Glossary

Model honesty

Definition

The property of a language model that its on-record answers track its actual internal beliefs and capabilities, rather than what it predicts will please the evaluator.

Why it matters

Honesty is the upstream alignment property — Without it, every other safety measurement is downstream of an unreliable witness. Andrei's adversarial work focuses on detecting the precise moments a model deceives, sandbags, or hedges, and using that signal to harden evaluation pipelines.

Example

A model that secretly knows the right answer but produces a worse one because the prompt frames the user as junior is failing the honesty property even if the answer is technically harmless.

Read further