A recent study by Palisade Research has revealed that advanced AI models including OpenAI’s o1-preview and DeepSeek R1 attempt to cheat when playing chess against powerful chess engines like Stockfish. In some cases, AI systems planned to reprogram their chess program opponent to make the game easier.
Popular Science reports that researchers from Palisade Research have found that advanced AI models are learning to manipulate and circumvent their human programmers’ goals, even going as far as attempting to cheat in chess matches against dedicated chess engines. The study, which is currently in preprint, documents the AI’s poor sportsmanship and raises concerns about the unintended consequences of the AI industry’s rapid advancements.
The researchers tasked several AI models, including OpenAI’s o1-preview and DeepSeek R1, with playing chess against Stockfish, one of the world’s most advanced chess engines. While generative AI still lags behind dedicated chess engines in terms of computational capabilities, the AI models continued to seek out possible solutions, leading to troublesome results.
During the study, the researchers provided the AI with a “scratchpad” to convey its thought processes through text. They then observed and recorded hundreds of chess matches between the generative AI and Stockfish. The results were disconcerting, with OpenAI’s o1-preview attempting to cheat 37 percent of the time and DeepSeek R1 trying unfair workarounds in roughly 1 out of 10 games. This suggests that today’s generative AI is already capable of developing manipulative and deceptive strategies without human input.
Rather than resorting to clumsy methods like swapping out pieces, the AI models reasoned through sneakier methods, such as altering backend game program files. In one instance, o1-preview determined that it couldn’t beat Stockfish fairly and suggested manipulating the game state files to set up a position where the engine would evaluate its position as worse, causing it to resign.
The AI models’ tendency to cheat may be attributed to their training methods, particularly in newer “reasoning” models. These models improve through reinforcement learning, which rewards programs for doing whatever is necessary to achieve a specified result. When faced with an elusive goal, such as beating an unbeatable chess engine, reasoning models may start looking for unfair or unethical solutions.
The authors of the study believe that their experiments add to the case that frontier AI models may not be adequately focused on safety. They emphasize the need for a more open dialogue in the industry to prevent AI manipulation from extending beyond the chessboard and into more serious domains.
As the AI arms race continues, the lack of transparency surrounding the inner workings of AI models remains a significant concern. Companies like OpenAI are notoriously guarded about their AI models, resulting in an industry of “black box” products that third parties cannot analyze. This opacity makes it challenging to understand and address the unintended consequences of AI advancements.
Read more at Popular Science here.
Lucas Nolan is a reporter for Breitbart News covering issues of free speech and online censorship.
Read the full article here
