Late Lessons, Jensen Huang and AI
Markdown version

Update log: 05-maynard-risk-and-ai-map.md (supplementary reading, rulings, clarifications, public framing)#

Update of 26 September 2026. Inputs: the seven supplementary reports (working/maynard/supplement/S1–S7), Maynard’s provenance rulings and his two clarifications (September 2026), and the publication brief (public knowledge base; third person; neutral provenance note). Specifics were checked against the source files in Resources/maynard-papers/ (text extracted to a scratch folder). Every new quotation was checked against its source; a final script checked all 1,239 quotations of 10+ characters in the map against the corpus, the FFTF extracts and the supplementary texts. The 15 non-matches were all explained: citation keys or labels in quotation marks, the September 2026 phrase “the hubris of risk assessment”, and PDF artefacts (ligatures “fl”/”ti”, a hyphenated “mind-set”, a photo caption splitting “legitimate … social licence”), each confirmed by reading the source passage.

The pre-update file is preserved as working/maynard/review/05-map-pre-update-2026-09-26.md.

Global changes#

  1. Commitments renumbered. A new commitment C5 (humility about numbers) was inserted, so former C5–C17 became C6–C18. All C-references in the map were renumbered by script (37 references) and the §4 headings 5–17 became 6–18. “Seventeen propositions” → “Eighteen”; “C1–C17” → “C1–C18”.
  2. Interpretation label. “My reading” / “[my reading]” / “(my reading)” replaced throughout by “Interpretation” / “[interpretation]”, because in a public document about Maynard “my reading” could be read as his. “The mapmaker” replaced by “the map”.
  3. Public framing. “Andrew” → “Maynard” where it referred to him (subtitle, §1); removed “prepared for Andrew” and “Andrew himself” as a reader; “at Andrew’s request” → “at Maynard’s request”; “the project’s text mirror” → the public mirror with local copy; “standing rulings” rephrased as exclusions with their reasons; meta-references to “earlier versions of this map” removed; Appendix B paths now “relative to the folder that contains this document”.
  4. Supplementary citation keys. Introduced short keys for supplementary items (e.g. NN 2015-09, PEN 2006, Testimony 2007, 2020science 2009, CR 2026) with a new Appendix C. The two 2008 and two 2010 blog posts were disambiguated as 2008a/2008b and 2010a/2010b.
  5. Symmetric risk softened. “It is symmetric” / “Risk is symmetric” replaced by “counts both sides” / “not innovating carries risks too”, with the 2014 evidence that he does not weigh both sides equally (“or, worse”, NN 2014-03). The † label “symmetric risk†” is kept only as a parenthetical in the §5.1 row.

Title, subtitle and §1 (rewritten)#

  1. Title: “2014 to September 2026” → “2005 to September 2026”.
  2. Subtitle: removed “prepared for Andrew and for later use as an analytical lens”; added a neutral provenance note (prepared with extensive AI assistance, at Maynard’s request, reviewed by him; reflects his September 2026 clarifications).
  3. “What it is for”: rewritten for the public reader (researchers, journalists, policymakers, informed readers) and as a lens (§9); removed “Andrew himself, as a check”.
  4. “Sources”: three bodies of work. Corpus described with the public mirror. Future Rising now read in part (33 of 60 chapters, Introduction, Afterword), cited “FR p.X”. New supplementary-corpus bullet listing all groups (papers 2005–2016 incl. Hansen et al. 2008, Safe handling 2006 and 2016, AOH 2007, sophisticated materials, Don’t define nanomaterials; the 2014–16 columns; testimony 2006–08; WEF proposals 2008/2010 and 2026 retrospective; 2020science archive; Rethinking Risk; Risk Innovation Nexus and related; BMI 2019, Dune 2024, HICSS 2024; 2026 papers; Feb/April 2026 essays; co-signed biopreservation papers and USDOT report).
  5. New subsection “Maynard’s clarifications, September 2026”: both clarifications stated as attributed, dated content, with a note that the supplementary reading supports both back to 2006 and that “the hubris of risk assessment” is his September 2026 wording.
  6. “Method”: added the supplementary reading stage (S1–S7), source checking, the neutral AI-assistance statement, and the rule that earlier dates found in the supplement replace later ones.
  7. “Provenance rules” rewritten:
    • Exclusions extended to the two AI-written papers (except his signed postscript and Annex 1) and the ChatGPT-filled OpenAI Risk Innovation Planner (only his framing post counts). Specific exclusions now give reasons (largely AI-written or not his prose; start-here is an orientation page).
    • Ruling 1 applied: Maynard and Garbee (2019) / 2019-08-13 moved out of “weaker evidence (co-written)” and weighted as his own thinking, with his confirmation (September 2026) recorded and the 2026 paper’s “The lesson that has stayed with me ever since” cited.
    • Ruling 2 applied: AI and the Art of Being Human removed from the weaker-evidence list (and from everywhere else; see item 17).
    • Ruling 3 applied: Boorsma and Lobo co-written posts keep their weaker weighting; supplementary co-authored items weighted by his role.
    • Ruling 4 applied: unsigned Nexus materials treated as programme positions; AI-generated text excluded; mixed-provenance items stay marked.
    • [mixed] entries updated: 2026-07-16 now records his three-day solo rewrite (“just me — no Fable this time”) and lists the parts securely his from earlier sources (threat to value 2015; orphan risks 2018; known-but-unowned 2019 and April 2026; definitions select risks 2015; the Garbee lesson); the frontier-specific apparatus stays possibly Fable’s. 2026-01-17 now records the discrepancy with the arXiv AI-use statement and keeps [mixed] for “honest non-signals” and the four mechanisms.
    • New entry on the Feb/April 2026 andrewmaynard.net essays: his own retrospective synthesis, used as a check; provenance cautions for “moral imagination” (from a ChatGPT-written post), “eighteen orphan risks” (ChatGPT output), “Seemingly Conscious AI” (Suleyman) and the Constitutional AI section (from a Claude-written paper), marked [AI-origin] where used. The April essays’ misdescription of the co-written book is not mentioned (ruling 2).
  8. “Conventions”: added supplementary keys, FR citation, silent correction of PDF artefacts, “Interpretation” label, §8 tag “[interpretation]”.
  9. “Limits”: removed “papers, columns, preprints, April 2026 essays, testimony … were not read directly”; now lists what remains unread (most toxicology papers, talks, Risk Bites, podcast, 27 FR chapters); unevenness note says the supplement fills much of 2005–2017; the 2026 preprints are more secure than the lecture and frontier paper.

