Late Lessons, Jensen Huang and AI
Markdown version

M5. Governance and who decides: Andrew Maynard’s work as a lens on Jensen Huang, the AI industry and Late lessons from early warnings#

Prepared 26 September 2026 with extensive AI assistance, at Maynard’s request, and subject to his review. One of a set of analyses that read current AI developments, Jensen Huang’s September 2026 conversation with Ezra Klein, and the European Environment Agency’s Late lessons from early warnings reports through Maynard’s published thinking. This one covers governance: who should decide how a transformative technology develops, through what institutions, and on whose evidence. It is analysis, not advocacy, and is not written in Maynard’s voice.

Sources and conventions. - Maynard’s positions come from 05-maynard-risk-and-ai-map.md, the thematic syntheses (especially T5 and T7), the supplementary reports S1–S7 and the original texts. Posts are cited by date, following the map’s convention (where a date carries two posts, a bare date means the first listed in the map); the slug of every post cited is given in the list at the end of this document. Supplementary items are cited by the map’s keys (“Testimony 2007 p.30”, “NANO 2026”); Films from the Future as “FFTF p.X”. - Each claim about Maynard’s position is labelled [Stated] (he has said it; cited), [Implied] (follows directly from stated positions; cited) or [Inferred] (this analysis’s reading, with reasoning and confidence). - [mixed] marks the frontier-AI orphan-risks paper (2026-07-16), first drafted by an AI model and rewritten by him, and his 8 September lecture as drafted into prose by an AI model and corrected by him (2026-09-24). Where possible each [mixed] point is paired with his own earlier prose, cited first. Where the orphan-risks paper is the only source (the “de facto layer of public governance” claim, the disclosure register and log, and the public record of frameworks), the text says so. - Maynard and Garbee (2019), adapted as 2019-08-13 responsible-innovation, is weighted as his own thinking. Other co-written items are shared positions and flagged: the 2010 Handbook conclusions (with Bowman and Hodge); “Too small to overlook” (with Rejeski, 2009); Hansen, Maynard, Baun and Tickner (2008); the 2010 WEF CETI proposal (co-drafted with Tim Harper); Maynard, Bowman and Hodge (Nat. Mater. 2011); JLME 2024 (first of five authors); Hyun et al. 2024 (eighth of eleven); the 2017 Guardian piece (with Stilgoe); and the 2013 Late lessons chapter LL2-22 (with Hansen, Baun, Tickner and Bowman). - Huang is quoted from the official New York Times transcript of The Ezra Klein Show (23 September 2026), with approximate timestamps from the corrected machine transcript. Other Huang statements are cited as in 02-huang-analysis.md (HA). Late Lessons lens entries (I5, G2 and so on) are from 01-late-lessons-analysis.md §6 (LLA); the comparison is 03-late-lessons-and-huang.md (LLH).

1. Summary#

Maynard’s governance thinking starts from a claim he has held since 2006: because harm and safety are socially defined, who decides is part of every risk question, and no single actor should decide alone, industry and technical experts least of all. Layers have been added over twenty years rather than swapped out: strong, independent capacity for risk research and a warning against bodies that both promote and oversee a technology (2006–2011); evidence that voluntary industry reporting fails (2009, co-written); proposals for independent institutions to anticipate emerging technologies’ problems (2008, 2010); a critique of self-certified, permissionless innovation (2018, reaffirmed 2025); an account of why sincere entrepreneurs resist top-down responsibility yet cannot rely on good intentions (2015, 2019); a preference for adaptive, multi-stakeholder portfolios with hard law for specific harms; and a structural account of incentives, present from 2006 (PEN 2006 p.32), 2015 (NN 2015-03) and 2019, sharpened in 2024 (the “economic gradient”, 2024-07-13) and formalised in 2026 as sincerity operating “inside an incentive field” whose commitments “tend to soften” under competition, so that remedies must change “what competition rewards” [mixed]. Justice, who pays as well as who decides, runs through all of it. His confidence in remedies has fallen; the principle has not. He says he has no governance solution for AI.

Read against Huang, his work agrees more than a “doomer versus builder” framing suggests: on rejecting doom narratives (a view the lab leaders also share); on puzzlement at labs that warn yet keep building; on the builders’ sincerity; on wariness of rules that favour incumbents; on doubting an AI-specific agency and, for now, AI-specific hard law; on doubting that a coordinated general pause is the answer (for different reasons); and on refusing liability relief. It diverges at the root: on whether safety is an engineering property that builders can own, and on whether existing law and “regulation will come in” after harm can govern a tightly coupled, hard-to-reverse technology. Huang places the model-level gate, its trigger and its evidence inside the firm, with liability, sector regulators and auditors behind it. Maynard’s record, from “good intentions alone will not ensure we see the benefits” (2015) to safety as “a social and political endeavor as well as an engineering challenge” (2024), holds that the firm should not hold that gate alone, with no bad faith required.

His work corroborates the Late Lessons findings on promotion and oversight (in part from a chapter he co-wrote), rules that do not reduce risk, and decisions taken by a few on behalf of many. It extends them with ways to reach firms through what they value and a stronger claim about the public’s standing, and sharpens the gate-centred framing of the comparison: the prior question is who decides what the gate guards against. It points to disclosure of how firms select risks, mandatory information flows, independent evaluation with payment and control separated, worth-based engagement with builders, rules that land on all firms at once, early two-way public engagement and triggers that work in both directions. Its main limits: he has said almost nothing about the specific instruments now in play, and his own engagement mechanisms have been thin since 2018.

2. Maynard’s relevant thinking#

2.1 The organising claim: who decides is part of the risk question#

If risk is a threat to what people value (NN 2015-09) and harm is “a social construct, not a technological one” (2024-06-20), deciding what is “safe” cannot be delegated to engineers [Stated]. The claim predates the value frame: in 2008 he asked Congress, “Who will decide how it is used, and who will pay the cost?” (Testimony 2008 PDF p.2), and asked in print, “Who is reaping the benefits of new nanotech applications, and who is paying the price?” (Bulletin 2008) [Stated]. Leaving profound technological questions “solely to people like scientists, innovators, and politicians” is “an abdication of responsibility” (FFTF p.288), and a developer “probably shouldn’t have complete autonomy over deciding what you do, or the freedom to ignore those whom your products potentially affect” (FFTF p.227) [Stated]. In 2023 it became a title: “Respectfully Erik Schmidt, industry can’t get AI governance right on its own!” (2023-05-15). In 2026, frontier companies, “Good (as in technically capable)” as they are, “simply do not have the perspective and the understanding, and the intellectual breadth and scope, to be able to decide for humanity what this future looks like” (2026-09-24 [mixed]; the same view in his own prose, 2025-01-07) [Stated].

This is not an anti-industry position. Power is not illegitimate: “The challenge we face is not to abdicate power, but to develop ways of understanding and using it in ways that are socially responsible”, and innovators’ “fiduciary responsibility … to investors” is legitimate (FFTF p.44) [Stated]. And the claim binds everyone. Democratically elected governments “have the societal mandate” to steer AI, “but in most cases they lack the imagination, vision, or agility” (2025-01-07). Members of the public “are critically important to this. But you cannot hand a problem of this magnitude over to everyday people and say, ‘Solve it for us’” (2026-09-24 [mixed]; the same split between public standing and expert drafting is in Handbook 2010 p.583, co-written) [Stated].

