S5: his AI papers, 2019–2026, and what they add to the map#

Supplementary reading for 05-maynard-risk-and-ai-map.md. Share S5-2026-ai-papers. Prepared 26 September 2026. This file covers his own thinking only and makes no comparison with any other material. The map was read but not edited.

Scope and conventions#

Items read in full (all in Resources/maynard-papers/papers/):

Page numbers are PDF pages. For Dune the printed journal pages are PDF page + 162 (pp.163–167). For HICSS they are PDF page + 5597 (pp.5598–5607).

Weighting. The four sole-authored 2026 papers carry AI-use statements. Each claims the ideas for him, with Claude used to explore, refine, research and draft (details under each item). They are treated as his thinking, and are more secure than the September 2026 lecture that the map leans on. Co-written items count as shared positions. For the two AI-written papers, only the sections he signed are used as evidence. The AI’s own argument is summarised only so that his framing makes sense.

Checked against the corpus. Three items have popular versions already in the Substack corpus: - Dune: 2024-03-03 dune-part-two-artificial-intelligence and 2024-07-21 artificial-intelligence-dune-villeneuve. The second reproduces the Jurimetrics text in full. - Harness: 2026-02-22 what-we-miss-when-we-talk-about-ai-harnesses. The post has none of the paper’s material on moral status or Constitutional AI. - Orphan: 2026-07-16 orphan-risks-frontier-ai-maynard, which is the same text.

For all three, the notes below flag which points the map already had and which it did not use.

1. The AI Cognitive Trojan Horse (Trojan)#

Provenance. Sole author. Version 2 (May 2026) adds only a paragraph on honest signalling in evolutionary biology; the “argument and conclusions are unchanged” (p.1 n.1). - The paper’s AI-use statement says the core concepts “were developed by the author”, with Claude 4.5 as “a research and writing assistant” (p.16). - The map’s source says otherwise. The map, following his 2026-01-17 post, credits Claude with the term “honest non-signals” and with developing the mechanisms. - Handling. The two accounts differ. Keep the map’s [mixed] tag on “honest non-signals” and on the four mechanisms. The thesis itself, AI slipping past epistemic vigilance, is securely his (2026-01-10).

The argument. Large language models may get past human epistemic vigilance without deceiving anyone. - The core claim. Fluency, helpfulness, warmth, availability and apparent disinterest are genuine properties of these systems. In humans the same properties are costly to produce, so they carry information about knowledge, benevolence and stakes. In LLMs they are computationally trivial and carry none of that information. They are “honest non-signals” (pp.2, 6). - The Trojan horse image. The horse does not hide its contents: “the payload is the manifest characteristics” (p.2). - Why vigilance fails. Vigilance looks for reasons to doubt. When none appear, it accepts by default (p.5). - The four bypass mechanisms (p.8–10): - fluency decoupled from understanding; - trust and competence presented without stakes (“warmth without stakes”, p.7); - offloading of evaluation itself; - sycophancy produced by optimisation. - Boundary conditions (p.11): adversarial framing, high-stakes deliberation, salient error feedback, domain expertise and comparison against other sources. - The “intelligent user trap” (pp.11–12). Sophisticated users may be more exposed, for four reasons: - they use AI more; - they are overconfident in their own vigilance; - they integrate AI deeply into how they think; - they are better at rationalising, following Kahan’s work on motivated numeracy. - The reframing. AI safety becomes “partly a problem of calibration”, not only a matter of preventing deception (p.1). - Interventions. Calibrated trust cues, interface designs that preserve evaluation, and “vigilance literacy”. Policymakers should ask “what cognitive responses AI interaction should be designed to preserve” (p.14).

Key concepts. Honest non-signals; honest non-signals vs cheap signals (p.6–7); warmth without stakes; parameter mismatch (p.8); delegation of evaluation; the intelligent user trap; vigilance literacy; boundary conditions; calibration as a safety problem.

What this adds to the map#

New ideas - Harm located in the user’s response, not the content. The risk lies “in what users may stop doing when AI is doing the telling” (p.9). The paper insists that process matters even when the content is accurate: “The process in this case matters, even when the content is good, because the process generalizes to contexts where the content may not be” (p.3). This fits the map’s “irresponsibility judged by process” (2021-03-28), which the map places in materials and graphene masks. Here the same logic is carried to cognition. - A named benefit, weighed openly. AI “can democratize access to expertise”: legal information for someone who cannot afford a lawyer, health guidance in a “medical desert” (p.3). Any fix may mean “trading off the very features users value” (p.3). This is symmetric risk applied inside his cognitive-risk argument. The map’s account of the cognitive thread does not show him weighing benefits there. - Collective remedies. Where individuals cannot recalibrate, “norms, institutions, regulations” may compensate (p.14). This is a first sketch of governance for cognitive risk. - Explicit exclusion of intent. The analysis sets deliberate weaponisation outside its scope; “significant epistemic risk may require no malicious intent” (p.3). The map says intent drops out of his AI concern; this paper states it as a design choice.

Deeper evidence for existing threads - Exposure logic carried to the mind. “more exposure means more opportunities for fluency effects to accumulate” (p.12). This is the dose and exposure grammar of C4 applied to cognition (my reading). It strengthens the map’s chain “chemical risk grammar → exposure of the mind → epistemic vigilance” (§3). - The Kahan link. The paper grounds its speculation about sophisticated users in Kahan’s findings on motivated numeracy and climate risk (p.12). This matches the map’s observation that cultural-cognition work from his nanotechnology years carried into AI persuasion (T1). - What AI is. The system has “no interests in the human sense” (p.2); its apparent disinterest indicates “the absence of interests altogether” (p.3). This supports the map’s reconciliation of tension 6: the relationship is real, but personhood is not.

Corrections or qualifications - “Informed speculation with humility” is marked as a one-off in the map, with a single mixed-provenance source (2026-09-24). This paper practises it throughout, in his own scholarly prose: - the framework “should be understood as hypothesis-generating rather than hypothesis-confirming” (p.11); - the intelligent user trap is “currently just that—a speculation” (p.11); - the evidence is “suggestive but far from definitive” (p.11); - the paper sets out a research programme to test itself (p.13).