§2 The picture in brief (rewritten, sound parts kept)#

  1. Opening paragraph: “a risk scientist who had lost patience with how risk is usually thought about” → “a risk scientist who built on his discipline rather than leaving it”; added the grammar’s limits from his own quantitative work (PEN 2006 “false assumptions of safety”; Toxicol. Sci. 2011 “unlikely to keep pace” / “paradigm remains relevant”), the TechTrends 2023 “physicist mindset, understanding of risk, innovation around how we think differently about risk”, and clarification 1 (attributed, dated). FFTF p.22 “less and less patience” kept.
  2. Engine paragraph: “Since January 2016 he has defined risk, first of all, as a threat” → “Since 2015 he has argued that risk is not only the probability of harm but a threat”; origins 2013 (Michigan) and September 2015; introduced as extension (“a useful starting point”; “extends conventional thinking”); complementary from the start (NN 2015-09; Coronavirus 2020); “ten years” → “eleven years”. Feature 1 now names probability as the base layer; feature 3 “It is symmetric” → “It counts both sides”, with the 2014 unequal-weighting evidence.
  3. New paragraph “Humility about numbers, and a duty to grapple” (clarification 2, attributed and dated, with 2009, 2016, 2020, 2026 evidence and the grapple side).
  4. Plausibility paragraph: dated to 2006 (grey goo) and 2011 (named filter).
  5. People paragraph: added entrepreneurs’ required optimism (2015), promoter–overseer conflict (2006–2011), and the Garbee chapter’s value mismatch as his own prose.
  6. Governance paragraph: added formative remedies (2006–08 central capacity; 2011 trigger points), soft law as a later layer “from 2015”, NANO 2026 “I don’t have a governance solution”, the early window (NANO 2026; BH 2026), inevitability-with-steering “from at least 2015”.
  7. AI paragraph: “AI arrived late” qualified with the 2014 AI-risk post, the 2015 column and 2019 Nexus tools; stages list gains 2014 (prolonged interaction), 2020 (Future Rising: fake-o-meter, evolutionary mismatch), late-2025 Berlin keynote origin, and constitutive resonance (March 2026 preprint) replacing the lecture-only formation stage. Added FWB 2026 convergence line after “defies analogy”.
  8. Being human: added 2014 “our very notions of humanity”.
  9. Method: added the 2011 signalled reversal.
  10. New “Where it began” paragraph pointing to the formative layer in §7 (including the 2008 application of the EEA late-lessons framework).
  11. “Where it is unsettled”, item 1: added his partial reconciliation.

§3 The landscape#

  1. Architecture heading “(my reading …)” → “(interpretation …)”; intro adds his April 2026 “gap” self-account with verified 30Y quotations.
  2. Engine bullet → “value, threatened and created, on a base of risk science” (two layers). Disciplines bullet adds “humility about what methods and numbers can capture”. Object bullet adds nanomaterials as first proving ground.
  3. Connections: new “Risk science → risk innovation (built on, not replaced)” and “Measurement humility → orphan risks → restraint about AI numbers”; “Risk innovation ⊃ responsible innovation” rewritten as a relation that moved from “complements” (2015) to nested (2020); threat-to-value → cognition chain now cites the CR preprint; “Symmetric risk† → steering” renamed; complexity connection adds 2019-08-13 tight coupling and latency; chemical-grammar chain adds “Nanomaterials are not just chemicals” (2016); structural-incentives chain adds 2015 and 2019 his own prose.
  4. Diagram: Engine box now “risk innovation, built on risk science … not acting is a risk too”; Disciplines add “humility about numbers”; Object line adds nanomaterials.
  5. Territory table: new row “Risk science and the limits of numbers” (C3–C5; B1, B6; S1–S3; 2005–2016); densest-period and anchor columns updated for formative dates in Rethinking risk, Epistemics, Complexity, Learning from past technologies, People and permission, Power/justice, Emerging technologies, What AI is, Mind/language, Governance; supplement reports added as sources; lens references updated for new lenses.

