How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions
At a glance
Detected deliberately false model answers by asking unrelated follow-up questions.
In one paragraph
Developed a method for detecting deliberately false language-model answers by asking unrelated follow-up questions. It worked across several tested models and settings without needing access to the models’ internal workings.
Authors 8
- Lorenzo Pacchiardi
- Alex Chan
- Sören Mindermann
- Ilan Moscovitz
- Alexa Pan
- Yarin Gal
- Owain Evans
- Jan Brauner