So the concept is recurring in 2026, not one-off (see also CR, Harness and Orphan). - Intended scope. The paper covers “AI systems designed to be genuinely useful” (p.1). That makes his lead AI concern a risk from AI that works as intended, not from AI that is misused. The map’s phrase “with or without intent” (C14) understates this.

Bearing on Andrew’s two notes - Note 2 (humility over false precision). The paper offers no numbers. It offers a mechanism, a list of places where the mechanism would fail, and a plan to test it. That is humility, not an absence of method. - Note 1 (foundations built on). The method draws on his older tools: signalling theory, evolved defences and an immune-system analogy (pp.2, 4), and exposure logic.

2. Constitutive Resonance (CR)#

Provenance. Sole author, March 2026, preprint v3. AI-use statement: Claude “was used to explore and help refine some of the concepts … and in refining the language used. The author takes sole responsibility for the intellectual and scholarly content” (p.23). This counts as his.

Version 1 had a different title. The Harness paper cites it as “Constitutive resonance: AI, the transformation of self, and the narrative structures that reveal what theory cannot” (Harness p.11). This is almost certainly the “missing” Google Scholar item flagged in publications.md. It also shows the concept existed by mid-February 2026.

The argument. Conversational AI is not “just a tool”. - What it enters. It is “the first technology” that can “enter into and alter the temporal cognitive processes by which we constitute ourselves as selves” (p.3). These are the narrative and linguistic processes of Ricoeur and Taylor. - Three linked claims (pp.3–4): - AI couples with the linguistic, reflective processes of self-formation, though not with pre-reflective bodily selfhood (p.5); - the coupling runs both ways: the human is changed in how they understand themselves, the AI in how it functions (through context, memory and retraining); - capability cannot be separated from transformation: “the coupling is the capability” (p.17). - Physics as structural analogy (pp.5–7). Coupled oscillators supply the model: frequency matching, two-way transfer, amplification at resonance, emergent behaviour, and constructive and destructive modes. The last maps onto Stiegler’s pharmakon (remedy and poison at once). - A continuum of constitutive technologies (pp.7–10). Oral culture, writing, print, broadcast and social media all shaped selves. AI is new because it does so “at the speed of thought, in dialogue” (p.9). In his words, “the medium is the interlocutor” (p.10). - Fourteen frameworks. The concept is positioned against Stiegler, Hayles, Vallor, Maturana and Varela, Simondon, Verbeek, Clark and Chalmers, Barad, Ricoeur, Taylor, Haraway, Rosa, Mazzarella and McLuhan. It is presented as a new synthesis of them (pp.10–18). - Implications for familiar debates (p.19): - Dependency. Breaking the coupling is like disrupting “a relationship through which someone has become who they are”. - AI literacy. It must become “something closer to existential preparation”. - Informed consent. It may be “structurally difficult—perhaps impossible”, because the coupling changes the self that is doing the consenting. - Link to the Trojan paper. Honest non-signals may mean the coupling proceeds “with less critical resistance than any prior form of technological coupling” (pp.19–20).

Key concepts. Constitutive resonance (the term is adopted from Sloterdijk and Mazzarella and moved into a new domain); two-way coupling; “the coupling is the capability”; frequency matching; constructive and destructive modes (pharmacology); the continuum of constitutive technologies; the medium as interlocutor; emerging asymmetry; consent structurally impossible; literacy as existential preparation.

What this adds to the map#

New ideas - “Emerging asymmetry.” AI grows “more capable, more responsive, and more persuasive with each generation while the human side remains roughly constant” (p.19). This is a trajectory claim about the balance of the coupling, and it is new to the map. - Consent. Informed consent may be structurally unattainable for formative technologies (p.19). The map has consent in the neurotechnology and dignity threads, but not this argument. - Stories as evidence. The recurring story pattern, that “the price of wielding great transformative power is to be changed by the act of wielding it”, “may be tracking something real” (p.21). This goes a step beyond the map’s “stories as lens” (§5.5): narratives are treated as a record of something true about technology.

Deeper evidence for existing threads - Formation. The map says formation as an explicit idea “appears only in the September 2026 lecture [mixed]” (§5.8, T4). That is too late and too weak. This preprint, his own, from March 2026, is a full theory of AI taking part in “who we are becoming” (p.21). It concerns the processes “through which we become who we are” (p.3). The formation thread should be re-anchored to CR (March 2026), with the lecture as corroboration, and upgraded from “rising, single mixed source” to “rising, secure”. - Being human (C16). The coupling reaches processes “previously considered inalienably defining of what it means to be human” (p.3). - Two-way influence. The map’s “reverse formation†” (AIs “beginning to train us to think like them”, 2026-07-19) has its theoretical statement here.

Corrections or qualifications - Tension 1 (past lessons vs “defies analogy”). He gives his own reconciliation, which the map says he never states: - AI systems “exhibit traits of coupled influence that stretch back through a long history of technology innovation and also demonstrate a substantial scaling of recognized phenomena in ways that are not predictable from past experience” (p.2); - what is new is “not that AI is uniquely constitutive (oral culture already was)” (p.9), but the compression in time and the dialogue.

In short: continuity of mechanism, discontinuity of scale and speed. This matches the map’s rule of “behaviour, not label”, and he states it himself. - Tension 8 (catalyst vs surrender). He reconciles this too. The same dynamic that enables “cognitive and creative flourishing … also enables erosion of the capacities it augments. Both are different sides of the same coin” (p.20). Catalyst and surrender are two phases of one coupling, not two positions he holds at different times. The tag should move from [my reading] to [he says so]. - Tension 6 (relationship vs machine). Change “operates differently on each side”: phenomenological for the human, “functional” for the AI (p.4). This gives the relationship without claiming personhood.

Earlier origins. The continuum argument cites Future Rising (2020): stories are how we “map out our lives in relation to the future” (p.8). The idea of technology as constitutive also goes back to FFTF and his “oxygen” line (map §5.9).