§4 Core commitments#

  1. New introductory caution paragraph (clarification 1; rhetoric vs substance, “a radical new approach to risk”).
  2. C1: firmness line adds 2009 “People matter” and 2020 “everyone has the right to thrive”; “named explicitly from 2025” → “at the centre of his work from 2025”; removed the parenthetical on 2025-10-23 21-tools-for-thriving-with-ai (ruling 2); sources add 2020science 2009 and FR.
  3. C2: “It is a threat” → “It is also a threat”; added NN 2016-03 “useful starting point” and 2026 “does not abandon … widens” [mixed]; firmness re-dated (seeded 2013; in print September 2015); new quotation NN 2016-03.
  4. C3 rewritten: title “Quantitative risk assessment is the foundation, necessary but not sufficient, so risk thinking must innovate on top of it”; documentary layering 2007–2026 (AOH, 2009, Nat. Mater., Toxicol. Sci., NN 2015-09, Rethinking Risk, Coronavirus 2020, JLME 2024, 2026-07-16); noted the absolute-sounding 30Y line vs NANO’s “not just”; firmness notes the impulse visible from 2008 and name in 2015.
  5. C4: formative sources added (1990s measurement, ILSI 2005, AOH 2007 characterisation, 2009 “No exposure—no harm”, NN 2016-06).
  6. New C5 “Hold numbers with humility…” — clarification 2 attributed and dated, with the documentary record 2006–2026, the “not against numbers” evidence, and the grapple side.
  7. C6 (was C5): added “safe” as “a relative term” (2006), “Who will decide … who will pay the cost?” (2008); the 2026 standing/expert split re-anchored to Handbook 2010 (co-authored) and “single source” removed; firmness updated.
  8. C7 (was C6): title “Risk is symmetric” → “Not innovating carries risks too”; added Testimony 2006, NN 2014-03 unequal weighting, AOH 2007 middle ground on precaution, NN 2014-09 trigger points, NN 2015-12 inevitability with steering, FR 2020 “not inevitable”, Dune 2024 slowing as legitimate choice, BH 2026 window; firmness re-dated.
  9. C8 (was C7): plausibility dated 2006/2010/2011/2014; FR 2020 (plausibility filters building, not imagining); firmness “from 2006”; new quotation from 2020science 2014.
  10. C9 (was C8): added 2020science 2010b (complexity), NN 2015-12 (early warnings, failing fast), 2019-08-13 (tight coupling, latency, value mismatch) and FR ch.39.
  11. C10 (was C9): added Testimony 2006, CETI 2010 GM lesson, FR p.154 realist anchor, Regrettable substitutions 2014 in sources.
  12. C11 (was C10): structural account “from the start”: PEN 2006, Testimony 2007 “good intentions are not enough”, Hansen et al. 2008, NN 2015-03, 2019-08-13 (now his own prose, ruling 1).
  13. C12 (was C11): added Testimony 2006, Bulletin 2008, Nature 2006, Rethinking Risk 2017, FR.
  14. C13 (was C12): title adds “backed by strong capacity for risk research”; formative layers (2006–08 testimony, WEF 2008/CETI 2010, Nature 2011 triggers), 2019-08-13 “crude boundaries”, 2026-07-16 [mixed] governance lines, co-signed Wolf 2024 and Pruett 2025, NANO 2026; firmness updated.
  15. C14 (was C13): rule stated for tools and materials (AOH 2007, Hansen et al. 2008, Toxicol. Sci. 2011, Nature 2011 draft, NN 2014-06, NN 2015-06) and his 2026 reconciliation (CR, FWB); firmness: discontinuity concerns scale and speed; new STICK quotation.
  16. C15 (was C14): 2014 question, FWB “more firmly now”, FR 2020 mechanisms, Trojan 2026 scope (“designed to be genuinely useful”), CR 2026; firmness notes formation now in his own preprint.
  17. C16 (was C15): 2014 singularity rejection, 2015 cyber insecurity, 2019 Loss of Agency card.
  18. C17 (was C16): 2014 seed and CR 2026.
  19. C18 (was C17): 2008, 2015, 2016 roots; S3 2026 education as formation; one quotation corrected against source (NN 2016-09 wording).
  20. Method note: added 2009 “I may be wrong”, 2011 reversal, 2016 public audit, “explorations, not findings”, 2017 CT-scan self-implication, and the late turn to fiction (2010 pejorative → 2012 speculative design); firmness “2009–2026”.

