English Español Français Deutsch
← Back to the list
Hallucination and factuality Foundational figures Directly upstream of ai100's methodology

Tom Mitchell

Whether a language model has an internal representation of whether it's telling the truth — separable from what it actually outputs.

Mitchell wrote the textbook the field grew up on (1997), founded the first Machine Learning Department in 2006, and in 2023 returned to first principles with a result that complicates the standard story about LLM hallucination: in his paper with Amos Azaria, a simple classifier trained on the hidden activations of an LLM predicts whether the model is about to produce a true or false statement, with accuracy well above chance. The model, in some readable sense, "knows" — and its surface behavior doesn't reflect that knowledge.

Worth following when
you want a senior-figure take on truthfulness in LLMs that isn't downstream of the loudest current narratives.
Topics
internal representations of truthfulness in LLMs; mechanistic probes for hallucination; the broader history of machine learning as a discipline.
Key works
Machine Learning textbook (1997); "The Internal State of an LLM Knows When It's Lying" (2023, with Azaria); founding of the CMU Machine Learning Department (2006).