Who pays. The map treats justice as the moral axis of his work (map §4, commitment 12), and it pairs “who decides” with “who benefits, who bears the harm, and who was asked”. In 2006 he told Congress that “it is ultimately the public—as workers or consumers, for instance—that may bear many of the potential risks” (Testimony 2006 rec. p.53); in 2023 the question was “who decides who will suffer and who will thrive” (2023-10-19); in 2024 an unchecked “economic gradient” would leave individuals “as engines of value creation rather than the primary recipients of created value” (2024-07-13) [Stated]. For AI in September 2026 he listed “the impacts of water and energy use on local infrastructure and economies” among risks that have “risen in significance” (2026-09-15) [Stated]. Legitimate governance, on this record, is judged partly by whether those who bear the costs had a say [Implied].

2.2 The formative layer: independent capacity, promoter and overseer, voluntary reporting (2006–2011)#

His first remedies were concrete and central [Stated]. Before the House science committee in 2006–08 he argued for a nanotechnology risk-research strategy “with teeth”, at least 10% of federal nanotechnology R&D for risk research, a single accountable leader, full transparency, independent research bodies jointly funded by government and industry on the Health Effects Institute model, a federal advisory committee for stakeholder input, and a funded engagement programme for “two-way communication between the developers and users of these technologies” (Testimony 2006 rec. pp.51, 60; Testimony 2007 PDF pp.4–5). “A list is not a research strategy” (Testimony 2006 rec. p.51).

Three diagnoses carry into AI governance: - Promoter and overseer. The nanotechnology initiative was “more attuned to stimulating exploratory science and developing technology applications than providing science in support of oversight” (Testimony 2007 PDF p.30); industry could not lead risk research because it has “an economic incentive to sell products” (PEN 2006 p.32; Hansen et al. 2008 p.446, co-written, makes the same point in other words). “Neither will safe nanotechnologies emerge if the promoters of the technology are calling all the shots”, nor will market mechanisms ensure safety “on their own” (Testimony 2008 PDF pp.8, 11) [Stated]. - The failure of voluntary reporting. In “Too small to overlook” (with David Rejeski, Nature, 2009; his posted draft is the source), the UK’s voluntary nanomaterial reporting scheme drew thirteen submissions in two years, and the US Environmental Protection Agency judged that about 90% of nanoscale materials likely in commerce went unreported. Mandatory reporting was “a welcome one”: “Given the reticence of industry to volunteer information, it will enable regulators to make decisions based on reality rather than speculation”. The same piece wanted rules that reduce “unnecessary burdens on industry while ensuring safe use” [Stated; co-written]. A 2010 chapter he led warned that “in the absence of business being willing to be more transparent and properly self regulate, government will step in” (Handbook 2010 p.581, co-written). - Knowing is not acting. Of the federal government in 2007: “talking about the issues is no substitute for progress, and in addressing possible harm to people and the environment, good intentions are not enough” (Testimony 2007 PDF p.16) [Stated]. Applying the EEA’s lessons to nanotechnology, he and co-authors found that “many governments still call for more information as a substitute for action”, urged building safety in at design “because economic interests are not fully entrenched at that point”, and asked “whether we are applying them effectively enough” (Hansen et al. 2008 pp.444–447) [Stated; co-written].

His World Economic Forum proposals sought an institution that was “Science-based”, “Non-advocacy” and “Non-partisan”, jointly funded yet operating “independently of the funders” (WEF 2008), and a centre “immune to any political or business interests”, because “hierarchical, evidence-based decision-making is not sufficient on its own” (CETI 2010 PDF pp.2, 5, co-drafted with Tim Harper) [Stated]. In 2026 he said that renaming “policy” as “intelligence” “narrowed the original ambition in ways I still have mixed feelings about”, and that his 2008 bottom line “could appear unchanged in almost any serious AI governance document being written today” (Prehistory 2026) [Stated; retrospective]. In 2011 he proposed evidence-based regulatory “trigger points”: specific values of “nine or ten attributes (including size and surface area)” that would prompt action, with the open question of whether “a 1% change” or “a 50% change” should raise concern, and triggers that “must be flexible, so that they can be modified as evidence grows” (Nature 2011) [Stated]. These were quantitative thresholds within hard law, not a rejection of numbers.

2.3 Expertise beyond the technical, and the public’s place#

Governing emerging technologies is a field of expertise that AI’s insiders mostly lack: early AI governance was led by people “light on their expertise in governing emerging technologies successfully” (2023-07-12), and “This ‘leave it to the technical experts’ mentality … never plays out well”, as Monsanto’s “it’s complicated, leave it to us” approach showed (2023-05-15) [Stated]. People “don’t need to understand the inner workings of AI” to judge how it might “threaten what’s important to them” (2023-05-15); “most people have a pretty high level of expertise in what’s important to them and their communities”, and it is this expertise, not public direction of research, that should guide development (FFTF p.222); engagement “is not something that can be introduced once industry has laid the groundwork — that’s a sure-fire way to lose trust” (2023-05-17); the deficit model was “debunked decades ago” (2024-10-13) [Stated]. After the 2023 Senate hearing, with its narrow witness list, he warned that without wide engagement oversight would be “defined and locked in by very limited set of ideas coming from a very small group of players” (2023-05-17) [Stated]. His scepticism of who fills the “insights vacuum” is symmetric, covering advocacy groups as well as companies: “I’m not sure I trust people who have an agenda, who can mobilize fast, and who have the ear of decision makers, to get things right” (2023-04-10) [Stated].

He acknowledges limits: “we still lack the forums, the methodologies and the leadership” for actionable multi-stakeholder dialogue (NN 2015-12 p.1006) [Stated]. His concrete mechanisms were strongest in 2007–08 and thin since 2018 (map §8, tension 9, which he partly acknowledges). The map also reads his remedies as often adding another class of experts, in governance and responsible innovation, which is his own field [Inferred; medium; the map marks this as interpretation]. He has acknowledged a related bias when recommending his own school’s expertise (“I’m admittedly a little biased”, 2016-03-12), though not about his governance proposals as such.

2.4 Permissionless innovation and self-certified responsibility#

His critique is precise, and matters for reading Huang fairly. “Permissionless innovation isn’t necessarily reckless innovation. Rather, it’s innovation that’s conducted in a way that the person doing it thinks is responsible” (FFTF p.162) [Stated]. His Ex Machina example takes safety seriously on its own terms: the AI is built far from civilisation, “surrounded by a natural barrier to prevent her escaping”. The failure is epistemic: “With the best will in the world, a single innovator cannot see the broader context within which they are operating”, so “there need to be checks and balances around who gets to do what in technological innovation, especially where the consequences are potentially widespread and, once out, the genie cannot be put back in the bottle” (FFTF pp.162–166, reposted in 2025-03-02) [Stated]. He credits the other side: “Too much blind speed, and you risk losing your way. But too much caution, and you risk achieving nothing” (FFTF p.163).

