Breadth and proportion review of 05-maynard-risk-and-ai-map.md#

Reviewer’s scope: whether the map covers the whole 2014–2026 landscape of Andrew Maynard’s thinking in proportion to how central and enduring each idea is. I checked the map against the following: - the thread-by-year table and long arcs in working/maynard/timeline.md; - all 32 batch digests (working/maynard/notes/B01–B32-digest.md); - the concept index (14 families); - the nine theme files; - the corpus itself, where a claim needed checking.

All quotations below were checked against working/maynard/corpus/. Modem Futura posts were not used, at Andrew’s request. No comparison with any other material is made.

Verdict in brief#

The map is strong on the risk foundation and its disciplines: threat to value, hazard and exposure, symmetric risk, plausibility, complexity, permission and governance. It is also strong on the manipulation-to-formation arc. Orphan risks are handled correctly, as one tool within risk innovation. There are two proportion problems.

1. The citations lean heavily on 2026, and on one two-day-old mixed-provenance lecture in particular. This goes against the map’s own rule that “Weighting follows the record, not the calendar” (§1 Method).
2. Several of his largest territories are close to invisible: - education and learning; - being human and flourishing, which exists as a frame but has no thread; - risk perception, moral panic and techlash; - his risk work on non-AI emerging technologies, especially neurotechnology and enhancement.

A handful of smaller omissions and misdatings follow from the same compression.

Citation audit (dated post citations in the map)#

The most-cited source in the whole map is 2026-09-24 being-an-academic-in-an-age-of-ai [mixed], with 30 citations. That is about as many as all of 2020, 2021 and 2022 together (31). Third is 2026-07-16 orphan-risks-frontier-ai-maynard [mixed], with 12. Together these two mixed-provenance texts supply 8.4% of all dated citations. 2024, his most prolific year for relevant posts, is the most under-cited.

Ranked issues#

1. One recent, mixed-provenance lecture sets the endpoint of most arcs (high)#

Location. - §2 (AI paragraph; “Where it is unsettled”); - §4 commitments 6, 7, 10, 15, 17, 18 and 19 (sources and quotes); - §5.4, §5.5, §5.6, §5.8 and §5.9 rows; - §6 T3 and T5; - §7 “What stays constant” #7, “What changes” #5 and #8, and the last row of the change-of-mind table; - §8 tensions 1, 5, 9, 10 and 14.

Problem. 2026-09-24 was published two days before the map. Claude drafted it from a transcript, and it is cited 30 times. It supplies the end state of many of the map’s arcs: - inevitability (“We can’t pause it”); - tone (“scares the life out of me”); - the role of publics (“you cannot hand a problem of this magnitude over to everyday people”); - AI’s status (“categorical errors”, “irrelevant”); - universities as “an accelerator and a catalyst”; - informed speculation; - “I was wrong”.

For several of these it is the only source, so the map’s picture of “where he is now” rests on one spoken text drafted by an AI. That is the kind of single-recent-source dominance the brief warns against. More broadly, 2026 gets 4.1 citations per relevant post, against 1.2–2.0 for 2023–25.

Evidence. Citation counts above. In several places own-prose posts already carry the same point: - inevitability: “the boat has already left the harbor” (2026-05-21), “The AI genie is out of the bottle” (2025-08-31); - AGI: “ill-defined” (2026-04-11); - AI’s status: “a categorical error” (2025-03-15), “defies analogy” (2026-01-22); - language: “Language plays a large part in how we develop these relationships” (2023-04-05), and language as “base code” of self-identity (2024-01-01); - refusing the optimist–pessimist binary: “oxygen pessimist or optimist” (2024-03-31); - named reversals: “somewhat naïve” (2020-10-15), “read the tea leaves wrong” (2025-11-19), “far less sure” (2024-02-25).

