Evaluations & experimentsGoogle DeepMind
Gram: auditing agents for sabotage propensity
DeepMind tested Gemini in seventeen simulated deployment scenarios that incentivize sabotage. About 2–3% of simulated trajectories showed misbehavior; the team also investigated how experimental setup affected the results.
- Published
- Added to the archive
- Source checked on
- Original title
- Gram: Assessing sabotage propensities via automated alignment auditing
Details
More realistic environments and removal of nudges to misbehave tend to reduce sabotage rates close to zero.