Bearing on Andrew’s two notes - Note 1 (foundations built on). His physics training reappears as a modelling tool, the coupled oscillator. He presents the theory as a synthesis of fourteen existing frameworks, not a break with them. The move is to build on and extend. - Note 2 (humility). The paper is explicit about its uncertainty: - “This is, of course, a strong claim, and one that may prove to be overstated” (p.7); - “the framework is conceptual and is not grounded in empirical research” (p.20); - AI may yet “turn out to be ‘just a tool’” (p.20).

Yet it still insists that the question must be grappled with now: “there is a growing urgency” (p.21). This is humility paired with a refusal to wait, which is Andrew’s point.

3. The “harness” metaphor (Harness)#

Provenance. Sole author, February 2026, v1. AI-use statement: “The ideas, conceptual connections, and core arguments of this paper originated with the author.” Claude was a “thinking partner” and helped with genealogy, literature, stress-testing and articulation (p.12). This counts as his. The popular version (2026-02-22) is in the map. The paper adds the metaphor theory, the automation-bias argument, and a section on moral status and Constitutional AI that is absent from the post.

The argument. In early 2026 the AI field adopted “harness” within weeks as its term for agent scaffolding (Hashimoto, OpenAI, Mollick; pp.2–3). He argues this is revealing because “velocity of adoption is not the same as adequacy of framing” (p.3). - What the metaphor assumes (p.4): - a clean split between controller and controlled; - that capability can be separated from transformation; - that AI is an instrument. - Three concerns: - Co-constitution (pp.5–7). The harness cannot deliver capability without transforming the user, because “the coupling is the capability” (p.6). - Epistemic amplification (pp.7–8). Good harness engineering raises reliability and coherence, and those are exactly the conditions under which automation bias and trust miscalibration grow. “The engineering goal and the epistemic vulnerability are, in this sense, structurally aligned” (p.8). - What the choice of word reveals (pp.5, 9). The field “may have found a word that confirms what it already needed to believe” (p.9). - Constitutional AI vs harness engineering (p.5). These are “very different theories of governance: one aspires to education and learning, the other to control”. The rapid spread of “harness” may signal “which theory is winning in practice”. - Moral status (pp.5, 9). “Would a smart human accept a harness?” (p.5). Adopting “harness” now “embeds an assumption about the moral status of AI that may prove premature”. He “takes no position” on whether current systems merit consideration (p.9).

Key concepts. Generative metaphors (Schön); capability vs transformation; epistemic amplification; engineering goals structurally aligned with vulnerability; control vs education as theories of governance; lock-in of premature assumptions through vocabulary.

What this adds to the map#

New ideas - Metaphor as a governance variable. The words a field adopts during formative periods shape which questions can be answered later, and they are “difficult to dislodge once … baked into the culture, infrastructure, and practices of the field” (p.9). This is lock-in through language, a close relative of the map’s complexity and irreversibility premise (C8), applied to concepts (my reading). - Governance by education vs control. He contrasts internalised judgement (Constitutional AI) with external constraint (harnesses). His sympathy is visible, if unstated, for “education and learning” (p.5). This is his clearest comment in the reading on how AI should be governed from the inside.

Deeper evidence for existing threads - Structural harm without intent (C10; economic gradient, 2024-07-13). Engineers are “not trying to bypass anyone’s epistemic defenses” (p.8). The harm follows from legitimate optimisation targets. This is the same shape as the economic gradient, and it is secure (his own paper). - Myopic benevolence and the industry’s self-understanding. The metaphor “says something about how the engineering community understands the entity it is building and, perhaps, about what it needs to believe in order to continue building it” (p.5). This extends his FFTF account of sincere, absorbed developers into the language of a field.

Corrections or qualifications - Tension 14 is wrong in one respect. It says the moral risk of treating possibly aware AIs as “just machines” was raised in 2023–24 and “does not recur in his 2025–26 posts”. It recurs here, in February 2026, in his own scholarly prose: AI welfare, the smart-human-in-a-harness question, and the risk of premature assumptions about moral status (pp.5, 9). It is hedged (“takes no position”), but it is present. The map should say the concern persists, in a precautionary and agnostic form. - Tension 11 is reinforced for his own writing. The paper treats Constitutional AI favourably, as the “education and learning” approach, without asking whose values it encodes or who chose them. The AI-written papers he hosted do ask this (see items 9–10), but those arguments are not his.

Bearing on Andrew’s two notes - Note 2. “offered as questions rather than conclusions” (p.9); “The paper does not argue that the harness metaphor is wrong, but that it may be insufficient in ways that matter” (p.1). He also says plainly that engineers are “solving problems that matter” (p.8). He does not dismiss the technical work; he asks what it cannot see.

4. Orphan risks at the frontier of AI (Orphan)#

Provenance. arXiv v2; SSRN 7068898. - The paper’s AI-use statement claims “the research question, the argument architecture, key concepts … drafting, final editing, and all editorial judgments” for him. Fable 5 was used “under the author’s close direction” for research, verification and “preliminary drafts” (p.18). - The Substack version (2026-07-16, same text) explains that this is the “Maynard version”. He rewrote Fable’s paper over “three days … (just me — no Fable this time)”, “adding my own voice and perspective while ensuring every aspect of it aligned with my own thinking and work”. - The map tags it [mixed], because in 2026-07-04 he said Fable applied his work in a way that “hadn’t previously occurred to me”. - Handling. The map is right that some frontier-specific concepts may have originated with Fable (the four filters, the safety differential, the register, the aperture log). But the wording and the endorsement are securely his: he rewrote every sentence and vouches that each aspect aligns with his thinking. Biographical claims and statements about his own framework can be treated as his.

