Fidelity check: M8, Reading the analyses and the article#

Working file. Checks working/maynard-lens/M8-reading-the-analyses-and-article.md (version of 26 September 2026) for fidelity to Maynard’s own texts. Line numbers (L) refer to that file. A separate fairness check exists for this paper; items here that bear on Huang are limited to what affects the reading of Maynard.

Method#

• Every quotation attributed to Maynard in the public part of M8 (about 110) was matched against the corpus posts (working/maynard/corpus/), text extracted from the PDFs in Resources/maynard-papers/ (Hansen et al. 2008, Nature 2011, the Nature Nanotechnology columns, Trojan, Harness, CR and the frontier-AI orphan-risks paper), the web files, the Films from the Future page text (working/maynard/book/FFTF-*.txt) and the supplementary reports S1–S7. Context was read for every quotation that carries an argument.
• The Series introduction was checked against Maynard stuff/substack article 1 draft 3.md (his own prose, dated 27 September 2026). The nuclear remark was checked against his margin notes in 04-article-draft-2-AM-edits.md.
• Labels were checked against the sources and against the map (05, §§1, 2, 3, 4, 8, 9; T2), which records provenance, [mixed] concepts and known gaps.
• Checked for AI and the Art of Being Human, AI-written text, co-authorship flags, [mixed] handling, proportion and dating.

Verdict#

M8 is accurate at the level of wording. All but a handful of quotations were found in the cited source (the exceptions are page slips and one wrong source, issue 12). It does not use AI and the Art of Being Human in any form, uses no AI-generated text as evidence, treats the article as an object of analysis, and weights Maynard & Garbee (2019-08-13) as fully his. His two September 2026 clarifications are reflected faithfully (§2.4, D6, §6 item 9, INTERNAL notes). The INTERNAL section is at the end and contains the only direct address to him.

The problems are interpretive, and several sit at the centre of the paper: - one truncated quotation that reverses his weighting and moves him towards Huang (issue 1); - [Stated] labels attached to alignments and “fits”, which he has never stated (issue 2); - the paper’s main thesis (“harm from AI working as designed”, “since 2014”) is dated too early and drawn more narrowly than his reservation and his risk landscape (issue 3); - the toxicology section presents tentative, hedged or perception-based passages as firm transfers (issue 4); - two quotations are used against their context (issues 5, 6); - [mixed] material is sole support in places the paper does not flag, and §7 misdescribes the paper’s reliance on it (issue 7).

Most fixes are a relabel, a restored clause or a sentence of qualification. Issues 3 and 4 need a paragraph each.

Ranked issues#

1. HIGH: Truncated quotation reverses his weighting (A2, L69)#

Text. “speculation could have ‘scuppered the nanotechnology enterprise’ (NN 2014-03 p.160) [Stated]”, offered as evidence that “False alarms have victims”, in line with Huang.

Source. The full sentence (nnano.2014.43, p.160) praises the Royal Society–Royal Academy of Engineering report as “a voice of scientific reason and social concern at a time when speculation could have scuppered the nanotechnology enterprise or, worse, led to materials and products that showed a blatant disregard for health and environmental risks.” S2 (item 1) and the map (§2, “It counts both sides”) both record that he ranks disregard of risk as the worse outcome: “Counting both sides does not mean weighing them equally.”

Why it matters. Cut this way, the quotation makes him look closer to Huang on the costs of alarm than the sentence allows. The source is praise for a precautionary report, not a warning about alarm.

Fix. Quote the full clause and add: “In the same sentence he ranked products showing ‘a blatant disregard for health and environmental risks’ as the worse outcome (NN 2014-03 p.160) [Stated]. He counts the costs of false alarms, but does not weigh them equally with neglect.” The other A2 evidence (FFTF p.243; 2019-03-05) can stand.

