The Anthropic Institute is the research division of Anthropic concerned with the societal consequences of advanced AI. It publishes on its own path of the company's website and is the named author of Anthropic's June 2026 essay "When AI Builds Itself", which disclosed internal data on AI accelerating AI research and committed the Institute to researching the verification systems that a credible coordinated frontier slowdown or pause would require. It is a division of a for-profit company rather than an independent institution.
Establishment and leadership
The Institute's existence became public through an announcement by the Cosmos Institute on April 30, 2026, which described co-founder and policy head Jack Clark as heading it (Source: blog.cosmos-institute.org).
Accounts of who leads the Institute differ. Reporting on the June 2026 essay consistently identifies Marina Favaro as the Institute's head or lead — described as "Marina Favaro, head of the Anthropic Institute" and as "Marina Favaro, lead at the Anthropic Institute" in contemporaneous coverage (Source: tradingview.com; journalrecord.com). Clark's role as policy head of Anthropic and Favaro's as head of the Institute are not necessarily inconsistent, but no source reconciles them and neither formulation has been corrected. Both are recorded here rather than resolved.
Provenance note: the leadership divergence above is unresolved in the sources. Anthropic's own institute landing page (anthropic.com) is client-rendered and returned no readable content on attempted capture, so the company's own statement of leadership is not held. Citations should upgrade once that page or an Institute publication stating its own governance is ingested.
Publications
"When AI Builds Itself" (June 4, 2026) is the Institute's most-cited output. Written by Favaro and Clark with editorial support from Santi Ruiz and published at anthropic.com, it argues that AI is already accelerating AI research and development, and supplies internal Anthropic figures in support: more than 80% of merged production code Claude-authored as of May 2026, an eightfold increase in code merged per engineer against 2021–2024, an experiment-optimization speedup rising from roughly 3× to 52×, and Claude agents recovering 97% of a weak-to-strong-supervision gap end to end. It sketches three forward scenarios and closes with the commitment described below. The full argument and its reception are covered at When AI Builds Itself (The Anthropic Institute, June 4, 2026); the evidence it supplies anchors Recursive Self-Improvement (RSI).
"Measurements for understanding the pace of AI development inside frontier labs" (September 17, 2026) proposes three measurements of how frontier models are built and reports a snapshot of each from inside Anthropic: a prototype R&D Automation Index, on which Claude "leads" 26% of Anthropic's AI R&D work as of August 2026; coverage, review latency and escalation rate for internal AI agents, about 30,000 of which were working at any one time on the company's most-used internal platform; and the share of AI R&D compute allocated to safety, about 6% over one week in July 2026. It was co-authored by Favaro and Phillie Wright, with research direction credited to Clark, and the Institute's index page lists it as "Illuminating the Frontier" (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)).
An economic scenarios working paper is published at anthropic.com. It is named in contemporaneous coverage of September 2026 but is not held in the source record, so no account of its contents appears here (Source: New Developments Log/2026-09-13-2205-ai-developments.md).
Anthropic's Economic Policy Framework assigns the Institute a role in its own forward work, stating that the framework's third and most speculative tier of policy thinking is to be developed "in partnership with The Anthropic Institute, policymakers, and external researchers."
Verification and coordinated-slowdown work
The commitment stated in "When AI Builds Itself" is that the Institute will research and help build the verification systems a credible coordinated slowdown or pause would require, and will convene policymakers, researchers, civil society and other labs "in the coming months." The position is conditional rather than a present-tense call to pause: the essay argues a unilateral pause "accomplishes much less," and that an effective one would need multiple frontier labs in multiple countries stopping under shared triggers, each able to verify the others. It rests on a detectability claim that training runs are harder to conceal than missile silos (Source: When AI Builds Itself (The Anthropic Institute, June 4, 2026)).
This places the Institute at the most binding end of the instrument spectrum described at Frontier AI Governance, and distinguishes it from prohibition-first proposals such as ControlAI's ASI Security Bill: the Institute's output is lab-internal and verification-first rather than legislative. Its position is compared against other coordinated-slowdown proposals at Coordinated slowdown proposals compared.
Relationships
- instance-of: Frontier AI Governance — the Institute's verification commitment sits at the coordinated-slowdown end of that spectrum.
- depends-on: Anthropic — a division of the company, not an independent body.
- supports: Recursive Self-Improvement (RSI) — its June 2026 essay is the first frontier-lab disclosure of internal data on AI accelerating AI R&D.
- related: Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026), Anthropic R&D Automation Index — its September 2026 measurement proposals and internal snapshot.
- related: Jack Clark; When AI Builds Itself (The Anthropic Institute, June 4, 2026); Anthropic's Economic Policy Framework (June 2026); Coordinated slowdown proposals compared; Compute Governance (the enforcement and detectability theory its verification argument rests on); ControlAI (the legislative counterpart position).
Open questions
- Who holds formal leadership of the Institute, given that the Cosmos Institute announcement and press coverage of its June 2026 essay name different people.
- Whether the convening of policymakers, researchers, civil society and other labs promised "in the coming months" in June 2026 has occurred, and with whom.
- The contents of the economic scenarios working paper, which is published but not held in the source record.