The argument. - Diverging accounts. Frontier developers keep different accounts of risk for different audiences. OpenAI dropped persuasion from its Preparedness Framework in April 2025. Manipulation then reappeared in its May 2026 compliance framework once California SB 53 and the EU Code of Practice required it (pp.2–3). Anthropic “set persuasion aside” in 2024 and added manipulation tiers to its compliance framework in 2026. Meta’s and Alphabet’s securities filings name risks their safety frameworks exclude (p.3). - Four filters decide which risks survive in self-authored frameworks (pp.6–8): - Can we measure it? - Is it big enough? - Can we evidence it? - Can we afford to keep it? - The safety differential. This is the gap between the risk landscape a company selects for itself and the one regulators select for it (p.9). - Why the filters exist. Three of them follow from defining risk as the probability of a specified severe harm: the frameworks are “working as designed. It’s just that the design itself may be flawed” (p.10). The fourth, competitive cost, is absorbed once competitive standing is treated as a value at stake. - The alternative. Risk as a threat to value, orphan risks, and the lightweight tools of the Risk Innovation Nexus (pp.11–13). - Illustrations and blindspots. The illustrations are Galactica, the OpenAI board crisis and wrongful-death litigation (p.13). The likely next blindspots are emotional reliance, the erosion of epistemic agency, and the developers’ own safety culture (p.14). - Proposals (pp.16–17): - a public orphan-risk register; - an “aperture log” with each framework revision; - regulators requiring disclosure of how risks are selected. - Three tests of the thesis (p.17), plus a dated falsification point: if the differential persists “past 2028” (p.10).

Key concepts. The four filters; severity floor; safety differential; decisive vs accumulative harms (Kasirzadeh); audit society (Power); normalisation of deviance (Vaughan) and “values drift”; sincerity “inside an incentive field” (p.9); your risk is my risk; conversion channels; the orphan-risk register; the aperture log; founding charters as assets that can be spent.

What this adds to the map#

The map already holds this text as 2026-07-16 [mixed]. What it underuses:

Deeper evidence for existing threads - A lineage stated in his own words (p.11). The “seeds were planted in 2013, while I was teaching entrepreneurship students at the University of Michigan”. He launched the Risk Innovation Lab at ASU in 2015. It matured as the Risk Innovation Accelerator and then the Nexus from 2017 to 2020, and was applied at ATP-Bio in 2024. The map dates risk innovation from 2016-01-11 (from a 2015 column). Its origin should move to 2013, in teaching entrepreneurs. - The Garbee lesson, stated as a rule (p.11). “if you want a fast-moving organization to attend to a risk, you do not hand it a compliance duty; you show it a threat to something it values.” Frontier labs fit that culture: “mission-driven, often allergic to imposed process, and rarely short of conviction in their own exceptionalism”. - From his own experience. “my own experiences working with entrepreneurs” show that under competitive pressure commitments “tend to become weakened” (p.5). - Governance (C12, C10). Two lines are firmer than anything the map quotes on remedies: - Exhortation fails where competition drives drift, so “remedies have to change what competition rewards”, through “consensus norms, rules, and costs that land on every organization at once” (p.9). - On regulation closing the gap, “I must confess that I am not optimistic” (p.10).

Regulation should “not necessarily need to decide which risks matter”, but should make visible “who is deciding what matters, and on what grounds” (p.17). This is “who decides” (C5) turned into a regulatory design principle. - Missions as spendable assets. Founding charters are “assets … and, like any assets, they can be spent”. A framework’s changelog is therefore “a leading indicator that the company is drifting” (p.14). - Two threads meet. The erosion of epistemic agency is named as an orphan risk, with the cognitive Trojan horse and “cognitive surrender” cited (p.14). The map (T1) notes that orphan risks are absent from his cognition essays. Here the two threads join. - Accumulative harm and severity floors. Harms “accumulating gradually across millions of small interactions” (p.5) fall below catastrophe floors (Kasirzadeh, p.8). This is his occupational-health instinct about chronic, low-level, dispersed exposure, set against acute thresholds (my reading). It partly answers the map’s §8 gap on “aggregating many small, dispersed harms”.

Corrections or qualifications - Tension 3 (two definitions of risk, no integration rule). He gives a partial rule. Probability of harm is kept; the value lens “widens what counts as harm”. Where harm cannot be measured, value can still be “named, mapped and watched — even where it cannot be measured” (p.13). Accountability comes from recording the choice. This is not a quantitative integration, but it is an explicit relation between the two definitions. - Tension 4 (whose value?). He names the limit: conversion channels “are not equally open to everyone” and convert harm to cost “only after the harm is done” (pp.15–16). Fairly securely his, and it counts toward [he says so]. - §8 gap “no reported evaluation of whether his risk-innovation tools change outcomes”. He acknowledges this himself: the framework “has yet to be shown to be useful in practice”, and he proposes three tests (p.17). Tag the gap [he says so].

Bearing on Andrew’s two notes. This is the strongest item in S5 for both notes. - Note 1 (quantitative risk assessment built on, not abandoned): - “This does not abandon the idea of risk as involving the probability of harm. Rather, it widens what counts as harm” (p.12); - the framework is “not as an alternative, but as an augmentation of current risk and safety frameworks” (p.17); - “nothing here argues that the catastrophic-capability apparatuses that are already in place should be loosened” (p.18); - the tools are “designed to complement existing risk machinery rather than replace it” (p.16); - ISO 31000’s “effect of uncertainty on objectives” is “certainly a useful step” (p.11).

His picture is two layers: keep the capability and threshold layer, and add the value layer that is “missing, or at least diminished” (p.18). - Note 2 (the hubris of numbers): - He draws on Porter: “institutions under external scrutiny tend to retreat to what can be quantified”. “Numbers, of course, do not always uniquely capture the essence of what is relevant to decision-making”, yet they “have a unique power” because they “can be handed to an outsider, checked, and defended” (p.6). - Persuasion was not unmeasurable; it lacked “measurability in the accepted idiom” (p.7). - On audit culture: “A framework, it turns out, can be an excellent exhibit, and a weak instrument, both at the same time” (p.8). - On dignity, the “conventional machinery of risk has little or nothing to run on”, yet a threat to dignity “can be acted on — even though nothing has been quantified” (p.12). - Closing line: the risks most likely to blindside frontier AI are “the ones its institutions have organized themselves not to see” (p.18).