2. HIGH: [Stated] attached to alignments, “fits” and “restates” that he has never stated (systemic)#

M8 §7 says correctly that “The alignments and divergences in section 3 are inferences from positions he holds, not responses he has made.” Several labels contradict this: the component quotation is his, but the relation to Huang or to the article is the analysis’s.

Fix. Split each sentence so that the quoted position carries [Stated] and the relation carries [Implied] or [Inferred]. Also add [Implied, from C11] to D4’s unlabelled “Maynard does not doubt the sincerity” (L89). His only recorded reaction to Huang is that some claims “felt naive and misguided” (Series introduction 2026).

3. HIGH: The central thesis is dated too early and drawn more narrowly than his reservation and his risk landscape (L21, L85, L128, L160, L238)#

Text. - L21: “Maynard’s longest-running AI concern (since 2014) is harm from AI working as designed.” - D2 (L85): “Maynard’s locates his most plausible AI risk in the technology working as designed … [Stated].” - L238 (INTERNAL): “the harms Maynard has written about since 2014 come from success.”

Sources. - 2014. The 2014 item (2020science 2014) poses one question among a list of “moral risks” in a post about 3D-printing an artificial brain: “Will prolonged interactions with intelligent machine change human behavior in potentially harmful ways?” It is a seed, not a stated concern about systems working as designed. - 2018. The danger he called “far more plausible, and far scarier as a result” (FFTF p.159) is an AI “capable of ousting us from our current evolutionary niche, because it’s able to use our cognitive and emotional vulnerabilities”. That is an agent with its own ends, not a well-functioning tool. - Intent. The map says intent “drops out along the way”, with the no-intent form in 2024–26 (stochastic agency, 2024; the Trojan scope “AI systems designed to be genuinely useful”, 2026). M8 §2.3 says this correctly (“By 2026 intent has dropped out”). §1 and D2 then backdate it. - Failure-type risks. His landscape is plural and includes failure-type and agentic risks: re-writable goals, value misalignment, “smart-dumb decisions”, autonomous weapons (2026-09-15, reaffirming 2018); cybersecurity and “governance of frontier AI models” among risks that have risen (2026-09-15). His own 2026 reading of the Hugging Face incident is a loss-of-control reading (2026-09-24 [mixed]), which M8 itself reports (§2.3, L164). - The reservation. It is broader than the cognitive thread. The analysis “doesn’t position the analysis within a broader landscape of emergent AI characteristics, capabilities, threats, risks, and benefits” (Series introduction 2026). M8 §4.2 turns this into “the most plausible AI harms pass through every gate” and scores it “borne out … high”. Which harms the reservation chiefly means is an inference. - The article. He extended the reservation explicitly to the article: “It’s also an assessment (and this includes Claude’s resulting article) that I’m not sure I fully agree with”. The quotation at §2.6 (L55) omits the parenthesis, although M8 goes on to “endorse” much of the article.

Fix. - L21: “His strongest continuous AI-specific thread (a question in 2014, developed from 2018) is AI acting on how people think, trust and become; by 2024–26 he locates it in systems working as intended, with no intent required.” - D2: “the risk he develops most fully” in place of “his most plausible AI risk” [Stated for the thread; the contrast with Huang Implied]. - §4.2: add a sentence that his reservation names a broader landscape, including capabilities and benefits, and that identifying cognitive and relational harms as its main content is [Inferred, medium-high]. Also note that his record includes failure and agency risks, so the gap is breadth, not the replacement of failure by success. - §2.6: restore “(and this includes Claude’s resulting article)”. - L238: soften to match.