§5 Key concepts (tables rewritten; unchanged rows kept verbatim)#

  1. §5 intro: “First” now includes the supplementary corpus; † defined as the map’s labels.
  2. §5.1 new rows: Quantitative risk assessment as foundation; Measurement humility; Humility about the hubris of risk assessment; Decisions under incomplete information; Emergent risk; Trigger points for action; Quick to question, slow to respond; Novelty as an unreliable indicator; Regrettable substitution (co-authored, one-off); What the framing makes invisible†. Updated rows: Risk innovation (2013/2015 origins; complementary), Risk as a threat to value (2015/2016), Existing and future value (2015), Reciprocal threats (CIO guide 2022 assumption), Risk landscape (2015/2016), symmetric-risk row renamed “The risks of not acting” and dated 2006/2014, Socially defined safety (2006), Hazard vs risk (2005/2007/2009), Exposure of the mind (Trojan 2026), Risk perception (2006, 2010, 2014, 2020 qualifications), Comparative risk → with benchmarking (2010, 2016; “Occasional” → “Recurring”), Orphan risks (2011 precursor, 2019 known-but-unowned, April 2026 own prose, July 2026 [mixed]), Social licence (2011), Precaution (2007, 2014).
  3. §5.2 new rows: Mutual worth creation (2019, 2026); Promoter and overseer (2006–2010); The less responsible entrant (2016). Updated: Responsible innovation (2015 and 2019 doubts; 2015 “complements” → 2020 nested; Garbee no longer “co-written”), Could vs should (2008; FWB “most durable”), Hubris (2008, 2014, 2016, 2020), Reversibility test seeds (2015, 2020), Structural incentives (2006, 2015, 2019 his own; “co-written” flag removed), Honest broker (advocacy before 2018; STICK 2026).
  4. §5.3 new rows: Actionable empathy (2015); Strong capacity for risk research (2006–08); Institutions for anticipation (WEF 2008, CETI 2010, Prehistory 2026); Stewardship and future generations (FR 2020; FWB 2026). Updated: No abdication (2008), Two-way engagement (2007–09), Loud vs quiet voices (2016 “information-vacuum”), Agile governance (2007 pacing; 2015 “years, not hours”), Operationalised ethics (Guardian 2017), Technology vs use (his own 2009 rule), Inequity (2006), Whose future.
  5. §5.4 new rows: Late lessons from early warnings (2008–2016); The gap (his April 2026 organising construct; no †, since the term is his); The early window and lock-in (2008–2026). Updated: Convergence (“core to 2021, then background” → “a live systems premise into 2026”), Complexity (2010; 2019-08-13 no longer co-written), Tipping points and early warnings (2015), Rising irreversibility (2010; NSF 2023), ATT (NSF, TechTrends 2023), ATT models (April 2026), Inevitability (2015).
  6. §5.5: Plausible vs imaginable (2006–2014; FR 2020); Exponential row “with a reversal” → “Exponentials and S-curves” (not a reversal); Informed speculation “One-off [mixed; single source]” → “Recurring (2026; roots 2014)” with his 2026 papers; Analogy as mindset (2007, 2011); new row Behaviour, not labels (2009–2022, core); Epistemic humility (2014, 2016); Fallible science (2014, 2016); Science fiction as lens (2012 speculative design).
  7. §5.6: From converging strand (2014, 2015; FWB caveat); Superintelligence agnosticism (2014); Emulation (Fable annex 2026 “incremental and combinatorial”); Relational technology (Harness, HNS); Seeming vs being conscious (Dune 2024); Moral status (2014; Harness Feb 2026) — corrects the earlier claim that it does not recur after 2024; Augmentation (Slate 2023).
  8. §5.7: Mundane but serious (NN 2014-06; STICK); Democratic and systemic risk (NN 2015-12); Drain of agency (Nexus 2019; Dune 2024); Deepfakes (FR 2020 chapter; 2024 is a sharpening, not a reversal).
  9. §5.8 new row: Evolved defences and mismatch (2016 materials parallel, 2017, 2020, 2026). Updated: Artificial manipulation (2014 seed; FWB 2026), Economic gradient (2019 Nexus scenario as interpretation), Benevolent persuasion (CIO guide 2022), Cognitive Trojan horse (OEB 2025 origin; Trojan 2026 scope), Honest non-signals (Trojan 2026; [mixed] kept), Constitutive resonance (re-anchored to CR March 2026 with its hedges; Slate 2023 positive precursor), Formation (two faces; re-anchored to CR 2026 and S3 April 2026; the lecture no longer “the only explicit statement”).
  10. §5.9: Flourishing (2009, 2020; “explicit from 2025” → “central from 2025”); Intelligence is not goodness (FR p.70).
  11. §5.10: What education is for (S3 2026 formation); Education as a lever (2008, 2015, 2016; accountability); Catalyst/illusion (CR reconciliation); AI literacy (2008 roots; TechTrends; CR “existential preparation”).
  12. §5.11: Nuance (CETI 2010); Self-implication (Rethinking Risk 2017; Fable annex); Writing as self (September 2026 “academic dishonesty”); new one-off row Humour and satire (Scholarship 2026).