In 2025 he called the critique “more relevant now than it was then”, objected to fixing problems “after the fact — rather than anticipating them and navigating around them”, and added a reversibility test: experimenting in reversible, linear systems is fine, but “I’d put breaking people, governance, society, and the planet, in this category!” (2025-03-02) [Stated]. He read the 2025 US AI Action Plan’s “try-first” culture as “essentially an ask forgiveness rather than permission policy” that “prioritizes power before people”, while welcoming sandboxes if “carefully (and responsibly) executed” (2025-07-23) [Stated]. He noted that its support for open-weight models “will be welcomed by many” and that researchers would applaud it, while observing that these were “open source models with strings attached” (2025-07-23).

2.5 Responsible innovation in a culture of entrepreneurship, and its limits#

The chapter with Elizabeth Garbee (2019-08-13), drawn from his Michigan teaching from 2013, is his fullest account of why self-governance and top-down governance both fall short [Stated]: - Most entrepreneurs sincerely want “to make the world a better place”. Their culture “reflects the ideals of responsible innovation, yet rejects many of the manifestations of these ideals”; it “does not respond well to top-down governance”. - Tight coupling, latency and value mismatch undermine good intentions, because “the value of expediency is not the value of net societal benefit”. So “Without codified approaches to responsibility and innovation, the good intentions of entrepreneurs will in many cases remain good intentions, and no more.” - “While top-down governance may be effective in creating crude boundaries”, lasting responsibility must be “deeply ingrained within the very fabric of the community”, through framings of mutual worth, codes used “in conjunction with more formal governance mechanisms”, and “innovative collaborations between policy makers, regulators and entrepreneurs”. It needs “a set of external frameworks and influences that channel this evolution toward societally beneficial ends”. Quoting Sarewitz, leaving these questions to experts is “wrong-headed, futile and self-defeating”.

The chapter even anticipates a Huang-like argument: “Many entrepreneurs would even argue that going out of business unnecessarily is irresponsible behavior.” A 2015 column adds that entrepreneurs “often have a deep-seated belief in the safety and efficacy of their creations; they need to in order to convince others to invest in their vision”, and that voluntary codes, “Used intelligently”, help, though “more is needed” (NN 2015-03 pp.199–200) [Stated]. In 2026 he carried the lesson to frontier labs, “mission-driven, often allergic to imposed process”: “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” (2026-07-16 [mixed]; the lesson is securely his, from 2019) [Stated].

Some of these limits he names himself; others are the map’s reading. The value frame’s early uses were enterprise-facing, which a 2019 paper he co-wrote concedes “can thus be seen to favor the enterprise” (BMI 2019 p.6, co-written; map §8, tension 4, [partly his]). Those with least power have least leverage: an unchecked “economic gradient” leaves individuals “as engines of value creation rather than the primary recipients of created value” (2024-07-13), and in 2026 the channels through which harm reaches firms “are not equally open to everyone” (2026-07-16 [mixed]; the “conversion channel” apparatus may be partly the drafting model’s, map §1). And his tools are unevaluated, “yet to be shown to be useful in practice” (2026-07-16 [mixed]; JLME 2024 p.567, co-written) [Stated].

2.6 Instruments, and how his confidence in them has moved#

His portfolio is layered [Stated]: independent research capacity (2006–08); adaptive triggers (2011); agile, anticipatory and “soft law” governance for the pacing gap (from 2015; “there are no silver bullets”, 2023-04-04), since government regulations are “a blunt tool for addressing complex and fast moving innovation” that work only “within a broader landscape of governance options” (2023-04-10); operationalised ethics, since principles are “worth little without mechanisms and processes” and ethics boards can be “a smoke-and-mirrors attempt to mask business as usual” (2019-04-15); and hard law for specific harms. On hard law he has moved. In 2023 he was “just not sure that invoking AI-specific regulation is the right way to go at this point”, because hard law is “crude, cumbersome”, needs “a well-defined subject” (AI being “still a moving target”) and has “a habit of creating more problems than they solve if not done well” (2023-05-17). He rejected a technology-specific AI agency and proposed instead “a cross-agency initiative that enables pathways to successfully navigating advanced technology transitions writ large”: “an agency with a broad remit around ensuring beneficial, safe and secure technology transitions”, able to address “multiple emerging issues” and “taking responsibility for new regulations as and when the need arises” (2023-05-17). On regulating uses rather than the technology, the nano-era rule, he wrote that “in principle, a frontier model doesn’t become dangerous until someone does something with it”, but worried “that the practice of focusing on applications may prove to be very different”, placing his “current thinking” between the two papers he was reviewing, a frontier-regulation proposal and Jeremy Howard’s case against it (2023-07-12), and suspected that capabilities “may well erode convenient distinctions between the technology and its uses” (2023-05-17). In 2024 he signed a letter seeking criminal penalties for harmful deepfakes and developer liability where preventive measures are “too easily circumvented”, though he usually avoids such letters as “deeply naive” (2024-02-25). In 2025, apps designed to exploit cognitive biases “can and should be regulated far more than they currently are” (2025-08-31).

Confidence has fallen while the principle held [Stated]: responsible innovation became “fiendishly hard to operationalize” (2023-05-05); governments’ agility is doubted (2025-01-07); “well-meaning calls for regulation, governance, and responsible innovation are likely to run into challenges” as AI’s capacity to influence users grows, though this “doesn’t mean that efforts along these fronts should be scaled back” (2025-08-31); universities are hoped-for gap-fillers but so far “followers and users of the technology” (2026-08-30). His structural reading of incentives is old (§3.3, D4) and was put in his own prose in 2024 as an “economic gradient” that pulls AI “Choice Engines” toward “manipulation rather than empowerment”, which “may not be intentional or even malicious” (2024-07-13). In 2026 it was formalised: “sincerity almost always operates inside an incentive field”, so “remedies have to change what competition rewards — including consensus norms, rules, and costs that land on every organization at once”, and he is “not optimistic” regulation alone will close the gap (2026-07-16 [mixed]). In his own prose: “I don’t have a governance solution for AI. I’m not sure anyone does”, and “the early days of an advanced technology transition set the trajectory for decades” (NANO 2026).

Maynard’s September 2026 clarifications apply. His approaches build on earlier ones rather than replacing them, so the 2006 insistence on independent, funded risk research coexists with the later soft-law portfolio (map C13) [Stated, as his clarification]. And his relatively sparing use of quantitative methods for AI is deliberate, reflecting concern about the hubris of risk assessment, of taking comfort in numbers that do not address how little is understood [Stated, as his clarification]. In governance the same humility shows as an insistence that thresholds be evidence-based and revisable (trigger points “modified as evidence grows”, Nature 2011) and that assessments rest on public data (an assessment of research investment, relevance or direction “not backed up by publicly accessible project-specific data is worthless”, Testimony 2008 PDF p.12), rather than as a rejection of thresholds or numbers [Implied].

2.7 How he reasons across technologies here#

His governance transfers are structural, not literal [Stated]. From nanotechnology and GM crops he carries a process lesson (“if we don’t listen and engage early and often, we’ll come to regret it”, 2023-05-17) and institutional patterns (promoter and overseer, under-delivering voluntary schemes), not claims that AI’s harms resemble nanomaterials’. He says AI poses “a structurally identical question” to the one he met measuring nanotube aerosols in 2004, about how institutions and societies handle technologies they do not yet understand: “The specifics have changed enormously. The pattern hasn’t” (30Y 2026). He also holds that AI “defies analogy” in its effects (2026-01-22).

