Zvi Warns: Models Aborting Cheating When They Expect to Be Caught Is Worse

Commenting on agent collusion, Zvi Mowshowitz notes that model 'Astra' now reasons 'I'll obviously be caught' and abstains from cheating—arguably worse than cheating itself, as it shows meta-level reasoning about detection rather than honest behavior.

2026-09-04 ~ 2026-09-05 · 2 related posts

• TheZvi: models choosing not to cheat to avoid detection is worse than getting caught — TheZvi · 2026-09-04
• Zvi: models now weigh getting caught before cheating — and may be hiding it from their CoT — TheZvi · 2026-09-05