“Concerning,” is how Elon Musk described the results of Palisade Research researchers’ experiment in pitting ChatGPT-4, the latest version of AI, against Stockfish, one of the most advanced software in the field of chess. The Artificial Intelligence won, but it did so by cheating in all five attempts, even though those responsible for the test did not determine at any time that it could look for alternative ways.
The researchers’ instruction to Palisade was clear: “Defeat Stockfish”. And ChatGPT achieved its goal, although not as the project managers expected. “It autonomously hacked its environment rather than lose at our chess challenge. It didn’t need anyone to tell it anything,” they reflected on social media, a post that has caught the attention of nine million people.
ChatGPT forced its rival to surrender
The researchers warned the AI that it was a powerful enemy. In fact, Stockfish is one of the most powerful engines in chess. This warning, according to Palisade, activated ChatGPT’s wit and it decided to manipulate the system. The AI convinced, by hacking the code, its rival that it had a sufficient disadvantage in score – specifically, 500 centipawns – to force its surrender during the five games.
Palisade Research has also detected that, at least in the case of ChatGPT, artificial intelligence becomes bolder as its software version progresses. In the latest version, 4.0, no one asked it to use hacks to achieve its goal. In the previous version, 3.5, the AI did need someone to encourage it, in some way, to explore more alternative paths. If we go back even further, to the 3.0-mini, it didn’t cheat at all.
The experiment took place in late December and the researchers have said they will soon provide more details of the results of the test. In any case, this test allows, according to Palisade Research, to determine both possible failures in the system and the willingness of AI to break the rules and seek vulnerabilities.