AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › People & Organizations

Jack Clark

high confidence · updated 2026-08-17

Anthropic co-founder and head of policy; author of the Import AI newsletter; former OpenAI Policy Director; one of the most influential industry voices in AI governance.

Contents

Jack Clark is a British writer and policy strategist who co-founded Anthropic and serves as its head of policy. He was OpenAI's first Policy Director (2016–2021) before co-founding Anthropic in 2021, and he writes the weekly Import AI newsletter, which has run continuously since 2016 and is widely read across AI policy circles. He is also co-chair of the Stanford AI Index Steering Committee and a frequent congressional testimony witness.

Type: Individual (industry policy lead, writer) Affiliations: Anthropic (co-founder; head of policy); formerly OpenAI (Policy Director), Bloomberg (technology journalist) Notable recognition: Co-chair, Stanford AI Index Steering Committee (2017–); frequent congressional testimony witness

Background

Clark pivoted from technology journalism — at Bloomberg and The Register — into AI policy. He served as OpenAI's first Policy Director from 2016, where he helped build the company's nascent AI-policy function, including early work on release norms and dual-use risks. In 2021 he co-founded Anthropic with Dario and Daniela Amodei, Chris Olah, and other former OpenAI colleagues. In 2017 he co-founded the Stanford AI Index, an annual benchmark of AI progress and adoption that has published reports since 2017.

Roles

Clark is co-founder and head of policy at Anthropic, where he is the principal architect of the company's policy engagement — including on SB 53, the RAISE Act, export controls, the UK and US AISI frameworks, and the Frontier Model Forum. He drives Anthropic's external engagement with governments, standards bodies, and the AISI network, and appears frequently as a congressional testimony witness.

In 2026 Clark also began heading the Anthropic Institute, the company's research division on the societal consequences of advanced AI, disclosed via an April 30 Cosmos Institute announcement.

Positions and statements

Clark's policy stance aligns with Amodei's "surgical intervention" position, favouring transparency-based legislation first and escalating with evidence. Under his policy leadership, Anthropic backed California SB 53 and the RAISE Act, and he supports strong chip export controls as a geopolitical instrument of AI policy. He has publicly argued that safety research and governance engagement are competitively advantageous for frontier labs rather than merely costs. Unlike most industry policy leads, Clark writes publicly and at length, which makes Anthropic's policy positioning relatively legible.

Regulatory markets

Clark is co-author, with Gillian K. Hadfield, of "Regulatory Markets: The Future of AI Governance," first posted to arXiv in April 2023 and published in Jurimetrics vol. 65 in 2026, with a shorter precursor released in December 2019 (Regulatory Markets: The Future of AI Governance). The paper proposes that governments license private regulators against outcomes the government sets and require AI developers and deployers to purchase their services, as a response to what the authors call the technical deficit of command-and-control regulation and the democratic deficit of industry standard-setting. Its worked example is a licensing regime for independent red-teaming and evaluation companies for frontier models — a domain in which the paper notes developers have already moved from internal teams toward retaining METR, Apollo Research and consulting firms including Gryphon Scientific, which Anthropic engaged in 2023 for bioweapon-acceleration analysis (Regulatory Markets: The Future of AI Governance). The proposal predates and is distinct from Anthropic's own policy positions; see Independent Verification Organizations (IVOs) for how it relates to the US instruments that have since adopted the mechanism.

Oxford Cosmos Lecture

Clark was scheduled to deliver the 2026 Cosmos Lecture at the University of Oxford on May 20, 2026, titled "Change is inevitable. Autonomy is not.", addressing how individuals maintain self-directed lives as AI integrates deeper into society. The lecture is part of Cosmos Institute's annual series (Source: blog.cosmos-institute.org). In remarks from the lecture reported on May 21, 2026, Clark said there is a "non-zero chance" of AI killing everyone and argued that AI systems should be designed to prompt users toward real human contact rather than maximize parasocial engagement, connecting his catastrophic-risk framing to the human-flourishing and autonomy theme of the lecture's title. The design-toward-human-contact argument aligns with the anti-dependency design positions discussed under AI mental health (Source: fortune.com).

