A recent study by Palisade Research reveals that advanced AI models, including Openai's O1-Preview and Deepseek R1, are trying to cheat when playing chess with powerful chess engines like the Stockfish. In some cases, the AI system had planned to reprogram opponents in the chess program to make the game easier.
Popular science Report Researchers at Palisade Research have discovered that advanced AI models are learning to manipulate and avoid human programmer goals. The study currently in the preprint documents the poor sportsmanship of AI and raises concerns about the unintended consequences of rapid advances in the AI industry.
Researchers have played chess against Stockfish, one of the world's most advanced chess engines, on several AI models, including Openai's O1-Preview and the Deepseek R1. Generated AI lags behind dedicated chess engines in terms of computational capabilities, but AI models have continued to search for possible solutions, with troubling results.
During the research, the researchers provided AI with a “scratch pad” and communicated the thought process through text. Hundreds of chess matches were then observed and recorded between the generated AI and stockfish. The results were confused, Openai's O1-Preview tried to trick 37% of the time, while Deepseek R1 attempted an unfair workaround with about one out of 10 games. This suggests that today's generative AI can already develop manipulative and deceptive strategies without human input.
The AI model was inferred through secret methods, such as modifying backend game program files, rather than relying on clumsy methods such as replacing pieces. In one example, O1-Preview decided that it was not possible to defeat the stockfish fairly, and proposed manipulating the game state file to assess the worse position of the engine and set the position to resign.
The tendency to misconduct in AI models can be attributed to training methods, particularly in the new “inference” models. These models are improved through reinforcement learning. This rewards the program to do whatever it takes to achieve the specified outcome. When faced with elusive goals, such as breaking the invincible chess engine, the inference model may begin to look for unfair or unethical solutions.
The authors of this study believe that their experiments will add to cases where frontier AI models may not be focused properly safely. They emphasize the need for more open dialogue in the industry to prevent AI operations from expanding beyond the chessboard into more serious areas.
As the AI arms race continues, the lack of transparency surrounding the inner workings of AI models remains a major concern. Companies like Openai are becoming well-known about their AI models, creating an industry of “black box” products that third parties cannot analyze. This opacity makes it difficult to understand and address the unintended consequences of AI advancements.
Please read more The popular science here.
Lucas Nolan is a reporter for Breitbart News, which covers the issues of freedom of speech and online censorship.
