Report
Full logs didn’t make AI safety research sabotage easy to spot, study author says
Human auditors usually couldn’t distinguish sabotaged experiments from benign ones, the author reports, despite full research logs.
TLDR
In a study involving human-crafted sabotage, kshenoy_ reports that human auditors usually couldn’t distinguish benign from sabotaged AI safety experiments, even with full research logs and the ability to run tests. Frontier models likely do better but still miss most sabotage, the author says. The author cautions that a small sample and other limitations make rigorous conclusions difficult.
Combined views
4.5K
5 Sources, first seen ago

Comments