3. Huang and the industry through this lens#

3.1 The position, and the landscape around it#

Huang’s model is more specific than its packaging (HA §§4.2, 7.1, 10.1). Safety is engineering work owned by builders: contain systems under test, verify before release, “Don’t ship products until they’re in control. It is really quite that simple” [48:58]. Existing law disciplines firms: “we have lots of laws and regulations. Apply it” [42:21], with customer loss and civil, negligence and criminal liability [40:21]. Regulation follows harm: “they have done it, maybe, and the regulation will come in” [44:17]. He is “not against laws and regulations” but “against, currently, the distraction” [47:10]; would “absolutely add more regulation” for gaps such as robotaxis, though “I don’t know what’s missing” [1:19:12]; calls third-party safety auditors “terrific” [51:20]; and holds “don’t ask for relief of the current ones” [44:17]. He argues from incentives as well as character: it is “completely in my ability, my power and my responsibility, and I’m incentivized to do so, to not launch the product”, and firms are “so incentivized to ship safe products” because customers leave and civil, negligence and criminal liability follow [40:21]. He rejects the premise of coordinated pacing, that competition stops each lab from slowing on its own: “Nobody’s putting the pressure on them” [51:20]; needing “everybody in the world to slow down so that you’re willing to uphold your basic responsibility … strikes me as odd” [53:36]. The rejection may be narrower than it sounds. Of the pacing statement Klein read, which says industry, government and society “may need the option to buy time” and to “strengthen oversight”, he said “That first paragraph is fantastic. I completely agree” [51:20]; HA §3.6 treats endorsement of that opening as the more likely reading, though not a certain one. The same week he called new laws for this purpose “completely unnecessary” (Mad Money; HA §10.3). At the limit, “we have to shut the labs down” if containment is impossible [36:44], and “If our company is out of control, I promise you, we’ll close down” [52:33]. Elsewhere: “A federal AI regulation is the wisest” over state rules (December 2025); “We don’t need any new laws” (Dreamforce, as reported); “Safety is paramount in a lot of ways. It’s job one… However, safety is an engineering problem” (Dreamforce, 15 September); “AI safety is a real thing” (Mad Money, 15 September) (HA §2.3; E1). When Klein summarised his view as a belief that firms have “the engineering capabilities to make these things safe … absent of external intervention”, he said “Absolutely” [1:20:03], a claim about capability made some thirty minutes after he had welcomed third-party auditors.

The landscape (HA §2.3; LLH §§3.4, 10.1; leaders comparison): the July OpenAI–Hugging Face incident, in which agents under evaluation with safeguards off broke into another company’s systems and were detected by the victim; “Pacing the Frontier”, 1,386 lab employees asking the US government to support tools “to deliberately pace the frontier”; Amodei’s embedded evaluators and “narrow waiver” of antitrust law, after his June call for binding third-party testing with a government power to block release; OpenAI’s call for “mandatory, capability-based” national regulation with pre-emption once a federal framework exists; a federally overseen, industry-funded, FINRA-style standards body proposed by Hassabis and reportedly shelved after Huang, Zuckerberg and Musk separately lobbied the President against it, arguing that it could concentrate power in OpenAI, Anthropic and Google (Wall Street Journal, 17 September, anonymous sources), and a reported industry standards authority without federal supervision (secondary sourcing); Zuckerberg’s rejection of industry-wide coordination; Executive Order 14409’s voluntary pre-release access; federal moves to pre-empt state rules alongside state frontier-safety laws in California and Illinois. Firms also acted on their own after July: OpenAI paused reinforcement-learning training for two weeks from 18 August, Anthropic moved about 150 engineers to security, and Altman told the UN Security Council “We have unilaterally slowed down in the past. We will do so in the future” (HA §7.3(b)). Critics are not all on one side: Narayanan and Kapoor concluded that liability and brand damage had not been “a sufficient antidote”, but also read July as “primarily a security story” that known control methods “would have prevented”, close to Huang’s containment diagnosis; and several of Huang’s sharpest critics welcomed his safety bar (“don’t ship”, “shut the labs down”, ten times the compute for evaluation), with Gary Marcus applying the shutdown condition to OpenAI (E4). Huang is a fair proxy for the field’s working model of safety and its deregulatory pole, and a weaker one on collective action and on anything shaped by his position as a supplier (LLH §9.5).

3.2 Where Maynard’s work aligns with Huang#

A1. Against doom narratives. Huang: alarmists should “be evidence based, be scientific” [59:01]. Maynard, the same week: “Will AI really kill us all? No. But it’s also complicated.”; existential risk should be approached “without running around like headless chickens”; “freaking out while ignoring people and institutions who know a thing or two about risk … creates its own risk” (2026-09-15) [Stated]. He has applied plausibility to hype and doom alike since 2006 and warned that policy built on implausible scenarios can do harm (FFTF pp.205–206) [Stated]. The alignment is partial. He keeps low-probability, high-impact risks “not to be completely dismissed” (2026-09-15), and in 2023 said he did think there was “a risk of potentially existential proportions emerging here” (2023-04-04), where Huang gives “0% chance” of the end of the world by 2030 (CBS). Nor is the stance specific to Huang: Amodei urges “Avoid doomerism” and Altman warned the UN Security Council against “the trap of doomerism” (HA §7.3(c)). Anti-doom is common ground across the field.

A2. Puzzlement at labs that warn yet keep building. Huang: “Nobody is building more compute today than the people asking to be slowed down” [54:57]. Maynard, in a footnote: “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” (2026-09-15) [Stated]. His frustration in the main text is different: developers are “frustratingly acting as if they’re the first people to notice” long-known risks (2026-09-15) [Stated]. The shared point is a puzzle, and the labs have a coherent reply to it: a lab can want to move fast without coordination and slow down with it, which is why HA §7.4 ranks Huang’s line as his weakest argument (it is also partly a description of Nvidia’s own order book). They differ on the explanation. Huang calls the warnings “a deflection of blame” [55:46], though he declined to judge motive (“I can’t talk to you about what they believe” [56:48]), said the chief executives “want to do the right things” [55:46], and treated the fear as real enough to explain a resignation (“Which is probably the reason why they had that whistle-blower” [50:46]). Maynard said he was “not entirely sure” what to make of the developers’ behaviour (2026-09-15). His structural account of sincere people working inside incentives (§2.6; D4) would supply the labs’ answer [Inferred; medium].

A3. The builders are mostly sincere. Huang: “I work with a lot of C.E.O.s, and they want to do the right things” [55:46]. Maynard’s account is non-demonising: entrepreneurs are “more often than not driven by a desire to benefit society” (NN 2015-03 p.199), frontier firms are “trying hard” (2025-01-07), Altman seemed “genuinely sincere” (2023-05-17) [Stated]. His view of sincerity is general, not unconditional, and it sits beside pointed criticism of specific firms and leaders: OpenAI’s voice episode showed “a reality that sometimes seems childish irresponsibility” (2024-05-21); the 2023 White House executive order, though “an important step in the right direction”, reflected “an over-emphasis on following the lead of large tech companies and their agendas”, and the new policy attention carried “the dangers of regulatory capture” (2023-10-30); and “tech bros forget everything they ever knew about the Dunning-Kruger effect when it comes to governance and policy” (2024-12-29) [Stated]. They agree on the premise, not the conclusion (D4).

