualifications. - His first governance remedies were top-down, expert-led and given real authority over research, and his adaptive triggers were hard law. Soft law is a later layer, which strengthens map tension 9. - PEN 3 is open policy advocacy that predates the book’s honest-broker stance. - The 2023 verdict that nanotechnology was a reasonably successful transition is kinder than the 2006–2008 record: highly relevant risk research at about 1% of the NNI budget, promoters overseeing risk, paralysis by analysis and weak engagement. Nano lessons for AI therefore include institutional failures as well as successes of process. - In 2016 he warned that careers built on assumed risk can lock in error. Hubris can form in risk communities too. - A co-authored 2014 study shows risk judgements shifting with the order information is given and with “free-of” labels. This qualifies “concern as signal” and adds the concept of regrettable substitution.

For the next stage, neutral lenses: - lessons applied, or only cited; - promoter and overseer as the same body; - what the “dose” is, and whether it is measured; - where the technology departs from known behaviour; - label versus behaviour; - regrettable substitution; - acting before economic interests entrench; - humility applied to boosters and risk professionals alike.