Jack Clark

persona · active · confidence: medium · reviewed 2026-09-21 · tier: quarterly

Who he is

Affiliation: Co-founder & Head of Public Benefit, Anthropic, leading The Anthropic Institute (announced 2026-03-11) — role check run 2026-09-21: the v1 line "Head of Policy" was eight months stale. He handed policy to Sarah Heck (formerly Anthropic's Head of External Affairs, before that Stripe) and now runs the company's in-house research-and-disclosure arm, which unites the Frontier Red Team (stress-tests models), Societal Impacts (how AI is used in the real world) and Economic Research (jobs and the economy), plus new teams on forecasting and on AI-versus-the-legal-system (Anthropic, 2026-03-11; title confirmed by TIME100 AI, 2026-08-27 and his Oxford speaker bio). Previously Policy Director at OpenAI (2016–2021); before that a technology journalist — "the world's only neural network reporter at Bloomberg" and "distributed systems reporter at The Register" by his own newsletter bio (v1's "TechCrunch" could not be confirmed this cycle and is flagged, not carried). Steering-committee member of the Stanford HAI AI Index (co-chairs are Ray Perrault and Yolanda Gil — he is a member, not a director; HAI). Author of Import AI, weekly since 2016 (issue 15 is dated 2016-11-07 in the archive), "read by more than 100,000 people every week." British, based in San Francisco.
One-line position: AI capability is accelerating faster than most observers internalize — he now puts a year on it: recursive self-improvement (AI that designs and builds its own successor without people) "by 2028, or perhaps sooner" — and the tractable response is mandated transparency from the labs plus third-party validation, compute-level governance, and, new in 2026, some coordination mechanism that can slow the field if the evidence demands it — because, in his words, "there's a collective action problem here" (NPR Morning Edition, 2026-09-15). The 2026 shift: from arguing for governance to manufacturing the evidence for it from inside the lab.

Discipline & technical bet

Not a technical researcher; a measurement-and-policy operator who has spent a decade turning lab observations into public argument. The concrete wager he is making now is institutional, not architectural: that an in-house think tank publishing "information that only Anthropic knows" can move policymakers before a crisis does. The Institute's May-2026 agenda names four problems — "Economic diffusion," "Threats and resilience," "AI systems in the wild," "AI-driven R&D" — and promises to "publish those findings, to help external organizations, governments, and the public make better decisions" (agenda, 2026-05-07). The technical claim underneath is that AI is already automating AI research: "more than 80% of the code we merge into Anthropic's codebase was authored by Claude" (as of May 2026), "the typical engineer was merging 8× as much code per day as they were in 2024," and Claude "leads" 26% of Anthropic's AI R&D work as of August 2026, up from under 1% in February ("When AI builds itself," Favaro & Clark, 2026-09-18; "Illuminating the Frontier," Aug 2026). His bet is that these internal numbers, published on a cadence and eventually checked by "independent third-party evaluators," are the missing instrument in AI governance — the AI Index idea, rebuilt inside a frontier lab.

Key claims (Says)

Notable predictions — with falsifiable checks

Revealed behavior (Does)

Feels

Fear he has chosen to keep: "The pile of clothes on the chair is beginning to move. I am staring at it in the dark and I am sure it is coming to life" — paired, in the same essay, with "I am a true technology optimist." Underneath both: a journalist's dread of being the one who saw it and didn't say so — hence the Institute, hence "It matters that leaders are honest about what they see."

Hears

Anthropic's Frontier Red Team, economists and safety researchers; the Amodeis; Congress and the national-security committees (ten years of hearings); the Stanford AI Index committee; a wide weekly arXiv diet that Import AI filters. Increasingly downstream of his own teams' data — the input and the output are now the same organization.

Sees

Frontier internals no outsider has: the share of a frontier lab's code and R&D written by its own models (80%, 26%), compute allocated to safety (about 6% of AI-R&D compute, 12% of AI-driven AI-R&D compute), monitor coverage of agent decisions (0.002% blocked), and the labor-market "observed exposure" data (young workers' entry into the most exposed jobs "decreases by about half a percentage point" a month while the unemployment gap "is small and insignificant"). He has become the source for the numbers everyone else argues about — which is the point, and the problem (see Incentive map).

Incentive map

Re-derived this cycle; conflict is higher than v1's "policy/comms, not P&L" reading. Who pays him: Anthropic, where he holds co-founder equity in a company valued at $965 billion that filed to go public on 2026-06-01. What he is selling: the Institute's output — and the Institute's founding premise is that AI is about to get dangerous fast. What that does to his claims: his timelines ("RSI by 2028," "powerful AI by late 2026 or early 2027") are also the story that justifies a near-trillion-dollar valuation and a $47-billion run rate; the alarm and the pitch point the same way. His regulatory asks (mandated transparency, third-party validation, frontier-only thresholds) are exactly the asks a well-capitalized incumbent can afford and a startup cannot — Sacks's charge is self-serving but not empty — while also being the asks most likely to constrain Anthropic itself, which is why the Pentagon fight is the best evidence in the file that the safety commitments are real. New this cycle: $40 million of company money in a political vehicle means his "positions" now have candidates attached, and a pending S-1 means a quiet period may soon limit what he can say at all. What he cannot say: "our models are less capable than we imply" (valuation), "the Pentagon has a point" (litigation), "state laws are a mess" (they endorsed SB 53), "RSI isn't coming" (the Institute's reason to exist). Credit where due: he printed the bearish creativity signal on his own timeline in August 2026 — the one public act this cycle that cut against his interest. Rule-25 discount: maximal on timelines and on anything sourced from Institute measurements of Anthropic itself; moderate on governance-institution claims, which are checkable against public bodies; light on the Import AI research summaries, which are about other people's papers.

Theories aligned with

What he's reacting against

Where he overlaps / splits (with Rich)

Track record

Empirical vs normative

Weak spots / open questions

Rich's take

Delta log

2026-09-21 — v1→v2 migration + validation (bulk run)

Sources

migrated v1→v2 2026-09-21 · backup: _archives/jack-clark.html.bak-20260921 · converts-from: people/jack-clark.md · AI & Society domain