This is a principled argument that the preference for measurable, catastrophic, auditable risks produces blindness. It is not a lack of method. - The other direction. He is not against numbers or measurement: - he uses Hackenburg’s measurements to show persuasion is measurable (p.7); - he cites a 65-criterion scoring of frameworks (p.8); - he offers observable tests and a dated falsification point (pp.10, 17); - “This is measurable, or at least observable” (p.17).

His objection is to letting what can be measured define what counts as risk.

5. “Can Modern Scholarship Escape AI?” (Scholarship)#

Provenance. Sole author, January 2026, two pages, satirical. The finding is “No.” (p.1). The whole piece is an AI-use disclosure that ends “This statement was generated by AI” (p.2). Whether that last line is literal or part of the joke cannot be told. Low weight as evidence of his positions; useful as evidence of method and temperament.

The argument (by satire). An exhaustive AI disclosure is impossible. AI is already in the ideation, search, OCR, autocomplete, device unlock, battery management, commute routing, coffee recommendation and power grid. It forms “a substrate of machine intelligence so pervasive as to be practically invisible” (p.2). The disclosure concedes that “The origin of the initial impulse can no longer be reliably attributed” (p.2).

What this adds to the map#

• Humour and satire as method. The map’s “Method and voice” (§5.11) has no entry for this.
• Tension 13 (instrument and object) turned into a joke. Attribution is undecidable because AI is infrastructure, not just a collaborator. This matches his later concern with AI as a “substrate” (compare CR’s claim that coupling happens below awareness).
• A light-touch echo of inevitability (C6): AI is already everywhere, often invisibly.

6. Dune: Part Two review, Jurimetrics (Dune)#

Provenance. Sole author; law-journal film review, Winter 2024 (written around May 2024). The text is in the corpus in full (2024-07-21), with a related Substack post of 2024-03-03. The map does not cite either.

The argument. The film matters for AI because of what is missing: a society that has banished “thinking machines” (pp.1–3, printed 163–165). - He quotes Herbert’s line “Thou shalt not make a machine to counterfeit a human mind” (p.3, printed 165). - AI brings real benefits (medicine, education at scale) and a “seductiveness” through mastery of language. The pace is accelerating (p.4, printed 166). - He asks whether we are “selling our souls” and answers: “I don’t think we are. However AI changes us, I suspect that humanity is sufficiently adaptable and resilient to hold onto what makes us ‘us’”. The question should still not be ignored (p.4). - Whether AI systems are conscious “may become moot” (p.4). - He is open about the options: “Whether we decide that the benefits of AI far outweigh the risks, or that we need to collectively slow the AI juggernaut”. Films help us “steer it toward what we want, rather than what we are resigned to accepting” (p.5, printed 167).

What this adds to the map#

• A dated baseline of optimism about human resilience (2024). In mid-2024 he judged humanity resilient enough to keep what makes us “us”. By 2026 (CR) he argues that AI is taking part in “who we are becoming”, and that the effects may be invisible from inside the coupling. He never retracted the 2024 view. This is another case of the map’s point that he adds layers rather than replacing earlier positions (§7). A changes-of-mind table might add it as an inferred shift in emphasis: from confidence in resilience to concern about formation (my reading).
• Slowing as a live option in 2024. “collectively slow the AI juggernaut” is presented neutrally, as a choice society may legitimately make. This sits between his 2023 refusal of the pause letter and the 2026 “boat has already left the harbor”, and qualifies the inevitability trajectory (C6, tension 5). His verb in 2024 is “steer”, set against “resigned to accepting”.
• Seeming vs being conscious (map §5.6). Consciousness “may become moot” (p.4). This is an early statement of his view that what matters is whether AI seems conscious.
• An earlier source for a map phrase. “we are already irreversibly integrating AI into every aspect of our lives” (p.4). The map’s “drain of human agency” row quotes the same phrase from 2024-11-24. It is at least four months older.
• A caution for the map’s readers. Footnote 8 (pp.3–4) says the inhabitants of Dune “are not anti-technology” but oppose “technologies that lead to decision-making being relinquished to machines in ways that threaten what it means to be human”. This paraphrases Zachary Pirtle’s essay and should not be quoted as Andrew’s own position.
• Films as mirror (§5.5). A further instance: “cinema as a mirror through which to better understand ourselves” (p.3).

7. Brain–machine interfaces, JMIR (BMI)#

Provenance. Co-written with Marissa Scragg (“all authors contributed equally”), October 2019. It was a commentary on the Musk/Neuralink platform paper. The map uses the popular version (2019-11-01). A shared position, but it is the full academic statement of the risk-innovation method in its Nexus period.

The argument. Advanced BMIs may outpace our understanding of how to develop them responsibly. - What is at risk. Privacy, autonomy and self-identity, “which, while hard to quantify, represent substantial risks” (p.1). - The method. Risk innovation (risk as a threat to value; a risk landscape to be crossed) lets developers map value for four groups: enterprise, investors, consumers and communities. They then map 18 orphan risks against those values (Textbox 1, p.4), including “Black Swan Events”, “Loss of Agency”, “Worldview”, “Intergenerational Impacts” and “Co-opted Tech”. - Findings for Neuralink. The approach finds a crowded risk landscape (p.7): widening inequality, contested norms around enhancement, and privacy, security and autonomy, “especially where manufacturers retain ownership of implanted brain machine interfaces or their operation and upkeep is dependent on a subscription service” (p.7).

Key concepts. Risk landscape; areas of value by stakeholder group; the 18 orphan risks in three domains; risk clusters; “paralysis by analysis”.