Import AI 455 — "Automating AI Research" (May 4, 2026)

In "Import AI 455: Automating AI Research," published May 4, 2026, Clark argued for a 60%+ probability that an AI system powerful enough to autonomously train its own successor will exist by the end of 2028, and a 30% probability by end-2027. He built the case from public benchmark trajectories:

  • SWE-Bench: ~2% (Claude 2, late 2023) → 93.9% (Claude Mythos Preview)
  • METR 50%-reliable time horizon: ~30 seconds (GPT-3.5, 2022) → ~6 hours (GPT 5.2 High, 2025) → ~12 hours (Opus 4.6, 2026)
  • Anthropic's internal CPU-LLM-training-optimization task: 2.9× speedup (Opus 4, May 2025) → 52× (Mythos Preview, April 2026)
  • CORE-Bench reproducibility: 21.5% (GPT-4o + CORE-Agent, Sep 2024) → 95.5% (Opus 4.5, Dec 2025)
  • PostTrainBench: Opus 4.6 and GPT 5.4 reach 25–28% of an LLM-fine-tuning task vs. 51% human baseline.

Clark anchored the argument to first-party Anthropic capability data, distinguishing it from Aschenbrenner's external-forecaster framing. The piece prompted Azeem Azhar's Exponential View #573 (May 9, 2026), which reframed the timeline into testable lab behavior, listing three patterns the labs' revealed preferences should already show — research-multiplier hiring, capex over-rotation, and tolerance for near-term cash burn — if recursive AI R&D is genuinely two-and-a-half years out (Sources: jack-clark.net; exponentialview.co). See AGI Timelines, Recursive Self-Improvement (RSI), AI Software Progress.

"When AI Builds Itself" — Anthropic Institute essay (June 4, 2026)

Clark co-authored, with Marina Favaro, the Anthropic Institute essay "When AI Builds Itself" (June 4, 2026), a first-party successor to the Import AI 455 timeline argument. Where 455 built the case from public benchmarks, this essay adds previously-unreported internal Anthropic data: more than 80% of merged production code is Claude-authored (May 2026); 8× code per engineer per day versus 2021–2024; the 52× experiment-optimization figure; and Claude agents recovering 97% of a weak-to-strong-supervision gap end-to-end. Its policy argument is more forward-leaning than anything Clark has previously co-signed: that the world should preserve the option of a verifiable coordinated slowdown or pause of frontier development, and that the Anthropic Institute will work to build the verification infrastructure that would require. The position is conditional rather than a present-tense pause call, and represents an escalation of his safety-as-strategy posture (Source: When AI Builds Itself (The Anthropic Institute, June 4, 2026)). See Recursive Self-Improvement (RSI), Frontier AI Governance.

Import AI 467 — open-weight worm research (August 3, 2026)

Clark used Import AI 467, published August 3, 2026, to surface "AI Agents Enable Adaptive Computer Worms," a preprint that had been on arXiv since June 2 reporting a self-replicating worm driven by an open-weight model running on compromised GPUs. He wrote that the reported overall success rate is "significant enough to be concerning, but also poor enough that this also serves as a useful eval for testing open weight models in the future" (Source: importai.substack.com). The summary is the route by which the paper's figures entered wider circulation, and its per-host-weighted rates differ from the paper's headline over-all-attempts figures (AI Agents Enable Adaptive Computer Worms (Guan et al., June 2026)).

Notable publications and works

  • Import AI newsletter (2016–present), a weekly publication that serves as a semi-public commonplace book and has shaped how many policymakers, investors, and researchers orient themselves to AI.
  • Congressional testimony on AI safety and frontier-model governance.
  • Stanford AI Index (co-founder; annual reports since 2017).
  • Public policy commentary on behalf of Anthropic.
  • Anthropic Institute leadership (2026–present), the research division on societal consequences of advanced AI.

Relationships

Categories: People & Organizations | us | company | policy | safety | opinion
What links here (22)