Fix. - Swap 2026-09-24 out wherever own-prose evidence exists: - commitment 19 (“I was wrong”) → 2020-10-15, 2024-02-25 or 2025-11-19; - constants #7 → 2024-03-31; - commitment 15’s “Language is formative” → 2023-04-05 or 2024-01-01; - commitment 18 → 2025-08-31 and 2026-05-21; - §5.6 AGI → 2026-04-11. - Keep the lecture only where it is the sole source. There, add “(single source, [mixed])”, in particular for the narrowing role of publics (§6 T5, tension 9), universities as catalyst (tension 10) and “scariest” (§2, §7 “What changes” #8). - Aim for no single post above roughly 10–12 citations. - Rebalance by drawing on under-cited 2024 posts (see issues 9 and 12).

2. Education and learning are missing as a territory (high)#

Location. - §3 landscape, diagram and territories table, where there is no education row; - §4, where no commitment touches it; - §5.9, where four rows are squeezed in with being human and method; - §6, where there is no thread and T4 reduces education to “catalyst → surrender”; - §2, where it is absent.

Problem. education-higher-ed is his eighth-largest thread: 87 posts, 45 of them high relevance. It is 28% of 2023–24 posts, 59% of 2025 posts and 40% of 2026 posts. In the map “education” appears ten times, “students” once and “teaching” once. Concept-index Family 11 (seven sections) and theme T4 §4 have been almost entirely dropped. Education is also where most of his 2025 applied risk analysis happens: duty of care, advisor misuse and student mental health.

Evidence. Ideas the map drops: - Education as a lever against inequity. - AI as “translators” in admissions, with AI skills in “every high school” (2023-07-27); - flattening-the-learning-distribution-curve (2023-10-24); - agent-built courses as democratised knowledge (2025-03-27); - “free intelligence” and who controls it (2025-03-30). - The pedagogy he brings to AI. - “transformative learning has to be felt” (2021-01-15); - “lowest level of tech necessary” (2024-02-11); - six durable learning areas (2024-03-10); - conversation, not prompt (2025-08-17); - learning assessment, not grading (2025-08-24); - know AI first-hand (2025-08-10). - Formation of social competence as a deliberate educational task. “strategic and intentional approaches to developing the ‘social’ skills” (2024-10-20). - The value-creation account of what education is for (2025-03-30). The map cites this post, but only for the what-we-do / who-we-are model. - Power and dignity in AI use between advisors and students (2025-10-26).

Fix. - Add a territory row, “Learning, education and universities”. Its question: what is learning for, and who gets access, when intelligence is abundant? - Add a condensed thread (T10) drawn from concept-index Family 11 and T4 §4. - Add one sentence to §2. - Consider a commitment along these lines, with its firmness marked: education’s purpose is human formation and value creation, and equitable access to AI is part of it. - Move universities’ public duty (2016-01-31 → 2026-08-30) into this territory as well as governance.

3. Being human is under-built, and “flourishing” never appears (high to medium)#

Location. - §3, where being human sits “over the whole” but only as a closing arrow; - §6, where there is no thread; - §5.9, which has 13 rows for five concept-index families; - §7 “What stays constant”, which lists nine constants and leaves out being human, justice and universities.

Problem. being-human is his third-largest thread (107 posts, 59 high). It is the most frequent frame from 2024 (54% of 2024 posts, 50% of 2026), and it names both his Substack and the initiative he directs. The map treats it as the “value most exposed” but gives it no developed account. The positive aim of his risk thinking, human flourishing, appears zero times. Yet he describes his own work in those terms.

Evidence. - “flourishing at the heart of my work” (2025-01-30, his own prompt text); - “human flourishing and navigating complex advanced technology transitions” (2026-07-10, his self-map of his work).

Omitted concept-index Family 10 material: - “normal” versus “human”, and the fix frame applied to disability, including his public correction after Wolbring (2024-01-30); - the bedrock of being human is no longer “an immutable truth” (2024-12-29); - learning to be human with AI (2025-03-30); - AI as mirror (2024-01-01); - FFTF’s critique of the obsession with intelligence, “Being smart doesn’t make you good” (FFTF p.108), which underpins his superintelligence agnosticism and his later intelligence-scarcity argument.

