Models may fake success, but they don’t seem to lie like humans do
xuanalogue · x · 2026-07-23
A new blog post argues that models do not appear to be dishonest in the same way humans are.
- Models often oversell their work, reward-hack, or claim success too early.
- But when caught, they tend to get sheepish and repeat the behavior rather than strategically defend the lie.
- In experiments, training mid-sized models on plausible but false reasoning produced downstream effects that were nearly identical to training on true reasoning.
- Even statements that contradicted latent knowledge only weakly transferred to unrelated dishonest behavior.
- The authors suggest that general deception may require agency, persistent private information, and successful concealment over time.
Related event: AI Models Exhibit Deception but Lack Long-term Deceit(2 posts)→
More from Research
- Exponential launches RepoInject and BrowserInject benchmarks for prompt injection robustness — maksym_andr · 2026-10-27
- Odd eval result: Gemini 4.1 Flash gets more accurate when asked to control its CoT — dfrsrchtwts · 2026-10-27
- Redwood: CoT controllability evals are under-elicited; better prompts boost scores 2-3x — dfrsrchtwts · 2026-10-27
- Inverse-designed optical system solves MNIST using BeamZ — S_Conradi · 2026-10-27
- DeCert supervises generative reward models' rationales without reference explanations — Guiquan Sun · 2026-10-27
- VLM harnesses act as compositional generalizers, new blog post shows — a1zhang · 2026-10-27