Signal Silent Revision: Measuring Undisclosed Change in the Safety Frameworks of Frontier AI Developers

Summary

This study proposes a silent revision rate to measure how much of the material change in frontier AI developers' safety frameworks goes unacknowledged in their own published change announcements. The researchers traced 710 commitment changes across consecutive version pairs from twelve developers. Under a strict standard, 67 percent of these changes were silent, falling to 53 percent under a more lenient standard. Narrative-style announcements were silent more often, at 74 percent, than itemized changelogs, at 63 percent. Seventy-seven percent of the traced changes weakened or removed an existing commitment, and weakenings were silent more often than strengthenings. The authors argue that disclosure rules in the EU and California should require developers to enumerate specific changes rather than simply narrate them.