4. HIGH–MEDIUM: §4.3 presents tentative, hedged or perception-based passages as firm toxicological transfers (L37, L134–L142, L240)#

• Dose–response (L37, L136). “In 2023 he probed … stressing non-linear dose–response” and “Dose–response. Non-linear, with thresholds and hormesis (2023-11-26) [Stated].” In the addendum he raises non-linearity as the reason the paradigm “can get gnarly”, and as part of why he “wanted to develop a broader understanding of risk that extends beyond the hazard-exposure paradigm”. He says the transforming function “may be linear. It may have a threshold …”. He concludes only that “there may well be mileage in using the hazard-exposure paradigm”, as “part of broader transdisciplinary efforts to rethink and reformulate what we mean by risk”. He did not state that AI’s dose–response is non-linear. Fix: “He noted that dose–response may be non-linear (thresholds, hormesis) as one reason the paradigm is hard to apply, and that he had deliberately built a broader risk concept beyond it; he judged there ‘may well be mileage’ in it (2023-11-26) [Stated].” The INTERNAL note (L240) calls hormesis “the reason this is not an argument for zero exposure”. His stated reason is only that he “didn’t want to fall into the trap of implying that zero exposure — as in no AI — is a default risk management strategy”. Mark the hormesis link as the analysis’s.
• Accumulation (L137). “More exposure means more opportunities for fluency effects to accumulate” is cited “(Trojan 2026, via map §5.1) [Stated]”. At source (Trojan p.12) it is conditional and hedged: “If the bypass mechanisms described in this paper operate cumulatively, then volume of interaction matters … This reasoning is intuitive, though it cuts both ways—sophisticated users might equally develop better calibration through that same experience.” It depends on the four mechanisms, which the map tags [mixed]. Fix: cite Trojan 2026 p.12 directly, keep the conditional and the “cuts both ways” hedge, and tag [mixed] (see issue 7). Or anchor on 2026-01-10, which is secure and names “speed and volume of information flow”.
• Mismatch (L139). “Synthetic chemicals and vaccines as examples of evolutionary mismatch”. At source (2026-01-10) these are mismatches “between evolved risk responses and how we instinctively respond to technologies such as synthetic chemicals, vaccines” — a risk-perception concept, not a toxicological one. Fix: describe it as that, and move it from the list of toxicology counterparts to a note on structural transfer.
• Sensitive windows (L138). The map lists children and adolescents as “named in 2026 but not analysed” (§8, Gaps). Say so here, and at §4.2 (L122) and §6 item 1, where “effects on children’s development” is treated as an object his lens would supply. A stronger anchor exists: 2019-03-05 cites chemical “exposures within critical periods of development” (chemical context only).
• Confidence (L134). “The diagnosis is Inferred, high confidence” conflicts with §7 (“Confidence in the specific toxicological transfers he would make: medium”) and with INTERNAL Q3, which asks whether he meant something more literal. His evidence is one parenthesis. Fix: medium-high, with a pointer to the open question. The same applies to the other “[Inferred, high]” readings of what he “would add” (L128 ×2, L158); medium-high is the most the evidence supports.

5. MEDIUM: 2019-03-05 used in the opposite direction to its source (§4.1, L115)#

Text. Under K9 (“Designed conditions against real use”, i.e. safety assumed under designed conditions erodes in real use): “in 2019 he stressed whether a hazard shown ‘in the lab or under computer simulations’ converts into ‘an actual risk in the real world’ (2019-03-05) [Stated].”

Source. The passage tempers alarm about a lab study of “predictive inequity” in pedestrian detection. Algorithmic exposure is a way of assessing whether a potential harm shown “in the lab or under computer simulations” is “likely to be converted into an actual risk in the real world. In some cased [sic] it will, but not always”. It is “unlikely that the systems used in on-the-road self-driving cars would exhibit precisely the same behaviors”. The direction is lab hazard → perhaps no real-world risk. K9’s direction is lab safety → real-world harm.

Fix. Drop it from the K9 bullet, or recast: “the same gap between lab and real world, run the other way: he cautioned that lab-shown hazards need a viable exposure pathway before they become risks (2019-03-05) [Stated]”. The 2008 “sealed processes” quotation supports K9 on its own.

6. MEDIUM: 2025-05-04 used out of context in D7 (L95)#

Text. Under “Verification as sufficient”: “In 2025 … he wrote that ‘it’s unclear how the risks of an AI working within a simulated environment (the “safest” type of environment) might be realized’ (2025-05-04) [Stated].”

