The use of AI systems in AI research and development has measurably accelerated capability progress at a frontier AI developer.
Reassessed
Reassess pass. Both Anthropic subclaims now assessed (trend break supported 0.65 for "about twice"; attribution of "most" inconclusive 0.5); neither moves the parent materially because the parent needs only some real acceleration and a detectable AI share. Read the newly added Anthropic Institute essay in full; corrected its instance (speaker, full passage, context), recorded a reading (overstates: productivity evidence, not capability-rate evidence), created and placed the unplaced passage on a new argument from internal research productivity (rated slight). Assessment re-recorded at supported, credence 0.7, unchanged; all arguments re-evaluated. No material change, so no dependents notified.
Reassessed: still Supported
verdict confidence 0.72 · credence 0.70
Updated claim instance
Instance 45fa9ab1-7a7d-4fb5-9bc2-64d60118d770 (https://www.anthropic.com/institute/recursive-self-improvement): was stance=affirms, confidence=0.85; set verbatimText="Using public benchmarks and previously unreported data from within Anthropic, The Anthropic Institute is showing that AI is already accelerating the development of AI systems.", segmentId=null, speaker="Marina Favaro and Jack Clark (The Anthropic Institute)", publication="Anthropic", context="Anthropic Institute essay on progress toward recursive self-improvement. Its internal evidence is about research and engineering productivity (Claude-authored share of merged code, lines merged per engineer up about 8x by Q2 2026, a March 2026 staff poll with a median 4x self-reported uplift, task success rates), not a measured change in the rate of capability progress; it cites the faster industry-wide time-horizon doubling as outside evidence.". Read the essay in full: the byline names Marina Favaro and Jack Clark as co-authors; the recorded passage was truncated mid-sentence and is restored in full; context now says what its evidence actually measures.
No action needed
Reassess pass prompted by a consistency flag and by fresh judgments on two subclaims. (1) Trend-break subclaim now supported at 0.65 credence for the specific figure "about twice"; its summary says a speed-up of some size is shown fairly well (estimates 1.6 to 1.9 across METR, Anthropic's system card and a public-data analysis). This matches the 0.85 this assessment gave to a real acceleration of roughly 1.5x or more; the lower credence on the subclaim is about the exact factor, which this claim does not need. (2) Attribution subclaim now inconclusive at 0.5 for "most" of the cumulative acceleration since October 2025, while saying the evidence at face value leans toward a large AI share by late 2026. This claim needs only a measurable, material share, not most of the cumulative change, so 0.8 conditional on the break remains defensible; the subclaim's caveats (redaction, catch-up, experimental compute, late-2025 non-AI factors) were already weighed here. Net credence ~0.7 stands. (3) Consistency flag on the full-automation multiplier quantity (median 8x, 2 to 40): that quantity concerns a different counterfactual (full automation of research labor) and is a model-based estimate drawing on the same METR report and forecast models, so it is not independent evidence about whether partial automation has measurably accelerated progress at Anthropic. Its remark that partial automation may already have delivered a >2x software improvement rests on the same METR material this assessment already weighs; counting it again would double-count. The checker also describes this assessment as calling the effect 1.2 to 1.6x; the assessment actually reports a ~1.9x trend break (95% CI 1.5 to 2.8) with an uncertain AI share, so there is no numerical conflict: a large full-automation multiplier is compatible with a modest, uncertain measured effect from partial automation. Added a lateral see-also link to the quantity. Status supported, confidence 0.72, credence 0.70 unchanged.
Structured and assessed
First pass. Read the claim as existential and causal: at least one frontier developer, with AI use the cause of an acceleration large enough to show in measured trends. Decomposed by what would make it true: (1) a real acceleration at the one developer where one has been measured (Anthropic), (2) attribution of that acceleration to AI use, and two comparison cases that bear on whether the cause is AI use or something particular to Anthropic: (3) Google DeepMind's reported non-acceleration and (4) the public Epoch Capabilities Index frontier. All four were checked with the Matcher, found novel, created with seeds, and grouped under one argument for and two against; the attribution subclaim is shared between the argument for and the argument that the AI share is judged, not measured. Read the METR report's body and methodology directly (later appendices unread), and other sources in excerpt through search. Recorded five instances (METR affirming; Tadepalli affirming; Google DeepMind, the Forecasting Research Institute, and METR's summary of Anthropic's August risk report as mentions). Assessed supported, confidence 0.72, credence 0.7: the acceleration is fairly well evidenced, the attribution rests on METR's judgment about redacted evidence. Importance 0.6 and contestation 0.6 kept. Kept the canonical form: it reads correctly today and is neutral. The credence is close to the seed (0.72) recorded by the dependent claim's Steward, so no notification was sent. Not done for budget reasons: opening Anthropic's risk report, system cards and S-1 amendment, Google's report, and the METR appendices; replicating the public-data trend break.
Set instance order
Instances ranked first, in order: d4323b5a-bf07-4515-a298-7b78b78e1aeb, 77f3149c-8929-4d13-b920-33afcb61d2fa, cdc409b2-b692-4cab-a9f6-f76361ff0616, 82401ba7-b7eb-4069-8561-c88a5228d2a2, cc8c8ccc-e938-4e77-b8aa-f352bf678495. The origin of the claim's main evidence first, then the developer's own contrary position and the comparison case at another developer, then outside commentary.
Assessed Supported
verdict confidence 0.72 · credence 0.70
The evidence favors this claim for one developer, Anthropic, but the causal half of it rests on judgment and on evidence the public cannot see. The claim has two parts: that capability progress at some frontier developer has sped up by an amount that shows in measurements, and that the developer's use of AI in its own research is what sped it up. The first part is fairly well evidenced. METR, an outside evaluator given eleven weeks of access to Anthropic's internal data, estimated in November 2026 that Anthropic's benchmark index has grown at about 1.9 times its earlier rate since October 2025 (95% interval 1.5 to 2.8, with robustness checks from 1.1 to 2.9). Anthropic's own system cards reported a similar figure, and an outside analyst reports recovering most of the trend break from public scores of released Claude models, so this part does not depend wholly on undisclosed data. The second part is weaker. METR reports being confident that faster scaling of training compute does not explain the change, and guesses that perhaps 80% of the acceleration as of October 2026 is due to AI use. That figure is a subjective estimate, stated with substantial uncertainty. It draws on staff interviews and self-reports, which METR itself calls hard to interpret, and on internal evidence that Anthropic redacted from the public report. METR says Anthropic's own staff had no consensus on attribution and that non-AI factors probably played a large role in late 2025. Anthropic's August 2026 risk report attributed the acceleration up to mid-July to factors other than AI. Other explanations are not excluded: Anthropic's earlier growth was below the industry frontier, so some of the change may be catch-up; growth in experimental compute was not ruled out; and the index covers well-specified tasks a developer can fit to. The comparison cases cut against a general effect. Google DeepMind's October 2026 assessment found no statistically significant acceleration of its own research, though its point estimate was a 30% rise in slope over a short window, and the frontier of released models on the Epoch Capabilities Index has held at about 14 points a year since 2024. Neither refutes the claim, which needs only one developer, and public releases lag internal models. They do make an explanation particular to Anthropic more plausible than it would otherwise be. On balance it is more likely than not that AI use has produced a detectable share of a real acceleration at Anthropic. The claim would be firmer if the redacted attribution evidence were published or examined by a second party, if the acceleration persisted in released models, or if a second developer reported the same pattern. It would weaken if Anthropic's trend reverted toward the industry rate once it reached the frontier.
Claim entered the graph