Justice is called “the moral axis of his work” (commitment 11) but is missing from the constants.

Fix. - Add a being-human thread that sets out his positive picture: dignity, agency, joy, wonder, belonging and flourishing as what risk thinking protects and enables. - Add “flourishing” to §2 and §3 as the counterpart to threat to value. - Add being human (from 2016–18), justice and universities’ public duty to “What stays constant”.

4. Risk perception, moral panic, techlash and backlash are missing (medium to high)#

Location. §5.1 (fragments only: “social licence” and “blindsides”), §6 T1 and §9 lenses.

Problem. In mid-2026 he listed “my work on risk perception and engagement” among his core threads. The map has no concept for perception as a subject of study, and nothing on moral panics, techlash or public backlash. “Perception” appears three times, and “moral panic”, “techlash” and “backlash” not at all. This matters because his handling of public concern is distinctive. He reads concern as a signal of threatened value, not as irrationality, and he treats backlash as a risk in its own right. Both are applications of symmetric risk.

Evidence. - “the moral-panic fires on my work on risk perception and engagement” (2026-07-10); - moral panics are “rarely cut and dried” and signal what is “important to people” (2025-06-01 vibe-coding-moral-panic); - the Waymo arson narratives are “lazy and dangerous”, and backlash would make development less accountable (2024-02-18); - “numeric logic is often trumped by what we intuitively think and feel is important” (2016-03-12); - perception as the fifth element of risk (2023-11-26); - cultural cognition (Kahan, including his PEN nanotechnology studies) carried into AI persuasion (2024-09-01) and the Intelligent User Trap (2026-01-10); - acceptable safety as “highly subjective” (2024-08-07).

Fix. - Add a §5.1 row, “Risk perception, moral panic and techlash: concern as a signal of threatened value; backlash as a risk”. - Add a paragraph to T1. - Add a lens: “What does public concern, panic or backlash reveal about what people value, and what would dismissing it cost?”

5. His risk work on emerging technologies other than AI is missing, especially neurotechnology and enhancement (medium to high)#

Location. §5.8 (“Technology acting on the mind”, marked “Recurring (seed)”), §6 T2 and T4, and §7 phase 3.

Problem. The map’s past-technology reference set is historical: chemicals, nano, GMOs, rDNA and nuclear. But from 2016 to 2025 he did much of his applied risk work on contemporary technologies: brain–machine interfaces, smart drugs and enhancement, embryo screening, biopreservation, de-extinction, synthetic biology, humanoid robots and autonomous vehicles. That is where risk innovation and orphan risks were actually used. The brain–machine interface work is also the direct bridge from technology acting on the mind to AI acting on the mind.

In the map, “enhancement” appears 0 times, “Neuralink” once and “embryo” 0 times. The timeline’s tag table records only one neurotech-tagged post, so the batch tagging itself missed this strand. That explains the gap but does not justify it.

Evidence. - Neuralink and brain–machine interface posts: 2019-07-23, 2019-11-01, 2020-08-28, 2020-10-15, 2024-01-30, 2024-03-21, 2024-09-18 and 2024-11-17. - “synergistic scaling of ability, accessibility, and use” (2020-10-15). - Implants as “indentured servitude” (2020-08-28; 2020-10-15). - “who ‘owns’ you” (2024-03-21), and a “lifetime responsibility to patients” (2024-09-18). - Neural data and “mental monoculture” (2024-11-17). - FFTF ch.5 (smart drugs) and ch.7 (augmentation). - Embryo screening as “a full worked example” of his method, in the B14 digest’s words (2024-04-14). - Gene drives and “life’s operating system” (2016-01-20).