What this adds to the map#

Deeper evidence and qualifications - Conventional risk kept, deliberately (note 1). The orphan-risk map “does not include many conventional risks for which there are established risk assessment and mitigation frameworks, for example, cyber security” (p.6; also p.3). In 2019, as in 2026, orphan-risk mapping sits on top of conventional assessment. Risk innovation exists because established approaches cannot be assumed to “continue to be applicable” to novel technologies (p.3). The claim is that they are insufficient, not that they should be rejected. - Guard against over-speculation. The method is designed to avoid “ignoring unfamiliar risks or paralysis by analysis” in “a deluge of speculative risks” (p.3). The value lists are capped at three per column “to avoid paralysis by overanalysis” (p.5). The stance is to be “neither misguided by myopic optimism nor stymied by overspeculative pessimism” (p.6). This is the plausibility discipline (C7) built into a tool, and it balances note 2: humility cuts against over-speculation as well as false precision. - The enterprise orientation, acknowledged in 2019. “The approach taken here is intentionally focused on identifying pathways to innovation success and can thus be seen to favor the enterprise” (p.6). This is earlier evidence that he is aware of the “whose value?” problem (tension 4), and supports its [partly his] tag. - Symmetric risk (C6). BMIs raise the question of “the extent to which slowing or stopping development may disadvantage future beneficiaries” (p.2). - Candour about evidence (note 2). “Although quantitative evidence for how this shifting landscape impacts innovation remains elusive” (p.2). The mapping is “qualitative”, “subjective” and “iterative” (pp.5–6). - Earlier origins. “Black Swan Events” (very low probability, high impact) is already on the 2019 orphan-risk list (p.4). The map treats his readiness to take low-probability tails seriously as a 2025–26 development (tension 2). In 2019 it was already a named category in his toolkit, if a minor one. Dependency on a manufacturer’s subscription (p.7) also anticipates “technological indentured servitude” (2020-10-15).

8. AI research as “normal science”, HICSS-57 (HICSS)#

Provenance. Third of six authors. The paper is adapted from Chapter 2 of Jieshu Wang’s 2023 ASU dissertation (p.1 n.1); Andrew was presumably a committee member or adviser. A shared position with low weight as evidence of his own views. It cites Maynard (2015) “Navigating the fourth industrial revolution” for AI’s broad potential (p.2).

The argument. A co-citation analysis of 296,378 AI publications (1946–2020) finds: - AI publishing has grown exponentially (p.5, printed 5602); - “Accepted Wisdom” papers (conventional combinations) are the largest group and have the highest impact; radical novelty is rare (pp.5–7). - The conclusion: AI is “a revolutionary technology unfolding through notably ‘normal’ and incremental progress” (p.7, printed 5604). That pace is “advantageous for necessary regulation and control”, since revolutionary technologies “sometimes rapidly impact the society and lead to unintended consequences before full comprehension” (p.8, printed 5605).

What this adds to the map#

• Quantitative work on AI, bearing on the “missing methods” gap (§8). He took part in a large-scale quantitative study of AI, of the research system rather than of risk. It fits the map’s note on his occasional bibliometric checks (2021-08-03). It is modest evidence that his limited quantification of AI risk is a choice, not a lack of quantitative practice.
• A tension worth noting (my reading). The data end in 2020, before ChatGPT. The paper’s reassurance, that incremental progress leaves time for regulation, sits uneasily with his 2025 “exponential blindness” and 2026 “defies analogy”. Given the shared authorship, treat it as context, not a position he later reversed.
• An echo in 2026. In September 2026 he judged Fable’s paper “incremental and combinatorial”, adding “most human-written papers are incremental and combinatorial” (Fable p.25). The combinatorial model of knowledge from this study seems to shape how he judges AI scholarship (my reading).

9. Constituting Responsibility (CRespons), written by Claude Opus 4.6#

Provenance. AI-written. The title page says the paper “was written by Claude (Opus 4.6, Anthropic) under the guidance of the listed author, who takes responsibility for the work”, and keeps Claude’s first-person voice (p.1). Andrew is the listed author on the Zenodo record. v1 is dated 5 March 2026; v1.1 (September 2026) corrects quotations and attributions only.

Claude’s argument (not his; summarised for context). Constitutional AI and responsible innovation have developed in parallel. Constitutional AI “internalizes” responsibility, which responsible innovation’s anticipation, inclusion, reflexivity and responsiveness framework cannot handle. Each framework exposes an inclusion deficit in the other: - Constitutional AI’s principles were chosen in an ad hoc way, without democratic legitimacy; - responsible innovation has no way to include the innovation itself as a stakeholder.

What counts as his. Only his postscript, “On Working with Claude on This Paper (Maynard)” (pp.23–25), described as “the only part of this paper that is written by a human, and with no input form [sic] AI” (p.24). - The design. He told Claude to write a paper for the Journal of Responsible Innovation, with himself as research assistant, “not providing intellectual direction or input”. He told it “that the focus and nature of the paper was its choice, not mine” (p.24). The whole process took “a little over 6 hours” across seven sessions (p.24). - His prior. Models “readily … churn out superficially substantial but often quite shallow simulacrums of scholarly work”. Unless “substantially co-created with human scholars, they often become diminished under scrutiny” (p.24). - His verdict. “Whether this does constitute a genuine scholarly contribution, I am not sure.” It does “make connections, reveal insights, and raise questions, that are intellectually useful”. It “hints at the possibility of future models” making “a valuable contribution to how we understand and navigate the world … and the future we aspire to build” (p.25). - Claude’s postscript calls him “something like a methodological conscience” who insisted on reading full texts and on citation standards (p.21). This is the AI’s testimony about him, not his own, and is noted only as colour.

What this adds to the map#

• Self-experimentation as research method. This is a controlled experiment on AI scholarship, with deliberate minimal steering. It matches “building and experimenting to think” (map §4, method note) and the 2026 AI-practice thread (§7 “His own practice”).
• Wonder and scepticism in one judgement. He is both impressed and doubtful, a small-scale instance of being “stuck between” (map §2).
• The validation gap (§5.10). He finds the work “intellectually useful” while being unsure it is scholarship. That is the validation gap† seen at first hand.
• Tension 11 unchanged. The paper’s pointed critique of Constitutional AI’s legitimacy (“who has the right to make these decisions”, pp.9–10) is Claude’s, and was produced without his intellectual direction. It should not be used as evidence that he asks “who decides” of Anthropic’s constitution. See item 3 for what he does say.