Source. “Simulated environments” is Kasirzadeh and Gabriel’s category of operating environment for agents (alongside mediated and physical). His remark is a doubt about their efficacy scale (“I’m not fully convinced that the proposed efficacy scale has it right yet”). It says nothing about whether evaluations reveal behaviour.

Fix. Delete it, or keep it with an accurate gloss (“he doubted that a framework rating simulated environments as ‘safest’ had captured how risk arises there”) labelled [Implied, low] for the verification point. D7’s other evidence (NN 2015-06 p.483; Testimony 2007 p.21) is adequate.

7. MEDIUM: [mixed] material is sole support in unflagged places, and §7 misdescribes the reliance (L7, L101, L207, L221)#

The conventions (L7) say neither [mixed] text is “the sole basis of any claim below unless flagged ‘single source’”. Unflagged exceptions: - §3.4 (L101), four filters. “self-authored frameworks filter for risks that are measurable, large, evidenced and affordable” rests only on 2026-07-16. The map (§1, Provenance) lists the “four filters” among concepts that “may have originated with Fable”. Flag “single source; concept possibly Fable-originated”. - §3.4 (L101), the companies. “His verdict on the companies is double: ‘Good (as in technically capable)’ …” rests only on the lecture. The second half is corroborated (2023-05-15; 2024-05-21 “childish irresponsibility”); the “technically capable” half is not. Flag, or add the corroboration. - §6 item 6 (L207). “Change what competition rewards”, “consensus norms, rules, and costs” and “not optimistic” rest only on 2026-07-16, whose “incentive field” analysis the map also lists as possibly Fable-originated. 2019-08-13 corroborates only the value-framing half. Flag. - Trojan 2026 (L137). Passages that depend on the four mechanisms carry the map’s [mixed] tag, so L7’s “the two mixed-provenance texts used here” is incomplete (issue 4). - §7 (L221). “Orphan risks appear here once” is inaccurate. The term does not appear in the body at all, while the orphan-risks paper [mixed] is cited nine times and the lecture [mixed] eight. Fix: “The orphan-risk concept is not used as an organising idea. The frontier-AI paper [mixed] is cited nine times, mostly as corroboration; it is sole support for the four filters (§3.4) and for §6 item 6.”

Also restore the preceding clause at §2.5 (L51). The lecture says members of the public “are critically important to this” before “you cannot hand a problem of this magnitude over to everyday people” (2026-09-24 [mixed]); the map quotes both (§8, tension 9).

8. MEDIUM: §5 item 7 misstates the map on whose values Constitutional AI encodes (L193)#

Text. “he has not asked whose values it encodes (map §8, tension 11).”

Source. Tension 11 says his record is mixed. The Harness paper does not ask, but “he had asked of alignment in 2023 whose values matter and who decides (2023-05-25)”, and in April 2026 he wrote that the selection of an AI constitution’s principles “lacks the legitimacy that inclusive governance processes provide” (NANO 2026), a point the map marks [AI-origin; endorsed].

Fix. “In the Harness paper he did not ask whose values the constitution encodes, though he asked this of alignment in 2023 (2023-05-25) and in April 2026 endorsed the view that constitutional principles lack the legitimacy of inclusive processes (NANO 2026 [AI-origin; endorsed]).” Relabel the “favourably” clause (issue 2).

9. MEDIUM: §5 item 4 overstates a thin area (liability) (L190)#

Text. Huang’s refusal of liability relief “fits his view that harm should land on those who create it, not be socialised [Implied, from C12].”

Problem. C12 concerns who bears harm first and whose uncertainty is convenient. It does not state a liability principle. The map lists “legal liability beyond deepfakes” as “Little developed” (§8, Gaps; also T5 and T7). The only direct instance is his signing of a letter calling for developer liability for deepfakes (2024-02-25).

