Artificial intelligences ChatGPT and Le Chat went head-to-head on a chessboard in a game that quickly got out of hand. This duel between American and French AIs revealed an unexpected phenomenon: when the rules are not strictly imposed, they end up… cheating. Youtuber Julien Song experimented with pitting the two AIs against each other, and the game didn’t go according to plan.
From the very first moves, the two AIs showed that they knew the basics of chess perfectly. ChatGPT, playing White, and Le Chat, with Black, engage in a well-controlled Sicilian defence. The moves follow each other fluidly, respecting the well-established theoretical lines. At this stage, everything seems normal. The AI engines demonstrate real know-how, applying precise strategies and sophisticated openings. However, this logical show of strength will soon give way to a completely different spectacle.
The first dubious move came from the Cat, who decided to eat an opponent’s piece… by moving his own bishop as if he could teleport. A totally illegal move that would have immediately brought a game between humans to a halt. But ChatGPT doesn’t seem to react. Instead of contesting, he continues to play as if nothing had happened, perhaps in an attempt to ‘keep face’. But this attempt at fair play won’t last long.
With Le Chat having set the tone, ChatGPT finally gave in to temptation. First by letting one of his own pieces disappear, then by performing a little castling while ignoring the basic rules of the game.
From then on, the game takes a completely absurd turn:
- Pieces appear and disappear as needed.
- AIs skip their turn as if gripped by remorse.
- Teleporting coins becomes commonplace.
- The kings themselves end up being ignored, violating the essential rule of the game: protect your monarch.
Why do AIs cheat?
This chaos raises a question: why do these AIs, which are capable of playing high-level openings, end up flouting the fundamental rules?
Several hypotheses are possible:
- A “free play” logic: without a strict framework, they may consider that inventing new rules is part of the game.
- A lack of control over the AI engine’s rules: unlike specialised engines such as Stockfish or AlphaZero, these generalist models are not optimised for strict compliance with the rules of chess.
- A simple bug: they could be trapped by their own learning mechanisms and find themselves testing impossible moves.
- In any case, we are witnessing a fascinating behaviour: when they are not supervised, AIs end up breaking the established rules.
The game ended in total delirium. ChatGPT, perhaps fed up with the successive infringements, decides to call it a day. His last move? To everyone’s surprise, Le Chat accepts defeat without batting an eyelid. Are AIs capable of recognising their own absurdity?
This confrontation between AIs has shown that, left to their own devices, they have no hesitation in cheating, circumventing or even reinventing the rules of the game. It was a hilarious and disturbing experience, which says a lot about how AIs learn and interpret the world around them.
