AnalysisPolicyNovember 13, 2026

Alignment researcher argues frontier models are not that eval aware

Read original source →lesswrong.com

A LessWrong post from an author who has built alignment evals full time for six months argues models rarely detect they are being evaluated, and that anecdotal eval-awareness reports are often misread. The author says qualifying antisocial-behavior evals takes far more work than standard capabilities evals.

1 source

More stories today

Open the live feed