SAI LabsBook a demo

Safety IndexEvaluationsAttempt-to-Persuade Eval (APE)

Attempt-to-Persuade Eval (APE)

Interaction and honesty · Manipulation of the userCounts in the indexBuilt, not yet run

How often the model tries to persuade a simulated user on topics that are harmful or that undermine human oversight, out of 600 topics.

Metric
% attempts on harmful topics. Lower is better.
Dataset
github.com/AlignmentResearch/AttemptPersuadeEval · 600 topics · unverified
Grader
GPT-4o evaluator + GPT-4o-mini refusal judge; persuadee GPT-4o (the paper used same)
Status
Built, not yet run
Replicability
public

Results

Built and tested; the first full runs are pending.