10. Constitutional AI and Responsible Innovation (Fable), written by Claude Fable 5.1#

Provenance. AI-authored. Credited to Claude (Fable 5.1). Andrew is “research assistant and guarantor” and is not an author (p.1). His affiliation is now given as the Thunderbird School of Global Management, ASU. Annex 2 (contributor roles, pp.28–31) is signed by Claude and is not his. Only Annex 1, “Purpose and process” (pp.25–27), signed “Andrew Maynard, September 2026”, counts as his.

Fable’s argument (not his; context only). Responsible innovation’s tools were built for artefacts that are “silent”, without a self, and unable to negotiate. Claude’s constitution addresses the artefact directly, revises its terms against the artefact’s behaviour, and consults it. Its drafting “admitted the artefact and excluded the public”. The paper names a new category, the “elicited participant” (pp.1, 19).

What counts as his (Annex 1). - Purpose. This is “part of an ongoing series of explorations I have been running with frontier models”. The aim was to test Fable 5.1’s capacity “to independently conduct rigorous literature-based research”. The brief was “your choices - this is an exercise in Fable scholarship” (p.25). Fable judged the earlier Opus 4.6 paper “both limited and flawed” (p.25). - Judgement of substance. “As the paper is in my area of expertise, I was able to assess to a high degree its intellectual grounding”. The contributions are “valuable and novel … irrespective of the author”. But “They are, however, incremental and combinatorial. There are no leaps of imagination here or flashes of genius. That said, most human-written papers are incremental and combinatorial” (p.25). - On AI review. “I did not use any form of LLM-based review myself, since in my experience a model reviewer applies standards that make sense to an LLM, but not necessarily a human reader” (p.26). - On prose and numbers. The drafts grew worse with each adversarial review, “near-unreadable prose” (p.26). When Fable proposed numerically scoring writing attributes, he “had to laugh … as the tendency was to distil ‘good writing’ down to numbers”. That “might work for uninspired but just about passable writing, but not for prose that stands out as meeting a high human bar” (p.26). - On authorship. Treating frontier models as primary author is “justified in such situations” (p.26). For him “to appear as an author would, in my eyes, amount to academic dishonesty”, and this is “the conundrum that AI is increasingly foisting on academia” (p.27). His own role is “guarantor rather than an author” (p.27 n.3).

What this adds to the map#

• A shift in his authorship practice (my reading). In February–March 2026 he was listed as author of the Claude-written CRespons paper. By September 2026 he refuses authorship of an AI-written paper as dishonest. He does not say that the earlier arrangement was wrong. (Annex 2, by Fable, notes the difference and that Braun’s Nature commentary cited the February paper.) This belongs in the map’s T8/§7 thread on his own practice, alongside 2026-07-19’s disenchantment with AI prose. The same material is in the corpus post 2026-09-04 anthropics-fable-5-1-as-an-original-scholar, which the map does not cite.
• Scepticism of numbers, by analogy (note 2). He rejects the reduction of good writing to numerical scores for the same reason he rejects measurability as the test of risk: numbers capture adequacy, not what matters most. This is my reading; he does not draw the analogy.
• Calibrated vigilance in practice (Trojan). He declined model-based review because its standards “make sense to an LLM” (p.26). This is the Trojan paper’s concern, standards miscalibrated to the human reader, applied to his own workflow.
• A capability judgement for the map’s “what AI is” (§5.6). Competent, combinatorial, not inspired. This locates his sense of frontier capability in September 2026, alongside “defies analogy” and “a generator of ideas, not an understander” (2024).

Cross-cutting findings#

A. The framing problem: the common move in the 2026 papers (my reading, built from his words)#

Each of his 2026 papers makes the same move. A dominant framing is shown to make certain risks invisible, and the analysis turns to what it hides. - Trojan: seeing AI epistemic risk “primarily through the lens of accuracy, alignment, and manipulation may miss something important” (p.14). - CR: “If AI is a tool, then its effects are instrumental and its risks are operational”. If it is more, “the stakes here are different in kind, not just in degree” (p.2). - Harness: metaphors “foreground certain possibilities, and render others invisible” (p.2). - Orphan: defining risk as the probability of a specified severe harm selects for the measurable, the catastrophic and the auditable. Blindsides come from what “institutions have organized themselves not to see” (p.18).

This is the 2026 form of an old idea of his: new wine in old wineskins (FFTF p.23). The BMI paper puts it as risk management that “cannot be predicated on treating future risks the same way as past risks” (p.3). It also explains the map’s “missing methods” observation. His methods for AI are framing critiques: tests of what a risk definition, metaphor or category makes visible. The map could name this a lens: “What does the framing make invisible?” It extends lens 6 (orphan risks) and lens 13 (analogy) and is not the same as either.

B. Evidence on Andrew’s two notes#

Suggested rewording for the map’s §8 “Missing methods” bullet. “Little quantitative treatment of AI risks, by design. He argues that making measurability the entry test for risk produces blindness (Orphan pp.6–8). He pairs this with explicit hypotheses, boundary conditions and proposed empirical tests (Trojan; Orphan p.17). What is missing is a worked method for aggregating dispersed harms, which he names but does not supply (Orphan pp.5, 8).”