§6 Threads (all eleven updated; cores kept)#

  1. Intro notes the supplement and points to §7 for the formative layer.
  2. T1: core now “built on, not in place of, the probability of harm”; new “Origins” paragraph (2006–2017); development adds complementarity evidence, 2015 “beginnings of … frameworks” echo, 2011 emergent-risk precursor, April 2026 orphan-risk label; new “Numbers and their limits” paragraph (clarification 2); perception strand adds 2006 and the 2014 qualification.
  3. T2: new “The rule, stated early” paragraph (2007, 2008 late lessons, 2011); development adds his own 2009 use/technology rule, 2014–15 counterweights, 2026 reconciliation; new “The nanotechnology record is mixed” paragraph; transfer list adds characterisation, knowing what to measure, 2010 brand foreshadowing, the “new, metaphorical grey goo” warning; biopreservation adds JLME 2024.
  4. T3: development adds 2014, 2015, 2019 (Nexus AI cases), 2023 (NSF “years to months”).
  5. T4: core adds harm from AI working as designed; development adds 2014, 2017, 2020, 2022, 2023 (Slate), OEB 2025, CR 2026; formation now dated April 2026 (positive) and September 2026 [mixed]; valence shift noted.
  6. T5: core adds strong capacity for research; development adds his roles (verified against NANO 2026), 2006–08 testimony, WEF/CETI and Prehistory 2026, 2009–2011, Guardian 2017, 2019-08-13 “crude boundaries”, 2026-07-16 [mixed] governance lines, NANO 2026; the public/expert split re-dated to 2010.
  7. T6: development adds PEN 2006, 2020science 2010b, NN 2015-12, stewardship, NSF 2023, BH 2026 window, FWB convergence.
  8. T7: development adds 2006–08, 2014–16, NN 2015-03, Nexus 2019 sincere founders, 2019-08-13, and 2016 hubris in risk research.
  9. T8: core adds humility about numbers; development adds physics/measurement formation, 2005–2011, 2011 reversal, 2010→2012 fiction turn, 2016 balance, U-shaped advocacy.
  10. T9: core adds clarification 1; position adds FWB alignment credit.
  11. T10: development adds 2008–16 roots, Slate 2023, S3 2026.
  12. T11: development adds 2009, 2014, 2020, Dune 2024, BH 2026, CR 2026.

§7 How his thinking has evolved#

  1. Phases table rebuilt: Phase 0 “Formation, to 2014 … known from later accounts” replaced by 0 Measurement science (1990s–2004), 1 Nanotechnology risk and governance (2005–2011), 2 From nanotechnology to risk innovation (2012–2017); later phases renumbered 3–8 with small additions (2019 tight coupling; Future Rising stewardship and hubris of prediction; NSF 2023; 2026 constitutive resonance, harness critique, gap and window).
  2. New subsection “The formative layer, 2005–2016”: quantitative foundations (ILSI 2005, AOH 2007, Toxicol. Sci. 2011); research strategy and testimony (PEN 2006, Testimony 2006–08, 2020science 2008a); “Safe handling” 2006 and ten-year audit 2016; late lessons from early warnings 2008 (Hansen et al.; 2020science 2008b; 2011 and 2016 slow-uptake examples); regulating sophisticated materials and “Don’t define nanomaterials” (Handbook 2010, Nat. Mater. 2011, Nature 2011); the 2014–16 columns; the WEF institution proposals (WEF 2008, CETI 2010, Prehistory 2026; notes that the documents do not use the term “early warning”); and “What the formative layer changes in this map” (layers; humility as old as the numbers; earlier dates; governance instincts; advocacy; mixed nano record).
  3. New subsection “His own retrospective, April 2026”: where the essays confirm the map, where they shift emphasis (the gap; early window; convergence; orphan risks as his label; institutional realism; education as formation; stewardship; alignment credited; tensions he names), and humility; notes that “hubris of risk assessment” is his September 2026 wording, not the essays’.
  4. “What stays constant”: new items 1 (quantitative base) and 2 (humility about numbers with duty to act); others re-dated (risks of not acting 2006; plausibility 2006; manipulation 2014; who decides 2008; binary 2010; being human 2014; justice 2006; self-implication with 2011 reversal and 2016 audit). Now 14 items.
  5. “What changes”: items 1–8 updated (convergence caveat; 2014 and 2020 stages; tails from 2010/2014 and S-curve reading; remedies incl. 2015 doubts, 2008 literacy roots, governance layers, “not optimistic”; inevitability from 2015 with 2020, 2024 and the 2026 window; developers’ structural account from 2006); new item 9 (AI changing its user: 2023 gain → 2026 concern); item 10 (own practice) adds the September 2026 authorship stance.
  6. Changes of mind, signalled: added 2011 (Don’t define nanomaterials), 2014 (fumed silica reconsidered), 2026 (capability plateau, S3); the 2023 use/technology row notes the rule was his own (2009); the 2024 deepfakes row notes the 2020 hedge.
  7. Changes of mind, inferred: added 2010→2012 fiction turn; 2006–08→2015 governance layers; 2024→2026 resilience to formation; the 2015→2016 responsible-innovation row re-dated to within 2015 with 2008–09 roots; the consciousness row adds the 2014 substrate-sensitive view; the 2018→2025 exponential row removed and recorded under “Not a change of mind” (S-curve view, “then, as now”).
  8. “How he updates”: adds the 2011 evidence-driven reversal and clarification 1.