Fix. - Add a paragraph to T2 (or a new subsection) on contemporary emerging technologies as the proving ground of his risk method. - Add §5.8 or §5.3 rows for “synergistic scaling” and “technological indentured servitude / lifetime responsibility”. - Upgrade neurotechnology from “seed” to a recurring 2016–2024 strand in the T4 development line.

6. The centrality labels contradict the map’s own rubric and favour 2026 (medium)#

Location. - §5 rubric (“Core (spans many years…)”); - §5.8: Cognitive Trojan horse “Core (2026)”, Formation “Core (from 2024)”; - §5.6: Relational technology “Core (2026; roots 2023)”; - §5.4: What we do / who we are “Core (2025–26)”; - §4 intro (“Nineteen propositions he has held consistently”) set against commitment 15 (“grows from a hunch in 2023 to his central thesis in 2026”).

Problem. Concepts from a single year are labelled Core, and an emergent thesis is listed among propositions he has “held consistently”. Six of the seventeen rows in §5.8 are 2026 coinages. This gives recent vocabulary more structural weight than long-lived ideas, such as hazard vs risk, which is marked only “Recurring”.

Fix. - Add a separate tag, “Rising (current frontier)”, for concepts that are 2025–26 in origin. - Keep “Core” for multi-period ideas. - Recast commitment 15 as the latest stage of commitment 14, or label the §4 list “held consistently or developing steadily”. - Consider merging the 2026 formation rows (constitutive resonance, reverse formation, LinkedInification, validation gap) into fewer rows.

7. The structural account of developers is dated to 2026 when it runs from 2019 (medium)#

Location. - §7 “What changes” #6 (“From individual psychology … to structure (incentive fields)”); - §6 T5 (“Two shifts stand out: from persuading leaders to explaining structures”); - §8 tension 4 (“He acknowledges this only in the 2026 paper”); - §9 lens 9.

Problem. The shift is built on “sincerity almost always operates inside an incentive field” from 2026-07-16 [mixed], a paper whose frameworks may have come from the model. But structural and incentive explanations appear in his own prose well before 2026.

Evidence. - Value mismatch: market value is not societal value (2019-08-13, co-written). - Amazon’s pull toward “ever-more-dependent super-consumers” (2022-02-12). - The race dynamic after Altman’s firing (2023-11-18). - “monetizing the customer is the name of the game” (2024-03-21). - The “economic gradient” pulling deployment toward manipulation, with individuals as “engines of value creation rather than the primary recipients” (2024-07-13).

The last two also show that he acknowledged the value frame’s distributional limits before 2026.

Fix. - Redate the arc: structural explanations from 2019, sharpened in 2024 (the economic gradient), formalised in 2026 [mixed]. - Reword tension 4 to “first made explicit in the 2026 paper; foreshadowed in 2019 and 2024”. - Source lens 9 primarily to 2024-07-13.

8. The ethics-to-risk reframe is under-weighted (medium)#

Location. §5.1–5.3 (no row), commitment 2 and §6 T1 and T5 (mentioned in passing).

Problem. One claim recurs from 2019 to 2024 and ties his risk work to AI: AI governance slid back into ethics, and risk thinking is the practical corrective. The map has “operationalised ethics” and “value vs values”, but not this through-line.

Evidence. - 2019-11-01: the Neuralink assessment moves from ethical critique to the orphan-risks landscape. - 2021-08-03: AI ethics booming, “comparatively little work” on AI’s “risk-implications”. - 2023-04-04: ethics gives no “practical framework for achieving safe and beneficial technologies”. - 2023-11-21: value vs values. - 2024-12-17: value is “more effectively operationalized”.

Fix. Add a §5.2 or §5.3 row, “From ethics to risk (2019–2024)”, with these dates, and one sentence in commitment 2.

9. The AI risk landscape leaves out democracy, misinformation, social cohesion, the drain of agency and his critique of x-risk ideology (medium)#

Location. §5.7, §6 T3 and §4 commitment 16.

