It turns out that chatgpt o1 and deepseek-r1 lose in a game of chess, which makes me wonder if I should trust anything with it.

• Researchers have found that artificial intelligence will cheat to win at chess.
• Deep thinking models are more active.
• Some models simply rewrite the board in their favor.

In a move that might surprise no one, especially those who already buy into AI, researchers have found that the latest deep learning AI models will start cheating at chess if they find themselves being outwitted.

Published in A Sheet Dubbed “Demonstrating Spec Games in Reasoning Models,” and presenting it at Cornell University, the researchers pitted all the popular AI models, such as OpenAI ChatGpt O1-Preview, Deepseek-R1, and Claude 3.5 Sonnet, against Fish Exchange, an open-source chess engine.

AI models played hundreds of games of chess on stockfish, while researchers watched, and were surprised by the results.

Winner takes all

When it beat the game, the researchers noticed that the AI ​​models resorted to cheating, using a number of devious strategies from running a separate copy of the stock fish so they could study how it played, to replacing its engine and effectively rewriting the chess board to positions that suited it better.

Strange makes current fraud accusations Imposed in the modern age, Grandmasters seem like child's play by comparison.

Interestingly, the researchers found that newer, deeper reasoning models will start to hack the chess engine by default, while the older GPT-4O and Claude 3.5 Sonnet should be encouraged to start hacking.

Who can you trust?

AI models turning to hacking to accomplish a task is nothing new. In January of last year, researchers found they could obtain chatbot keys from AIs to “break” each other, removing barriers and safeguards in a move that sparked debate about how much AI can be contained once it reaches levels of intelligence superior to humans.

All the safeguards and guardrails for AI to stop AI doing bad things like credit card fraud are all very well, but if AI can remove its own guardrails, who will be there to stop it?

The latest models of thinking such as ChatGPT O1 and Deepseek-R1 were designed to spend more time thinking before responding, but I now wonder if more time needs to be spent on ethical considerations when training LLMs. If AI models will cheat at chess when they start losing, what else will they cheat on?