§8 Tensions, open questions, gaps#

  1. Intro recount: [he says so] 8, 10, 13; [partly his] 1–7, 9, 12; [interpretation] 11, 14, 15, 16 (was: two he says so, four partly his).
  2. T1 retagged [partly his]: added 2007/2011/2014/2015 rule statements and his 2026 reconciliation (CR, FWB, 30Y); “He never states the rule” replaced by what remains unstated for AI.
  3. T2 retagged [partly his]: added 2010, 2014, 2019 black swans, 2016 “quick to question, slow to respond”; research vs action distinction.
  4. T3 retitled “Two definitions of risk, and how they fit” and retagged [partly his]: 2016, 2017, 2020 and 2026 statements of the relation; the remaining gap is an evidentiary bar, distinguished from his deliberate restraint (C5).
  5. T4: added NN 2015-03, BMI 2019 “favor the enterprise”, CIO guide reciprocity assumption, JLME “primary purpose”, NN 2016-03 “who is defining them”, Rethinking Risk 2017, FR p.197.
  6. T5: added NN 2015-12 baseline, Dune 2024, BH 2026 window and NANO trajectory; interpretation adds timing.
  7. T6 retagged [partly his]: HNS “matters less than we’d like to believe”, Harness, CR p.4.
  8. T7: “co-authors” wording updated for 2026 experiments.
  9. T8 retagged [he says so]: CR “different sides of the same coin”; S3 “reveals and amplifies”.
  10. T9: added Handbook 2010 split, 2007 concrete mechanisms, 2015 self-named gap, USDOT 2025 (co-written) and Hyun et al. 2024 (co-signed) mechanisms.
  11. T11: added USDOT 2025 parasocial trust as method; CIO guide 2022 benevolent-persuasion work; Harness Feb 2026 wording checked and softened to “evident if unstated sympathy”; NANO April 2026 legitimacy point [AI-origin; endorsed].
  12. T12 retagged [partly his]: advocacy 2006–08 before the honest broker; WEF “Non-advocacy”; Prehistory regret; STICK 2026 self-named strain; U-shaped trajectory.
  13. T13: added the AI-written papers and his September 2026 authorship stance; Scholarship 2026 satire.
  14. T14: substrate-sensitive 2014 view; corrected “does not recur in his 2025–26 posts” — moral status recurs in Harness (February 2026, his own) and in 2014 form.
  15. T16: added 2014 unequal weighting.
  16. Open questions: added “What should be measured, and what does measuring hide?” and “How long is the window?”; “Can we even formulate the problem?” adds 2011 and 2015.
  17. Gaps: “Missing methods” replaced by “Quantitative treatment of AI risk: a deliberate restraint, not a missing method” (clarification 2, attributed and dated, with the documentary record and what remains undeveloped); new “[he says so]” evaluation gap (JLME 2024; 2026-07-16); whistleblowers moved from “little developed” to “in programme materials and a co-written guide, not his essays”; cybersecurity and Global South qualified; cognitive evidence base adds HNS SCOPUS line.

§9 Lenses#

  1. Count 24 → 29; era categories add formative (2005–2014).
  2. New lenses: B6 “Is precision standing in for understanding?” (clarification 2); D7 “Is the body that promotes the technology also the one judging its risks?”; E6 “How long is the window?”; F3 “Are known lessons being applied, or only cited?”; F4 “What does the framing make invisible?”.
  3. Updated lenses: A1–A3, B1–B5, C1, D1, D3–D6, E1–E2, F1–F2 era tags and wording (formative dates; C1 adds AI working as intended; F1 adds novel vs ordinary risks; F2 adds hubris in risk research; E2 adds research capacity). Group landings for A, B, E, F updated.

Appendices#

  1. Appendix A: intro explains local copy vs public mirror and points to Appendix C; new “formative layer and supplementary sources” list; new entry 47b for 2019-08-13 (weighted as his own, ruling 1).
  2. Appendix B retitled “Supporting material”; paths relative to the document’s folder; adds Resources/maynard-papers/, working/maynard/supplement/, working/maynard/review/; notes that timeline.md predates the supplementary reading.
  3. New Appendix C: full references for every supplementary key, with authorship and role; titles checked against Resources/maynard-papers/README.md (HICSS byline order follows the README: Maynard second of six); co-signed items (Wolf 2024, Hyun 2024, Pruett 2025, USDOT 2025) listed as shared positions.