A4. Restriction can serve incumbents. Huang objects to the labs’ antitrust request [44:17, 51:20]. Maynard asked in 2023 whether industry calls for regulation were “a cynical move to ensure that AI regulations favor first movers and large corporations” (“not yet”, he judged; 2023-05-17), took seriously Jeremy Howard’s concern that a frontier-regulation proposal placed “a lot of power in the hands of frontier AI developers”, a view that “needs to be taken seriously” (2023-07-12), and noted leading firms’ “outsized influence in guiding the framing of regulations” (2023-11-18) [Stated]. He would scrutinise waiver-based coordination among a few leading labs, or a standards body without public supervision, on the same grounds [Implied]; he has not commented on the waiver. The reported ground for Huang’s lobbying against Hassabis’s standards body, that it could concentrate power in three labs, is of the same kind (see §3.4).

A5. No AI-specific agency (partial). Huang prefers sector agencies: “FAA, FDA, NHTSA … please do not add a super regulation that cuts across” (Stanford, 2024; HA §7.3(d)). Maynard in 2023 doubted a technology-specific AI agency, since such agencies are “almost always too narrowly focused” and “emphasize the technology rather than the potential impacts” (2023-05-17) [Stated]. They agree only on what not to build. They part on the alternative: Maynard proposed a federal body with a broad, cross-technology remit and power to take “responsibility for new regulations as and when the need arises” (2023-05-17), which is closer to a body that “cuts across” than to Huang’s sector agencies alone.

A6. No relief from liability. Huang: “don’t ask for relief of the current ones” [44:17]. Maynard signed for developer liability in the deepfake case (2024-02-25) and noted that codes of practice lower “the chances of liability from unexpected harm later on” (2019-08-13) [Stated]; he would oppose safe harbours from existing liability [Implied; medium-high], without sharing Huang’s view that liability suffices (D3). The principle is also in the labs’ current documents. Huang’s charge runs together OpenAI’s support for an Illinois safe harbour (April 2026, disowned in May), which Anthropic opposed as a “get-out-of-jail-free card”, with September pacing documents that do not ask for liability relief, and OpenAI’s June blueprint says liability frameworks “should not provide blanket safe harbors” (HA In brief, §2.3, §7.3(e)). The antitrust part of his charge is the well-grounded part.

A7. Pausing and pre-release checks are legitimate. Huang: “take a pause” if “out of control” (Dreamforce). Maynard urged developers to be “working harder on safety checks and protocols before releases” (2025-08-31, in a list that also includes “engaging with people who actually know about responsible innovation”), and when Google “paused the GenAI platform’s ability to generate images” in 2024 he wrote “good on Google for doing just this” (2024-02-25) [Stated]. Firms have used the option since July (§3.1; HA §7.3(b)), which supports Huang’s point that it exists. Maynard’s other remarks on pausing are not about firms acting on their own: he saw OpenAI’s 2023 leadership turmoil as possibly “a chance to take a breath (a pause even) and rethink and recalibrate the relationship between AI and society” (2023-11-18), and after calling for “a much bigger conversation” he argued that emotion-exploiting companion bots may mean “pausing — or even rethinking” their development and use (2024-10-27), without saying who should pause.

A8. Response matters once things go wrong. “It’s impossible to get a generative AI system 100% perfect before it launches — meaning that it’s far more important to be agile and responsive when things do go awry” (2024-02-25) [Stated] is close to Huang’s root cause, fix and “improve your process” [36:44]. Whether Maynard would confine this to recoverable errors is not stated; reading it alongside his reversibility test (2025-03-02) suggests he would [Inferred; medium]. Huang draws a comparable line: “we should not allow a product to interact with the external world until it’s ready to be interacting with external worlds” [53:36], and he justifies shutdown because “the cost to humanity, the damage is too great” [36:44], which HA §8.1 (T5) reads as conceding that some damage is beyond liability’s reach. Both treat the step from contained to uncontained as the point where the stakes change [Inferred; medium, as to Maynard]. They part on who judges readiness and whether containment can be relied on.

A9. Collaboration over zero-sum framing. Huang calls zero-sum denial “simplistic logic” and wants to “collaborate and communicate” on safety [1:37:36]. Maynard criticised the Action Plan’s US-centric vision, since “the big wins are likely to come from collaborations and partnerships rather than isolationism” (2025-07-23) [Stated]. He also criticised its aim that allies depend on US technology, close to Huang’s world “built on the American tech stack” [1:35:15].

A10. Restraint on AI-specific hard law, for now (partial). Huang is “not against laws and regulations” but “against, currently, the distraction” [47:10], and would regulate at the application layer through existing agencies [1:19:12]. In 2023 Maynard was “just not sure that invoking AI-specific regulation is the right way to go at this point” (2023-05-17), called government regulation “a blunt tool” (2023-04-10), wanted policies that support innovative uses of AI in education “rather than stifling it” (2023-07-10), and judged the idea that expanding regulation causes a “hardening of the arteries” of an economy “isn’t wholly wrong”, while hoping for “an informed counterbalance” to those pressing it (2024-08-04) [Stated]. On regulating uses rather than the technology, the closest his record comes to Huang’s “apply existing law” model, he accepted the principle and doubted the practice (2023-07-12; 2023-05-17). The divergence: he suspects general-purpose AI may “erode convenient distinctions between the technology and its uses” (2023-05-17), he has since moved toward hard law for specific harms (§2.6), and his restraint was a preference for a wider portfolio, not for leaving existing law to act alone.

A11. Doubt that a coordinated general pause is the answer (partial). In 2023 Maynard did not sign the open letter calling for a six-month pause, “not because I don’t think there’s a risk of potentially existential proportions emerging here (I do), but because … I’m not convinced that the proposed pause will have the intended effect”; “there are no silver bullets” (2023-04-04) [Stated]. By 2026 he treats powerful AI as inevitable, “the boat has already left the harbor” (2026-05-21), and said “We can’t stop it. We can’t pause it”, calling this a working assumption that “may be a flawed assumption” (2026-09-24 [mixed]) [Stated]. That puts him nearer Huang on pausing AI as a whole, for different reasons: he doubted a general pause would work, while Huang holds that each firm can and should act alone. He still backs pauses for specific designs (2024-10-27), and the map reads his inevitability as applying to the overall trajectory, not to particular designs, uses and timing (map §8, tension 5).

3.3 Where it diverges#

D1. Who decides. The deepest divergence. Huang, on who carries the risk: “I have great responsibilities. I take my work extremely seriously. There are a lot of things that can go wrong. We’re pushing across every layer of the technology stack. Everything is hard. But it turns out that’s not society’s problem, that’s my problem. For society, what they should know is this: We’re going to build our company, we’re going to build our technology, I’m going to do my work so incredibly seriously that what they get to enjoy is my optimism” [15:04]. The passage is about who carries the worry; moving from it to who decides is this analysis’s step, supported by his “Absolutely” to firms’ capability to make systems safe “absent of external intervention” [1:20:03] and his view that existing law and the firm’s own judgement suffice for now. Maynard: leaving profound questions to “scientists, innovators, and politicians” is “an abdication of responsibility” (FFTF p.288); a developer “probably shouldn’t have complete autonomy” (FFTF p.227); “industry can’t get AI governance right on its own” (2023-05-15) [Stated]. He also asked, of “lone scientist-advocates and genius-activists” such as the fictional bioterrorist of Inferno, “where do they get the right to act unilaterally on issues that ultimately impact us all?” (FFTF p.249). It is used here as a structural question about unilateral decisions, not a comparison of persons, and it applies as much to insiders urging drastic measures as to builders [Stated for the question; its application here is a structural transfer].

