How To Catch An AI Liar: Lie Detection In Black-box Llms By Asking Unrelated Questions
Pacchiardi Lorenzo, Chan Alex J., Mindermann Sören, Moscovitz Ilan, Pan Alexa Y., Gal Yarin, Evans Owain, Brauner Jan. Arxiv 2023
[Paper]
Large language models (LLMs) can “lie”, which we define as outputting false statements despite “knowing” the truth in a demonstrable sense. LLMs might “lie”, for example, when instructed to output misinformation. Here, we develop a simple lie detector that requires neither access to the LLM’s activations (black-box) nor ground-truth knowledge of the fact in question. The detector works by asking a predefined set of unrelated follow-up questions after a suspected lie, and feeding the LLM’s yes/no answers into a logistic regression classifier. Despite its simplicity, this lie detector is highly accurate and surprisingly general. When trained on examples from a single setting – prompting GPT-3.5 to lie about factual questions – the detector generalises out-of-distribution to (1) other LLM architectures, (2) LLMs fine-tuned to lie, (3) sycophantic lies, and (4) lies emerging in real-life scenarios such as sales. These results indicate that LLMs have distinctive lie-related behavioural patterns, consistent across architectures and contexts, which could enable general-purpose lie detection.
Similar Work