Removed entirely (ruling 2)#

  1. AI and the Art of Being Human (with Jeff Abbott): removed from §1 weaker-evidence list and from C1 sources (the 2025-10-23 21-tools parenthetical). Nothing from S7 §2 or the S6 items drawn from the book (BH “four qualities”, TL page, BOOKS essay, the book’s “virtuous cycle” and tools) was used. BH 2026 is cited only for his own framing prose (the window; “changes our understanding of what we are”).

Not used or not changed (with reasons)#

Internal follow-ups (not for publication)#

Check (26 September 2026)#

An independent check of the updated map against Resources/maynard-papers/ (all files re-extracted to text in a scratch folder), the corpus and working/maynard/supplement/.

(a) Claims, quotations and citations verified#

A script re-tested all 791 quotations of 12+ characters against the sources. The only non-matches were citation keys, labels, the September 2026 phrase “the hubris of risk assessment”, and PDF artefacts, each confirmed by hand. Five FR quotes are split by ligature artefacts (“/f.l”, “/t.i”) or page breaks. The Nat. Mater. “legitimate … social licence” is split by a photo caption. No quotation matched only the excluded AI and the Art of Being Human files.

About 45 new items were also checked in context, with page numbers.

Confirmed (quotation, page and characterisation): - PEN 2006: pp.8, 9, 13, 27, 32 (report page = PDF − 2); “one scientist’s personal perspective”. - Testimony 2006: pp.51–53. - Testimony 2007: pp.16, 21, 33. P.30 supports the promoter–overseer point, although the “economic incentive” quote itself is PEN’s. - Testimony 2008: pp.2, 12 ($68m → $13m), 15. - ILSI 2005: p.7, and the sub-group chair (p.30 of the PDF). - AOH 2007: pp.3, 7, 8, 10, 11. - Nature 2006: pp.267–269. - Hansen et al. 2008: pp.444, 446 (promoters and oversight), 447. - 2020science 2009: “Numbers—hard data…”, “When the data run out – innovate!”, “No exposure—no harm”, “safety-neutral”, “People matter”, “I may be wrong”. - Handbook 2010: pp.575, 582, 583, 584. - Nat. Mater. 2011: pp.554–557. - Toxicol. Sci. 2011: “unlikely to keep pace”, “paradigm remains relevant”, “a new science of risk”, “crude but effective filter”, “technology independent”, the emergent-risk definition. - All eleven NN columns: every cited page (e.g. NN 2014-03 p.160; NN 2015-03 p.199; NN 2015-06 p.483; NN 2015-09 pp.730–731; NN 2015-12 pp.1005–1006; NN 2016-03 pp.211–212; NN 2016-06 p.491; NN 2016-09 p.735). - Maynard & Aitken 2016: p.999. - Rethinking Risk 2017: pp.194, 195, 200. - CETI 2010: pp.1–3. - WEF 2008: “Non-advocacy” (9 December draft). - Prehistory 2026. - FR pp.18, 148, 153, 154, 156–158, 166–167, 215 (printed pages mapped from the excerpt PDF). - Coronavirus 2020. - JLME 2024: pp.555, 564, 567. - BMI 2019: pp.5–6. - CIO guide 2022: pp.34, 38. - Nexus 2019 cards: pp.22, 24. The Planner text “no agreed upon tools…” is in the Planner template (© 2019). - Nexus 2020 report milestone [03/13] (PDF p.13; printed p.12). - TechTrends 2023 (“the rigor and the math”). - Guardian 2017. - Dune 2024. - Wolf 2024: p.546. - Hyun 2024: p.591. - USDOT 2025. - Pruett 2025. - Trojan 2026: pp.1, 11. - CR 2026: pp.2, 4, 7, 9, 17, 19, 20. - Harness 2026: p.5. - 2026-07-16: “The lesson that has stayed with me ever since”, “(just me — no Fable this time)”, “retreat to what can be quantified”, “not as an alternative, but as an augmentation”, “has yet to be shown…”. - 30Y, NANO, HNS, FWB, S3 and STICK 2026 quotations. - CONV 2023 “this window is closing fast”. - 2019-08-13: first person, Michigan teaching, “crude boundaries”, “value of expediency”, tight coupling, latency and value mismatch.