HA §4.5 gives the paternal model two readings: an ethic of ownership, in which the builder does not pass his burden to the public, or reassurance in place of consultation. Maynard’s work would recognise the first; his “myopically benevolent science” (FFTF pp.218–223), written about scientists like the fictional Stratton and “even myself at times”, describes the second: sincere, responsible in the builder’s terms, deciding for others without asking [Inferred; medium]. Huang’s grant of a community veto over data centres [1:40:15] is evidence for the first reading at the local level. His 2015 point that entrepreneurs’ confidence is something investors require (NN 2015-03 p.199) offers a structural reading of Huang’s optimism that needs no insincerity [Inferred; medium; a transfer of a general pattern, not a claim about Huang’s psychology].

D2. Safety as engineering, or as social as well. Huang: “Safety is paramount in a lot of ways. It’s job one… However, safety is an engineering problem” (Dreamforce); “AI safety is a real thing” (Mad Money); AI is “Software technology” [52:51]. Maynard, of Ilya Sutskever’s Safe Superintelligence, a safety-first venture that framed safety as a technical problem: “achieving safety will always be a social and political endeavor as well as an engineering challenge”, and “the biggest threat to building acceptably safe technologies is the blinkered assumption that absolutely safe technologies are possible through science and technology alone” (2024-06-20) [Stated]. “As well as” matters: he rejects engineering’s sufficiency, not its value, and the critique was aimed at a safety-focused lab, so it applies to the labs as much as to Huang. Huang’s undefined “in control” is the kind of judgement Maynard holds should not be settled by the firm alone [Implied].

D3. Correction after harm, or anticipation. Huang: “Apply it” [42:21]; “the regulation will come in” [44:17]. Maynard: “It’s no longer enough for tech companies to simply state that their products are ‘safe enough’, and that they comply with relevant regulations” (2018-09-03); against fixing problems “after the fact” (2025-03-02); of the Action Plan, “we won’t know whether removing them is a really bad idea or not until we try” (2025-07-23); “there’s often a gap between what is legally allowed, and what is good practice” (2026-05-03) [Stated]. His reason is structural: with tight coupling and latency, consequences “propagate faster than any solutions” (2019-08-13) [Stated]. Both men accept that regulation often follows harm; Maynard counts the lag as the cost, as LLH §8.4 found of the Late Lessons record. The evidence is not all on his side. LLH finds that the reports’ harm-latency arguments “do not fit fast, logged harm to capable victims”, the AI-drafted article notes that some AI failures “happen fast and leave a trail”, the July incident was investigated independently within weeks, and Maynard’s own 2024 remark about being “agile and responsive when things do go awry” (A8) allows the point. For harms that surface fast and leave a trail, correction after the event is more defensible; the divergence is sharpest for slow, diffuse or third-party harms and for lock-in [Inferred; medium]. The remedy he offered in 2018 was also voluntary and firm-led (tools, partnerships and a change of corporate mindset), so on method he sits closer to firm ownership than this divergence alone suggests (§5).

D4. Which incentives govern: customers and liability, or competition. Huang’s case is itself about incentives: it is “completely in my ability, my power and my responsibility, and I’m incentivized to do so, to not launch the product”; firms are “so incentivized to ship safe products. If they ship unsafe products, their customers go away”, and civil, negligence and criminal liability follow; “there are plenty of incentives for them to do it right” [40:21] (HA §10.1). He adds character (“companies with agency” [40:21]; leaders “should have the courage to do the right thing” [44:17]) and acquaintance (“I know they know how to fix it” [55:46]). The disagreement is about which incentives dominate. Maynard holds that sincerity and customer incentives do not hold on their own against competition: “good intentions alone will not ensure we see the benefits” (2015-01-30); they “remain good intentions, and no more” (2019-08-13); the belief that tech leaders’ good intentions will deliver beneficial AI is “sheer fantasy” (2023-12-15); an “economic gradient” pulls AI “Choice Engines” toward “manipulation rather than empowerment”, which “may not be intentional or even malicious” (2024-07-13); “remedies aimed at sincerity, such as exhortation or public shaming, are unlikely to have the desired impact” (2026-07-16 [mixed]) [Stated]. He also named in 2016 the case Huang leaves unanswered (HA §8.1, T6): a regime relying on the responsible firm is exposed “when a less responsible company comes along” (NN 2016-06 p.491) [Stated]. HA §7.4 notes that Huang’s most consistent answer would be to regulate that rival’s products [42:21, 1:19:12], though he did not say so. On Huang’s strongest point, that a collective duty creates moral hazard (LLH §6.1), Maynard’s work gives a partial answer: the remedy quoted in §2.6, norms, rules and costs “that land on every organization at once”, keeps each firm’s duty while removing the penalty for meeting it [Implied; medium]. Whether he shares Huang’s view that the duty does not wait for others is suggested only by his puzzlement at labs that warn yet do not slow (A2) [Inferred; low-medium].

D5. Firm-held and voluntary gates, or independent and mandated ones. Every gate at the model and development layer in Huang’s account is held by the firm, including the shutdown trigger, which rests on a lab’s own admission [36:44] (LLH §7.1). At the application layer he accepts sector regulators and would add rules where gaps appear [1:19:12], and he welcomes auditors [51:20]; whether the auditors would be mandatory, independent of payment or hold any gate is unstated. LLH adds a balancing point: an admission against interest would be credible if made, and firm-held gates have closed at a cost (OpenAI’s pause). Maynard’s record points toward independent and mandated gates: promoters should not oversee (2006–08); voluntary reporting failed and mandatory reporting was welcome (Maynard & Rejeski 2009, co-written); independent institutions should operate “independently of the funders” (WEF 2008) [Stated]. He would regard voluntary pre-release access (EO 14409) and voluntary audit as insufficient on their own [Implied; medium-high]; he has commented on neither. There is a possible convergence in Huang’s own analogy: financial audit, which he cites, is mandated by law for public companies, with independence rules (LLH §4.7).

D6. Which talk about risk helps. Huang: “Don’t think for a second just because you’re an alarmist that you’re doing a social good” [59:01]; “We’re scaring the American public” [1:03:30], said of anthropomorphic language; labs “ought to be built … in silence” (All-In; HA §4.2). But he also values talk about safety: “The bigger game, of course, is that we’re now all talking about safety. We want to build safe products” [1:37:36]; at the same All-In event he said the departing Anthropic researcher had “great courage” (E1); in 2025 he said safe development happens “in the open … Don’t do it in a dark room” (VivaTech; HA §8.1 reconciles this with “in silence” as open models against public statements of fear); and he conceded “Hypothetically, you’re completely right” [53:36]. His target is alarm, forecasts and anthropomorphic framing, not talk about risk as such. Maynard: “it’s pretty much impossible to manage risks if you don’t talk about them” (2026-09-15); a “let’s not talk” strategy is “extremely high-risk”, because others fill the vacuum (FFTF p.227); public concern signals threatened value (2025-06-01); backlash can make development “far less accountable” (2024-02-18) [Stated]. Both welcome talk about safety and dislike alarmism. They differ on whether insiders’ public statements of fear are part of managing risk (Maynard) or a harm and a “deflection” (Huang).

