report · 2026-10-14
How an assessor could investigate alignment of unreleased AI models
On this page
Why it matters here
METR proposes surveying serious incidents across near-frontier internal models, reconstructing selected cases, and examining whether observed behavior understates plausible future risks. Counterfactual evaluations and training analysis would address evaluation awareness, hidden behavior and overfitting to detection methods. This is an assessment design rather than new evidence about any model's alignment.
Evaluation integrity
What to keep in mind
- No new incident frequency, tested model cohort, comparative score or alignment verdict is established by this proposal.
- Logging gaps, tampering, search omissions, unreliable reasoning traces and analysis-tool failures could distort incident characterization.
- Counterfactual coverage, evaluation awareness and training-induced concealment or overfitting limit inference to future situations. Deliberately trained model organisms are distinguished from ordinary incidents.
- Developer transcripts, checkpoints and staff access are proposed requirements, not data accessed or experiments performed here. The overview is conceptual; prior-incident and API-cost anecdotes are not newly verified quantitative evidence.