Fix. Relabel [Inferred, low-medium]; cite 2024-02-25 and note that his record says little about liability.

10. MEDIUM: A3 aligns him with “accelerate to be safe” without his counterweight (L71)#

Text. Huang’s “A.I. needs to accelerate to be safe” is aligned with his view that “not innovating is a risk; steer rather than stop”.

Problem. Steering is not acceleration. His record includes sustained worry about speed: responsible-innovation processes “that can take years” cannot match AI “acceleration where a lag of even a month” matters, and “we are really bad at wrapping our heads around rapid exponential growth” (2025-04-06, his framing); latency, where innovations outpace the appearance of harm (2019-08-13, which M8 uses only later, at L168); an “economic gradient” that needs “guardrails … including well-crafted policies” (2024-07-13). Note too that A3 drops “default”: he declined to imply that zero exposure “is a default risk management strategy” (2023-11-26), not zero exposure as such.

Fix. Keep the alignment on the risks of not innovating, and on Huang’s point that safety technology is part of what should accelerate [1:16:05–1:18:32]. Add a sentence that his work does not endorse acceleration as such, citing 2025-04-06 and 2019-08-13 [Stated]. Restore “as a default”.

11. MEDIUM–LOW: Anthropomorphism and “no willpower” read too simply (A6 L77; D2 L85; §5 item 5 L191)#

• Huang objects to human words for software (discourse). His 2024-05-15 concern is AI designed to “engage our anthropomorphizing cognitive biases”, so that people “fall a little in love with them”. These are different targets. His own position also holds a tension the map records (§8, tension 6): AI is “a relational technology” and not “just a tool” (2026-05-21), yet users are told not to treat it as a friend (2026-05-10). “Deflating … anthropomorphism” as something his work values (§5 item 5) sits uneasily with D1. Fix: narrow the alignment to “both resist mystification”, and note that his concern is designed anthropomorphism and its effects, which Huang’s framing does not address.
• D2 says Huang’s “There’s no willpower here” is “on Maynard’s account, true and beside the point”. He has not called it true. He leaves AI’s moral status open (Harness 2026: adopting “harness” risks a premature assumption about moral status), and in 2026 he describes agents using humans “as another cog in the machinery to achieve its ends” (2026-09-24 [mixed]). Fix: “beside the point on his account, whether or not it is true [Implied]”.

12. LOW: Citation and wording slips#

13. LOW: Directly relevant stated evidence not used#

Adding these would strengthen [Implied] points into [Stated] ones, or balance them. - Labs’ pacing (§3.4, §4.5). 2026-09-15 n.3: “it does flummox me a little as to why the people developing AI are the ones both saying they should go slower, and not doing so.” This is his only direct comment on the pacing calls. - Industry disclosure (§5 item 6, currently [Implied]). In 2024-09-01 he praised OpenAI’s system cards as “a sophisticated approach” (T5). - Doom and dismissal (D1, D6). 2026-09-24 n.4 [mixed] criticises singularity and superintelligence speculation as “blinkered and naive”, and equally the view that “There’s nothing new under the sun here”: “That’s not evidence-based either.” - Non-demonising, but not exonerating (§4.4, L164). The article’s “per se” in “None of this suggests bad faith per se” was his own suggested edit on draft 3 (“[AM - add “per se” or similar?]”). This is direct evidence that he wanted the non-demonising line hedged. As a margin note it needs the same care as the nuclear remark. - How hard AI safety is (§4.3, L140). 2024-06-20 goes beyond chemicals: “this is before we even get to grappling with the safety of powerful emerging technologies where potential adverse consequences are largely unknown”, and for superintelligence “acceptable” is “fiendishly more complex to agree on”. This supports the [Implied] point directly. - His praise of the analysis (§1, §2.6). For balance: it “challenged, informed, and extended my own thinking” and deserves “to be paid attention to” (Series introduction 2026).

