A safety checklist in the prompt didn't measurably help frontier models
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A pre-registered test found that adding an eight-item safety checklist to frontier AI agents' instructions did not measurably reduce how often they fell for planted traps: 24.8% with the card versus 29.5% without, not statistically significant. It also did not harm legitimate work, which stayed within pre-set margins.