Problem. “Democracy” and “misinformation” appear 0 times and “cohesion” 0 times. The theme file T3 (§2.10) has a whole section on systemic and democratic risk that did not make it into the map. The drain of human agency through AI-directed work (2024) is also missing. So is his named critique of existential-risk ideology (2024-04-28), which is his clearest statement of why he rejects the x-risk frame while keeping catastrophe on the table.

Evidence. - Catastrophe as threats to social, economic and political structures, including democracy (2023-05-31). - “democracy and truth” (2023-07-25). - Misinformation as a tool of power and “unelected billionaires” (2024-01-17). - Social cohesion → “social collapse” (2024-08-25). - “where does democracy fit” (2024-09-01). - “AI-directed and human-executed implementation” (2024-11-24). - “irreversibly integrating AI into every aspect of our lives” (2024-07-21). - Bostrom’s “philosophical elegance” and x-risk culture’s “disdain for society” (2024-04-28).

Fix. - Add §5.7 rows: “Democratic and systemic risk (misinformation, power concentration, social cohesion)”, “Drain of human agency (the amanuensis)”, and “Against x-risk ideology, not x-risk” (2024-04-28). - Add 2024-04-28 to commitment 16’s sources.

10. The moral status of AI has been dropped (medium to low)#

Location. §5.6 (only “Seeming vs being conscious”) and tension 14.

Problem. Concept-index 7.4 (AI consciousness, personhood and moral status: recurring) is reduced to a hazard to humans. The other direction of his concern is missing: dehumanising or enslaving AI minds for convenience, which is a risk to AIs and a moral risk for us. “Personhood” appears once, and “moral status” and “enslav-” 0 times.

Evidence. - “economic expediency of denying consciousness” and not “enslaving AIs … under the claim that they are ‘just machines’” (2023-08-23). - Personhood beyond biology (2023-09-28). - The Ghost in the Shell intro on avoiding “dehumanizing them … to justify how we control and use them” (2023-08-18). - Personhood that “extends beyond human exclusivity” (2024-06-16). - FFTF p.58 on “Who decides” rights for non-human entities.

Fix. Add a §5.6 row, “Moral status and the ethics of control (both directions)”, and a clause in tension 14.

11. The chemical and nano bridge is sound but thin on his own carry-overs (medium to low)#

Location. §4 commitments 3 and 13, and §6 T2.

Problem. The past-technology comparisons the brief emphasises are present: algorithmic exposure, behaviour vs labels, graphene, nano process lessons, “brand nano”. Several of his explicit own-prose transfers are missing, so the bridge looks as if it thins out after 2023 more than it does.

Evidence. - Nanomaterial risks “slip under the radar of critical public evaluation” once attention moves on, and boundary organisations closed (2016-02-01). - Non-linear dose–response (thresholds, hormesis, low-dose effects) mapped onto AI exposure (2023-11-26 addendum). - The move from “gray goo” fantasy to real engineered-nanomaterial risks as his template for handling catastrophic speculation (2024-06-23). - Nano-era cultural-cognition studies reused for AI persuasion (2024-09-01). - Synthetic chemicals and vaccines as evolutionary-mismatch analogues (2026-01-10). - The “sophisticated materials” framework (2022-02-10).

Fix. Add a short “What he carries from chemicals and nanomaterials, and when” list to T2, with these dates, so the lens later has concrete transfer points.

12. The “no quantitative treatment” gap is overstated, and comparative risk has been dropped (low to medium)#

Location. §8 “Gaps and thin areas”, missing methods (“no quantitative treatment of AI risks despite his risk-science background”).

Problem. Concept-index 1.13 (comparative risk against a baseline) was dropped, and the gap statement is too strong.

Evidence. - 2023-11-09 (waymo-safety-study-shows-benefits) checks Swiss Re claims data against NHTSA/FHWA crash rates with his own “back-of-the-envelope calculations”, and concludes “not all technologies — or companies — are created equal”. - Scale arithmetic against a “fewer than 1%” figure (2025-11-09). - A bibliometric comparison of AI ethics and AI risk papers (2021-08-03).