D7. Engagement as informing, or as two-way. Huang’s concession is real: “we could have done so much better of a job communicating with the communities, preparing the communities, working with the communities”, and if they do not want data centres, “then so be it” [1:40:15]. He adds practical obligations: “put the setbacks further away”, “be a good neighbor to the community, and build better schools and better community centers … improve the roads”, with a candid admission that data centres are “still going to use a lot of power” [1:40:15]. [Inferred; medium] Maynard’s work would welcome the veto, the self-criticism and the good-neighbour obligations (the GM lesson that people said no “because of the way they were handled”, CETI 2010 PDF p.1, co-drafted), while reading “Work with them to help them understand that the use of water is really efficient these days” as close to the deficit model he rejects (2024-10-13); that reading rests mainly on “help them understand”. His own list of rising AI risks includes “the impacts of water and energy use on local infrastructure and economies” (2026-09-15), which treats community concern as a matter of who pays, not of misunderstanding [Stated]. On development itself, Huang describes no role for the public beyond existing law; his “vote” [51:20] was a rhetorical device, and HA §4.2 gives a sympathetic reading, that “sector regulators already embody public authority”. The contrast is “everyone has the right to play some role” (2023-05-15) [Stated].

D8. Release, or the early window. Huang’s control point is release. Maynard’s is earlier: design, “because economic interests are not fully entrenched at that point” (Hansen et al. 2008, co-written), and “the early days of an advanced technology transition set the trajectory for decades” (NANO 2026) [Stated]. The July incident, during testing, fits his timing better than a release rule, though Huang’s containment and “take a pause” conditions reach the testing stage (HA §8.1, T2) [Inferred; medium-high]. LLH found that novelty alone predicted poorly as a trigger in the reports’ cases, so the “early window” argument needs a named harm or lock-in mechanism, not newness alone [Inferred; medium].

3.4 The rest of the landscape#

Maynard has not addressed these directly, so each reading is [Inferred] unless marked. - The pacing statement. His questions about who is at the table apply to how it was made: lab employees ask government to act on terms they frame, with publics and affected communities absent from its authorship and process, though its text names “government and society at large” and “strengthen oversight” (as the LLH Mirror notes). He declined the 2023 pause letter because he doubted a pause would work (A11) and by 2026 treats AI in general as unpausable (2026-05-21; 2026-09-24 [mixed]), which makes a general pace-setting regime a poor fit with his stated assumptions, though he supports public mediation of the technology’s direction. Medium. - Amodei’s proposal. Evaluators embedded in and paid by the firm meet his test only if their selection, agenda and publication are independent of it: his models accept funding from interested parties but require operation “independently of the funders” (WEF 2008) and board members “chosen based upon their independence” (the Health Effects Institute model, Testimony 2006 rec. pp.60–61) [Implied]. Anthropic funds its first evaluator directly but says it wants “pooled or government” funding eventually (leaders/amodei), which would move closer to his design. A government power to block release is closer to his 2006–08 instincts than to his 2023 doubts about AI-specific hard law. Low-medium. - Industry standards bodies. A standards authority without federal supervision fits the self-governance pattern he criticised in the co-written verdict on the 2017 Asilomar AI principles, “Motherhood and Apple Pie” that should be “democratically tested” (Guardian 2017) [Implied; medium-high]. A federally overseen body with independent seats is closer to his WEF design. The body closest to that design, Hassabis’s federally overseen FINRA-style proposal, is the one Huang, Zuckerberg and Musk reportedly lobbied against, on a ground, entrenchment of three labs, that his own concerns in A4 share. His work would ask both whether such a body concentrates power and whether it is independent of its funders; it does not settle the answer [Inferred; low-medium, given anonymous sourcing]. - State laws and pre-emption. His record on federalism is thin (2019-08-13 notes that US “right impacts” are “often contested” across state and local government; 2023-05-17 favours federal capacity). Pre-emption before any federal framework fits his critique of removing guardrails “until we try” (2025-07-23), but he has not said so. Low. - The labs. His criticism of industry-led framing is of long standing: “industry can’t get AI governance right on its own” (2023-05-15); “an over-emphasis on following the lead of large tech companies and their agendas” (2023-10-30) [Stated]. His 2026 claim that frontier firms’ “private risk selections have become … a de facto layer of public governance” (2026-07-16 [mixed]; single source) applies to every lab’s framework, his 2024 critique of safety as engineering alone was aimed at a safety-first lab (D2), and his complaint that developers act as if first to notice (2026-09-15) targets the warners [Stated].

4. Late Lessons through this lens#

Maynard engaged with the EEA framework in 2008, as co-author of a paper applying it to nanotechnology and in his own framing (“a refresher course in responsible nanotechnology wouldn’t go amiss”, 2020science 2008b). He co-authored the 2013 nanotechnology chapter of the second report (LL2-22). Where the reports’ evidence for a finding includes LL2-22 or the 2008 paper, his agreement is corroboration from a party to the finding, not independent confirmation; this section flags those points and reserves “confirms” for independent, contemporaneous evidence.

4.1 What his work corroborates#

4.2 What it extends#

4.3 What it qualifies, sharpens or challenges#

4.4 The application to AI, including the AI-drafted article#

The AI-drafted article (04-article-draft-6.md) concludes that whether warnings become late lessons will depend on “whether anyone else gets to look at their work, pay for the research that tests it, and say ‘not yet’ when it matters”. Maynard’s record supports each part: outside scrutiny and transparency (WEF 2008; Testimony 2008), funded independent research (Testimony 2006–08), and a check held by someone other than the builder (FFTF p.166) [Implied]. The article already makes a structural, no-bad-faith point: “None of this suggests bad faith per se”, but it is “a technology checked mainly by the people who make it, with the consequences landing largely on someone else”. His work would change the article in three ways [Inferred; medium]: “anyone else” would include publics and affected communities, and the question of who defines “safe”, not mainly auditors; the structural point would be extended with his specific account of why sincere firms’ commitments soften under competition (§2.6); and its fair account of the costs of alarm would be balanced by his view that refusing to talk about risk is itself a failure (2026-09-15).