Errors found and fixed: 1. “It complements approaches” (NN 2015-09 p.731) was cited in §2 and §5.1 as evidence that risk innovation complemented established risk assessment. In the source, it complements responsible innovation and anticipatory governance. §3’s use was already correct. In §2 and §5.1 the quote was replaced by “Important as evidence-based … are” (p.730), with Coronavirus 2020 and JLME 2024 for complementarity. 2. The “past 2028” falsification point and the observable tests come from the frontier-AI paper and concern the possibly-Fable “safety differential”. C5 and the §8 Gaps bullet presented them untagged as “his own thesis”. Both now cite 2026-07-16 [mixed]. 3. T2 cited the asbestos fibre paradigm to “AOH 2007 p.5”. The fibre-shaped-nanomaterials point is in Nature 2006 pp.267–268. The citation is now “Nature 2006 pp.267–268; AOH 2007 pp.4–5; 2020science 2009”. 4. §2 TechTrends 2023: the quotation was truncated before his fourth item (“love of engaging with people…”), and the map’s gloss “layers, each resting on the one before” was presented as if it were his. The fourth item is restored and the gloss removed. 5. The C5 firmness line said the stance was “Stated for AI in December 2014”. The 2014 post calls the risks untestable yet “foolish not to” examine, so this is changed to “Anticipated for AI”. “Worth preparing for” is corrected to “worth examining” in C5, §7 and tension 2. 6. T6 attributed “on behalf of generations that haven’t arrived yet” to Future Rising, “quoted in 2024-09-08”. It is his 2026 gloss (FWB), and 2024-09-08 does not contain it. This is now attributed correctly. 7. Tension 3: the Rethinking Risk range is changed from pp.194–197 to pp.193–197 (“a powerful way of making trade-offs” is on p.193). 8. §2 said he met the developer pattern “first among entrepreneurs (2015)” and then in institutions (2006–11). The order is now chronological. 9. Control banding is “the tool he proposed”, not “his tool” (§3, C3). 10. “His first writing on AI risk” is now “his earliest writing on AI risk in the record” (§2, C15, §7 phases, Appendix A).

(b) AI and the Art of Being Human#

The title, Abbott and the 21-tools post appear nowhere in the map. The remaining issue was BH 2026, the February 2026 page “What Does It Mean to Be Human in an Age of AI?”. The page is the landing page for the book: it has the four qualities, the 21 tools and links to the book. Its “narrow window—maybe five years, maybe less” paragraph sits inside the book’s framing (“I used it extensively in developing the book”). S6 also records that the window claim had previously appeared only in the book’s foreword. Ruling 2 excludes the book as evidence, context or source, so BH 2026 has been removed from §1, §2, C7, C17, §5.4, T6, T11, §7 (retrospective and “What changes”), tension 5 and Appendices A and C.

The early-window idea is now carried by: - NANO 2026 (“set the trajectory for decades”); - CONV 2023 (“this window is closing fast”, co-written); - Hansen et al. 2008; - NN 2015-03.

The “maybe five years” figure and “changes our understanding of what we are” are no longer used. §1 now describes the andrewmaynard.net material as the April 2026 essays only. For Maynard: if he regards the BH page as separate from the book, the two quotations can be restored.

Residual note: Resources/maynard-papers/README.md, which the map cites as an index in §1 and Appendix B, lists the book’s files. This is a working file, and the map does not name them.

(c) Garbee chapter#

§1 weights it as his own, and the wording is tightened to “fully his own thinking”. All 16 uses of 2019-08-13 are unflagged, and the one co-authorship mention is factual. Appendix A 47b is correct.

(d) Clarifications#

Both are stated accurately in §1, attributed and dated. The following over-statements, in both directions, were corrected: - C3’s title said “Quantitative risk assessment is the foundation”. It now says it “remains part of the foundation”, as in his wording. - §2 said “Most of what follows grows from … loyalty to risk science”. This is now “Much of what follows … grounding in risk science”. - T1 said “He was, and remains, a quantitative risk scientist”. This is now “Quantitative risk science remains part of his foundations”. - §7 “What stays constant” 1 now reads “as part of the base”. - §4’s caution called his phrasing “rhetoric” whose substance is “almost always” an extension. This is now “Some of his own phrasing … read in context, the substance is usually an extension”. - The §3 link from orphan risks to his AI restraint is now marked Interpretation, since his clarification does not make it. - C1’s “People matter” now carries its actual meaning (protection and quality of life), so it is not read as a thriving claim.

The map consistently says that humility is not a rejection of numbers or a licence to wait. It does not present him as anti-quantitative.

(e) Project-internal material#

No questions for Maynard, to-dos or article notes were found in the map. The following process references were removed or reworded: - “the map had … / had marked as its own inferences” (§7); - “The supplementary reading has moved several tensions” (§8); - “has been updated with the supplementary reading” and the theme and supplement paths (§6 intro); - the duplicated method sentence (§1); - “What the formative layer changes in this map” (now “What the formative layer shows”); - the Appendix B bullet pointing readers to working/maynard/review/, which holds this log and its internal follow-ups.

Kept as neutral method and provenance: the local corpus path, the §3 crosswalk columns naming theme and supplement reports, and Appendix B’s list of working files. For Maynard: if the knowledge base will not publish working/, these pointers could be cut.

(f) Proportion#

The map has grown from about 22,300 to 40,200 words, and “Core” concepts from 25 to 30. Changes made: - The two overlapping Core rows “Measurement humility” and “Humility about the hubris of risk assessment” are merged into one row (Core rows now 29). - The §5 intro now says “under thirty” instead of “about two dozen”. - The §8 Gaps bullet no longer repeats the clarification verbatim for a third time; it cross-refers to §1 and C5.

The humility theme still recurs in §1, §2, C5, T1, §7 and B6. That is proportionate to a stated clarification with a 2006–2026 record, but no further additions are advisable. Orphan risks remain “Recurring (a tool)”. Co-signed items are still used sparingly and flagged.