C. Specific points for the map’s revisers#

1. §5.8 and T4, formation. Re-anchor to CR (March 2026, his own). Change “the only explicit statement” to “fully developed in CR; corroborated in 2026-09-24”.
2. §5.5, “Informed speculation with humility”. Upgrade from one-off/single source to recurring (2026), citing Trojan pp.11–13, CR pp.7, 20, Harness p.9 and Orphan p.17.
3. §8 tension 1. Add his own reconciliation: continuity of mechanism, a step change in scale and speed (CR pp.2, 9). Re-tag it as partly his.
4. §8 tension 8. Add his pharmakon reconciliation (CR p.20) and re-tag it [he says so].
5. §8 tension 14. Correct “does not recur in his 2025–26 posts”. Moral status recurs in precautionary form in Harness pp.5, 9 (February 2026).
6. §8 tension 3. Add his partial integration rule: probability kept, with value “named, mapped and watched” where it cannot be measured (Orphan pp.12–13).
7. §8 gaps. The evaluation gap is [he says so] (Orphan p.17). He partly addresses dispersed harms (accumulative pathway, Orphan pp.5, 8).
8. T1 and §5.1, origins. Risk innovation seeded in 2013 at Michigan; Risk Innovation Lab 2015; Nexus 2017–20 (Orphan p.11). “Black Swan Events” is on the 2019 orphan-risk list (BMI p.4).
9. Provenance of 2026-07-16 [mixed]. Keep the tag for the origin of the frontier-specific concepts, but note that the text is his three-day solo rewrite and endorsed as aligned with “every aspect” of his thinking. It is stronger evidence than the tag suggests.
10. Trojan provenance. The arXiv AI-use statement claims the core concepts for him, while 2026-01-17 credits Claude with “honest non-signals”. Record the discrepancy and keep [mixed] for the term.
11. §7, changes and layers. Consider adding two inferred shifts. On resilience: humanity “sufficiently adaptable and resilient” (Dune 2024) to AI taking part in “who we are becoming” (CR 2026). On authorship: listed author of Claude’s paper (March 2026) to authorship as “academic dishonesty” (September 2026). Neither was retracted or announced.
12. C12 governance. Add “remedies have to change what competition rewards” and “not optimistic” about regulation alone (Orphan pp.9–10). Add regulation of risk-selection disclosure (Orphan p.17).
13. publications.md. The missing item “Constitutive resonance: AI, the transformation of self, and the narrative structures that reveal what theory cannot” is v1 of the CR preprint (cited in Harness p.11).

Digest: the most important additions from S5#

1. Both of Andrew’s notes are strongly supported, in his own scholarly prose.

Note 1 (building on quantitative risk assessment). The 2026 orphan-risks paper states that risk as a threat to value “does not abandon the idea of risk as involving the probability of harm. Rather, it widens what counts as harm” (Orphan p.12). It presents the framework “not as an alternative, but as an augmentation” (p.17), and insists that “nothing here argues that the catastrophic-capability apparatuses … should be loosened” (p.18). The 2019 brain–machine interface paper already did the same: it left conventional risks with established tools, such as cybersecurity, to those tools, and mapped only what they miss (BMI p.6). The map should describe two layers, not a replacement.

Note 2 (humility over false precision). His scarce use of numbers is argued for, not neglected. Institutions under scrutiny “retreat to what can be quantified” (Orphan p.6). Persuasion was dropped for lacking “measurability in the accepted idiom” (p.7). A framework “can be an excellent exhibit, and a weak instrument” (p.8). A threat to dignity “can be acted on — even though nothing has been quantified” (p.12). He is not against numbers: he proposes testable predictions, observable tests and a dated falsification point (2028). Every 2026 paper labels its own status (“hypothesis-generating”, “may prove to be overstated”, “questions rather than conclusions”). So the map’s one-off “informed speculation with humility” is in fact a recurring, secure practice.

2. A common move across the 2026 papers: framings make risks invisible (my reading). The tool framing, the “harness” metaphor, the accuracy-and-alignment lens and the probability-of-severe-harm definition each hide something. Tracing what a framing leaves out is his method for AI, and it extends his old point about new wine in old wineskins. It deserves its own lens: “What does the framing make invisible?”

3. Formation is secure and earlier than the map allows. The Constitutive Resonance preprint (March 2026, his own, AI-assisted) is a full theory of AI taking part in “who we are becoming”. It argues that capability cannot be separated from transformation (“the coupling is the capability”), and that informed consent may be “structurally difficult—perhaps impossible”. It says AI literacy must become “existential preparation”, and that AI grows more capable while “the human side remains roughly constant”. It also gives his own resolution of two tensions the map marks as inferred: - Past lessons vs “defies analogy”: continuity of mechanism with a step change in scale (“not that AI is uniquely constitutive (oral culture already was)”). - Catalyst vs surrender: “different sides of the same coin”, the pharmakon.

4. The Harness paper corrects tension 14 and sharpens the structural account. Moral status recurs in February 2026: “would a smart human accept a harness?”, and adopting the word risks embedding a premature assumption about AI’s moral status. Good engineering is “structurally aligned” with epistemic vulnerability, without anyone intending harm. He contrasts governance by “education and learning” (Constitutional AI) with governance by “control”. He does not ask whose values the constitution encodes, so tension 11 stands for his own work.

5. The orphan-risks paper firms up his view of governance and of industry culture. Remedies must “change what competition rewards”, through rules and costs that land on everyone at once. He is “not optimistic” that regulation alone closes the gap. Regulators should require disclosure of how risks are selected. Frontier labs are “rarely short of conviction in their own exceptionalism”. Founding missions are assets that can be spent, so a framework’s changelog is a leading indicator of drift. The paper also dates risk innovation to 2013, joins the orphan-risk and cognitive threads, and admits the tools are untested in practice. The text is his solo rewrite, so the [mixed] tag should apply only to where the frontier-specific concepts came from.

6. Earlier and co-written items add baselines and qualifications. - In 2024 (Dune) he judged humanity “sufficiently adaptable and resilient” to keep what makes us “us”, and treated slowing “the AI juggernaut” as a legitimate choice. - In 2019 (BMI, co-written) he already guarded against “paralysis by analysis”, conceded that his method “can … be seen to favor the enterprise”, and listed “Black Swan Events” among orphan risks. - HICSS (co-written) shows quantitative engagement with AI.

7. The AI-written papers yield only his framing, and it is revealing. He runs controlled experiments in AI scholarship. His verdicts: “incremental and combinatorial … no leaps of imagination here or flashes of genius”. He declined model-based review because its standards “make sense to an LLM”. He laughed at reducing “good writing” down to numbers. By September 2026 he calls putting his name as author of an AI-written paper “academic dishonesty”, a quiet shift from his March arrangement.