67% of Material Changes to Frontier AI Safety Frameworks Are Made Without the Developer Saying So
Researchers assembled a versioned, hash-pinned corpus of every public safety-framework version from the twelve developers that have published one, plus each provider's changelog, redline or announcement, then traced 710 commitment instances across twelve consecutive version pairs and individually adjudicated 244. Under a strict standard 67% of material changes are silent (95% CI 62 to 72), falling to 53% lenient and 49% at section granularity, with narrative announcements running 74% silent against 63% for itemised changelogs. Most consequentially, 77% of traced changes weaken or remove a commitment, and in seven of eight pairs weakenings are more often silent than strengthenings, which matters because the EU and California now treat these documents as accountability instruments.
↳ Follow the thread