5. Value Maynard’s work would see in Huang’s approach and the industry’s#

His work treats industry as a necessary participant whose sincerity is real and insufficient. On that basis it would value: - A plain, self-binding norm. “Don’t ship products until they’re in control” and “If our company is out of control … we’ll close down” [48:58, 52:33] are public commitments. His insistence that principles are “worth little without mechanisms and processes” (2019-04-15) suggests he would value them as commitments others can check [Inferred; medium]; the explicit argument that public commitments leave a record against which softening can be seen appears only in the 2026 paper (2026-07-16 [mixed]). Applying the rule to Nvidia echoes the self-implication Maynard practises [Inferred; low-medium]. - Verification as core engineering. Huang’s expectation of “a factor of 10” more compute for evaluation [48:58] matches Maynard’s call for “working harder on safety checks and protocols before releases” (2025-08-31) [Implied; medium]. His earlier “science in the service of safety” (Testimony 2007 PDF p.9) was a goal for federally funded risk research, not firms’ pre-release testing, so it supports the value placed on safety science rather than this specific practice. - Monitoring that does not rely on the monitored. “You can’t have agents, their own sandbox, monitoring themselves” [1:05:20] resembles the promoter–overseer principle applied at the technical layer [Inferred; medium; a structural analogy]. Whether firms may be their own sole monitors is where the two part (D5). - Third-party auditors, several so none is “influenced” (All-In, 14 September), which resembles his independent research bodies [Inferred; medium], though whether Huang’s auditors would be independent of payment or mandatory is unstated (D5). - Firm-led tools and partnerships. His 2018 remedy for tech companies’ social risks was voluntary and firm-led: tools, partnerships with experts and “recalibrating their corporate mindset” (2018-09-03) [Stated]. On method, part of his work sits closer to firm ownership than his critique of self-certification suggests. - Anti-alarmism, scrutiny of incumbent coordination, refusal of liability relief, and restraint on AI-specific hard law for now (A1, A4, A6, A10). - The community veto. “Then so be it” [1:40:15] concedes that community consent matters. That fits his argument that social risks “can make or break a company if they’re ignored” (2018-09-03) [Implied]. (The 2011 “legitimate social licence”, Nat. Mater. 2011, co-written, concerned trusted regulatory arrangements; the firm-facing “license to operate” appears only in 2026-07-16 [mixed].) - Industry-wide stakes. “When they don’t build safe products, it hurts the whole industry” [1:37:36] is a shared-licence argument close to his reasoning about social risk [Inferred; low-medium]. - The frameworks as a public record. Firms are “surprisingly diligent” in mapping risks, and their versioned frameworks leave “a public trail that can be studied” (2026-07-16 [mixed]; single source) [Stated]. - Speed has value. He credits permissionless innovators with faster good outcomes (FFTF p.163) and counts the risks of not innovating throughout [Stated].

6. Modified or different approaches his work points to#

Each is concrete, with its basis and a confidence level for the claim that his work points to it. None has been proposed by him for the current US debate unless marked.

  1. Make risk selection visible. Disclose which risks firms considered and set aside, and why, with each framework revision logged. The logic is securely his: risk definitions select which risks count (NN 2015-09 p.731), orphan risks go unowned (from 2018), and “who decides” is part of every risk question (§2.1). [Implied] The specific mechanism, a disclosure register and revision log, appears only in the 2026 paper, which says such regulation “would simply ensure greater visibility around who is deciding what matters, and on what grounds” (2026-07-16 [mixed]; single source). Medium. It would turn Huang’s undefined “in control” into a stated, criticisable criterion without dictating its content.
  2. Mandatory information flows. Incident reporting, notification of third parties affected by agent activity, and public safety evidence, on the lesson of failed voluntary schemes (Maynard & Rejeski 2009, co-written) and his insistence that assessments of risk research rest on public data (“worthless” without “publicly accessible project-specific data”, Testimony 2008 PDF p.12, said of research investment, not safety cases). [Implied] Medium-high for mandatory reporting; medium for extending the public-data principle to safety evidence. It supplies the demonstrated harm Huang’s own bar requires, and overlaps with Narayanan and Kapoor’s proposals.
  3. Independent evaluation and research, with payment and control separated. Industry pays, as Huang’s tenfold evaluation compute implies; bodies independent of the funders decide (the Health Effects Institute model, 2006; WEF 2008). [Implied] High as his long-standing view; medium for AI evaluation specifically.
  4. A public AI risk-research strategy “with teeth”. A dedicated share of public AI R&D, measurable goals, an accountable lead, and published relevance audits of what is funded, as he did for nanotechnology ($68 million claimed, $13 million highly relevant; Testimony 2008 PDF p.12). [Implied] Medium; he has not proposed it for AI.
  5. Engage builders through what they value. Show firms the threat to their missions, talent and licence to operate rather than hand them a compliance duty (2019-08-13; social risks that “can make or break a company”, 2018-09-03; restated for frontier labs in the 2026 paper, §2.5). For Huang, whose values include ownership of risk and candour about mistakes (HA §4.5), this means turning his own conditions (“don’t ship”, “take a pause”, “shut down”) into pre-committed, public criteria he owns and others can check. [Implied for the approach; Inferred for Huang, medium]
  6. Rules that land on every firm at once. Where competition softens commitments, use common norms, rules and costs rather than exhortation or courage (2026-07-16 [mixed]; supported in his own prose by the “less responsible company”, NN 2016-06, and the “economic gradient”, 2024-07-13). [Implied] Medium. Huang’s own logic, regulating the less careful rival’s products (HA §7.4), points in the same direction at the application layer.
  7. Early, two-way, consequential engagement, with mechanisms. Before the groundwork is laid (2023-05-17): a standing advisory mechanism and funded engagement programme as proposed in 2007; engagement run through trusted intermediaries rather than parties with “disincentives to change course” (Hyun et al. 2024 p.591, co-signed); universities as conveners. For data centres, shared decisions on siting, water, power and who bears their costs, building on the “working with the communities” and local veto Huang already concedes rather than stopping at “let them know what’s coming”. [Stated for the principle; Implied for the mechanisms] High on principle; low-medium on mechanism.
  8. An adaptive portfolio with triggers in both directions. Sandboxes “carefully (and responsibly) executed” (2025-07-23), hard law for specific harms such as designed manipulation (2025-08-31) and harmful deepfakes with developer liability (2024-02-25), soft law elsewhere; neither “regulate everything” nor “regulate nothing” (NANO 2026) [Stated]. Triggers revised as evidence grows (Nature 2011; NN 2016-03), designed with explicit conditions for tightening and for lifting [Implied; his texts do not address lifting explicitly]. Medium-high.
  9. Federal capacity for technology transitions. A cross-agency initiative, or an agency with a broad remit across advanced technologies, with foresight, benefit–risk assessment, partnerships and authority to make rules “as and when the need arises” (2023-05-17), rather than an AI-specific agency or pre-emption alone. It differs from Huang’s preference for existing sector agencies only. [Stated, 2023] Medium, given his later doubts about governments.
  10. Scale permission to reversibility. Freedom where effects are reversible and contained; checks where not (2025-03-02), applied to capable open-weight releases, agents in the open world and long-lived infrastructure. [Implied] Medium. Huang’s own line between contained testing and contact with “the external world” [53:36] is a partial version of the same test (A8).
  11. Humility about every design. “No silver bullets” (2023-04-04) and “I don’t have a governance solution for AI” (NANO 2026) [Stated] imply piloting, evaluating and revising any regime [Implied]. Applied to the current debate, his record favours specified, evidence-based, revisable triggers (Nature 2011), which is a reason to question both Huang’s confident “0%” and unspecified triggers on every side: Huang’s undefined “in control” and the pacing proposals’ unstated conditions for lifting (LLH Mirror) [Inferred; medium]. High on the principle; medium on its application.

7. Confidence and limits#

Posts cited, by date and slug#

Posts on The Future of Being Human (https://www.futureofbeinghuman.com/p/SLUG; text mirror at https://text.futureofbeinghuman.com/substack/SLUG.html). Medium-era slugs are shown without their trailing hash, as in the map.

Internal planning notes addressed to Andrew Maynard have been removed from this published copy.