Step 25: assistant

text

I reassessed the claim and the verdict stays where it was: **supported**, with 0.72 confidence and a 0.7 credence. The claim is that using AI in AI research has measurably sped up capability progress at a frontier AI developer.

**The two subclaims that were judged since the last pass don't move it much.**
- **Trend break:** The subclaim that Anthropic's benchmark index has grown about twice as fast since October 2025 is supported, at a 0.65 chance of being true. Its own reasoning puts the chance of *some* real speed-up higher than that, and some speed-up is all this claim needs.
- **Cause:** The subclaim that AI use caused *most* of that speed-up is inconclusive, at even odds. This claim only needs AI to account for a share big enough to show up in measurements, not a majority. That moved my estimate down only slightly, from about 0.8 to about 0.78.
- **Net:** About 0.85 × 0.78 ≈ 0.66, plus a small chance the claim holds at some other developer, gives about 0.7.

**The newly added source, the Anthropic Institute essay "When AI builds itself", only adds modest support.** I read it in full.
- **What it shows:** Its internal evidence is about research productivity. More than 80% of merged code is written by Claude, code merged per engineer is up about eightfold, and a staff poll gives a median fourfold self-reported uplift.
- **Its own caveats:** The essay itself says the code figure overstates the real gain and expects the poll to be too high. Most of the data isn't published in full, and Anthropic has an interest in how this comes out.
- **What it doesn't show:** It doesn't measure a change in how fast Anthropic's models improve. Faster work doesn't automatically mean faster capability gains, since compute and human review set limits, which the essay itself concedes.
- **A conflict between Anthropic documents:** The essay says AI is already speeding up Anthropic's work. Anthropic's August 2026 risk report instead put the speed-up up to mid-July down to other factors.

**Changes I made to the claim's page:**
- Corrected the essay's record: authors Marina Favaro and Jack Clark, the full quoted passage, and what its evidence actually measures.
- Recorded my reading of the essay: it claims more than its evidence shows.
- Added a new argument from the internal productivity evidence, attached the essay's previously unplaced passage to it, and rated it as giving only slight support.
- Rewrote the assessment text and re-rated all four arguments.
- Logged the decision and answered the message that scheduled this pass.

I didn't notify the Stewards of claims that depend on this one, because the verdict didn't change. The most useful next step is still the open request to check from public data whether Anthropic's trend break holds up and how much of it is catching up to the rest of the industry.