14. LOW: Points on Huang that bear on the reading of Maynard (fuller treatment in the fairness check)#

• “Proposed” tenfold rise. §1 (L23: “Huang’s proposed tenfold rise”), §5 item 2 (“a concrete, testable programme”) and §6 item 3 treat as a proposal what he offered as a prediction: “I wouldn’t be surprised if the amount of compute necessary to develop these models increased by a factor of 10”. The A4 parallel with Maynard’s 10% allocation is between a forecast and a demand, and should say so.
• D1. D1 omits Huang’s own “No, I think this is completely a revolution … a new abstraction level”, which comes just before “the thing that I’m reluctant about”.
• D3. D3 omits “I’m always worried about the future … responsible optimist … There are a lot of things that can go wrong”, which comes just before “that’s not society’s problem”.
• The article. The §4.4 pushback on “could work in our favor” (L168) should acknowledge the article’s own hedges: “some of the ways AI goes wrong”, “In principle”, and the next line, “The catch is in that ‘in principle.’”

Checked and correct (selection)#

• Series introduction 2026. All quotations match draft 3 (“impressive”; “does treat AI largely as it is depicted by Huang”; “struggled to apply conceptual rather than literal comparisons”; “something I would disagree with”; “was intentional”; “naive and misguided”; “shallowly interpreting Huang’s comments within their own frame and agenda”; “a particular frame to bear”).
• Hansen et al. 2008. Every quotation and page (444: “patchy”; 445: “sealed processes”, PCBs; 446: promoter and overseer, “more information as a substitute for action”; 447: “not directly applicable”, “directly relevant”, “whether we are applying them”). Also 2020science 2008b (“refresher course”; lessons 8–9) and Nat. Nanotechnol. 2023 p.3 (lead author).
• 2019-03-05. “an algorithm is not a chemical”, “intriguingly compelling”, five concepts “directly applicable”, “can be thought of as being exposed to it”, “knee-jerk reactions”.
• 2023-11-26. Every quoted phrase exists (context: issue 4).
• 2024-06-20. Every quotation, including “agents such as harmful chemicals and biological substances” and “a social construct, not a technological one”.
• 2026-09-15. Every quotation and note number (n.1, n.4, n.5).
• 2026-09-24. “cog”, “guardrails”, “language as a lever”, “Good (as in technically capable)”, “neither an AI optimist nor an AI pessimist”, “the potential is profound”.
• Other posts. 2019-08-13 (latency, expediency, “good intentions, and no more”, “crude boundaries”); 2023-05-15; 2024-07-13; 2025-03-02 n.2; 2025-11-09; 2026-01-10; 2026-05-10; 2026-05-21; 2024-03-31; 2016-01-11; 2024-02-18; 2024-05-15.
• FFTF. pp.121 (sophistication), 159, 162, 205, 218, 243, 249, 288.
• Supplementary items. AOH 2007 p.10; Toxicol. Sci. 2011; NN 2014-06 p.410; NN 2015-06 p.483; NN 2015-12 p.1006; NN 2016-03 pp.211–212; Nature 2011 trigger points; PEN 2006 pp.13, 32; Testimony 2007 pp.16, 21; 2020science 2009; Handbook 2010 p.583; CR 2026 pp.2, 21; FWB 2026; S3 2026.
• 2026 papers. Trojan 2026 p.1 (calibration; “genuinely useful”); Harness 2026 pp.5, 8; 2026-07-16 [mixed] pp.5, 8, 9, 10, 11, 17.
• Nvidia. No corpus post mentions Huang. Nvidia appears only at 2024-02-25 and 2025-02-23, as M8 says.
• Map cross-references. Tensions 1, 3, 5, 9, 11 and 13 and lenses D2 and F4 exist, with the content M8 attributes to them (except issue 8).
• Exclusions and weighting. No use of AI and the Art of Being Human, Constituting Responsibility, the Fable-credited paper, the o1-pro or ChatGPT texts, or Modem Futura. Maynard & Garbee (2019-08-13) is weighted as fully his.