Understand this story
On October 24, 2026, the University of Chicago published a paper pointing out that large language models (LLMs) cannot reliably self-report the key factors in their own decisions. The paper states that LLMs cannot predict the key questions on which their own answers diverge, and that the key decision factors they self-report match the true factors only 13.4% of the time, lower than the 39.2% achieved after random shuffling. The study had models repeatedly predict 59 March Madness games and 15 Metaculus events, then used a reranker to compare answer pairs, with computation about 86 times lower than local LLM judging. The proportion of individual answers covering the true key factors was 28.6%, versus a baseline of 37.8%. Currently this finding has been reported by a single source, with no subsequent developments yet.
AI-generated from public reports
Latest development · Report
University of Chicago: LLMs' self-reported key decision factors are unreliableDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 24
Rohan Paul
University of Chicago: LLMs' self-reported key decision factors are unreliable
A new paper from the University of Chicago finds that LLMs cannot predict the key issues behind disagreements in their own answers; the key decision factors they self-report match the true factors only 13.4% of the time, lower than the 39.2% for randomly shuffled ones. The study had models repeatedly predict 59 March Madness games and 15 Metaculus events, then used a reranker to compare answer pairs, requiring about 86 times less computation than local LLM judging. A single answer covered 28.6% of the true key factors, versus a baseline of 37.8%.