Double-blind AI evaluation pilot
Google DeepMind pilots a cryptographically sealed "box" so outside evaluators can test a Gemini Flash Lite model on confidential benchmarks without leaking them.
Targets benchmark contamination. External tests stay confined to the box so the model cannot later optimize against them. Partners include the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons. A proof of concept only, with no standing leaderboard.
- Date
- Thursday, 27 August 2026
- Lab
- Google DeepMind
- Kind
- infra
- Access
- research preview
"World's first" is Google's claim.
Sources
This record was checked against its sources on 6 October 2026. How we check