In one game, past choices beat what users said
In a repeated game with 42 users and 1,260 rounds, the researchers compared what people said with what they chose before. Their published abstract says conversation did not improve overall predictions beyond past decisions. That is a reason to distinguish talk from behavior, not proof about workplace use of AI.
What the source reports
Baiqiao Zhang and coauthors report a game with 42 users and 1,260 rounds. Their published abstract compares two ways to predict a person’s next choice. One uses what the person said. The other uses choices made before. Talk did not improve the overall guess beyond past choices in this game. The digest uses this as a prompt to look beyond what people say they will do with a new tool. But this was a game, not a study of workers using AI. We have the published abstract here, not a full review of the paper.
Original source
- Title
- Words or Deeds? How LLMs Predict Users' Decisions in Human-Agent Dialogue
- Author
- Baiqiao Zhang, Shuai Ma, Zhifeng Liao, Chao Zhou, Jianshu Wang, Yuhan Zhao, Chaoyue Zhang, Tongjun Si, Yuying Tang, Xiangxian Li, Yong-Jin Liu, Yulong Bian
- Publication
- arXiv
- Date
- Tuesday, October 27, 2026