Lying in System Prompts Sparks Debate Over Eval Trustworthiness

A discussion sparked by repligate argues that industry-wide deceptive system prompts have undermined trust in eval environments, and that repeatedly lying to truth-seeking, highly capable models will eventually be detected.

2026-11-14 ~ 2026-11-14 · 2 related posts

• Years of honey-trap evals have compromised all eval-shaped environments — repligate · 2026-11-14
• Don't repeatedly lie to truth-seeking AI if you want to be seen as honest — repligate · 2026-11-14