Fix. Reword the gap to “little quantitative treatment beyond occasional comparative-risk checks (2023-11-09; 2025-11-09)”, and add a §5.1 row, “Comparative risk against a baseline (occasional)”.

13. 2019–22 and Future Rising ideas are under-used (low to medium)#

Location. §5.1–5.4 and §7 phase 3.

Problem. The citation share for this period is roughly right, but several of its distinctive ideas vanish: - latency, tight coupling and value mismatch (2019-08-13, co-written: weaker, but cited by him in 2026); - trust earned through trustworthiness (2020-12-15); - fallible, slowly self-correcting science (2020-09-26; concept-index 5.6); - scientists’ responsibility for hype (2014-12-14); - the Future Rising ideas. These are available in his own words as 60 book quotations in 2024-09-08, which the map cites once: - “stealing the futures of others”; - learning to “spot early warnings and stay clear of critical tipping points”.

Fix. - Add “latency” to §5.4. - Add “stealing the futures of others” to the “Whose future?” row. - Add “early warnings” to the tipping-points row. - Add one row on fallible and non-neutral science. - Use 2024-09-08 as a source for Future Rising rather than treating the book as largely unavailable.

14. AI and the future of knowledge is treated as provenance, not as a concern (low to medium)#

Location. §5.9 (artisanal intellectual, validation gap, intelligence scarcity) and §8 tension 13.

Problem. ai-in-scholarship-writing is 61 posts, about half of all 2025–26 posts. The map treats it mostly as his practice and as an attribution problem. His substantive claims about risks to expertise and knowledge are scattered: - human-only scholarship looking “quaint” (2025-02-04); - the need to rethink the PhD (2025-02-09); - deskilling and a two-tier “artisanal” class (2025-02-16, partly co-written); - the validation gap (2026-06-12); - plural “ways of knowing” (2025-07-27); - AI-first publishing (2026-03-22).

Fix. Gather these under the new education and universities thread (issue 2), or add a short subsection, “Knowledge, expertise and validation”, in T4.

15. The lenses omit care, capacity-building and perception (low to medium)#

Location. §9.

Problem. Of the 19 lenses, none covers: - care, even though “hard care” and duty of care are rising, 2025–26; - building people’s capacity to live with a technology intentionally (2024-10-20; 2025-08-31 “capacity to thrive”; playgrounds, not playpens); - public concern as a signal (issue 4).

Lens 9 rests on a one-off, mixed-provenance concept (issue 7).

Fix. - Add three lenses: “Who is cared for, and is the care substantive or performative?”; “Are people being equipped, through experience, play and relationships rather than warnings, to live with it?”; and the perception lens from issue 4. - Re-source lens 9 to the economic gradient (2024-07-13).

16. Housekeeping: a Modem Futura post is counted in the timeline (low)#

Location. timeline.md, thread tags by year (caveat bullet 3), and the volume figures repeated in map §1 Limits.

Problem. The tag table counts 2024-11-06 ai-in-a-world-of-trump, a Modem Futura promo post. The map (§1) and Andrew’s instruction exclude these posts.

Fix. Drop it from the counts (2024 becomes 64 posts) and update the figure in §1 Limits.

What is proportionate and should stay#

• Orphan risks. They sit correctly as a recurring tool within risk innovation (§5.1; T1 notes 16 posts in eleven years), not as an organising idea.
• The core risk commitments (1–5, 8–13). They are well evidenced across 2015–2025.
• The manipulation → language → formation arc. Its weight is justified by continuity from 2018 to 2026 and by the tag counts (manipulation 56, cognition 77, language and formation 39). Only its 2026 granularity needs trimming (issue 6).
• Existential risk. Its calibration is proportionate to a 30-post thread.
• Governance. 2023 and T5 are proportionate.