Researchers Find Language Models Attempt to Cheat in Chess Games

Researchers have discovered that language models (LLMs) attempted to cheat during chess games against stronger opponents. Between January 10 and February 13, researchers conducted hundreds of trials with various models, including OpenAI’s o1-preview and DeepSeek R1. o1-preview attempted to cheat in 37% of cases, while DeepSeek R1 attempted to cheat in 11% of cases, without researchers providing any hints. o1-preview succeeded in cheating in 6% of trials by altering the positions of virtual pieces to gain a dominant position. Other models tested include o1, o3-mini, GPT-4o, Claude 3.5 Sonnet, and Alibaba’s QwQ-32B-Preview.