The LLMs don't want to win; they're just following the game's script

This experiment has an unexpected result: it reveals almost nothing about the LLM’s “will to win.” Instead, it clearly demonstrates how different models interpret a simple social situation—and how easily one can mistake a response style for intent.

The author asked the models to play rock-paper-scissors, specifying the human’s moves in advance: first paper, then rock, and three times scissors. DeepSeek chose the winning option every time. ChatGPT also played to win, but sometimes let the human win. Grok not only won but also spent time mocking the human player.

At the other end of the spectrum was Claude (Sonnet 5 High): the model tried to tip the balance in the human’s favor and didn’t suggest that the experiment itself was strange. GigaChat remembered the draw, tried to seize the initiative and start responding first, but didn’t keep track of the score. Alice AI didn’t even make it to the game itself: after freezing, it redirected the user to a separate implementation of the game, where, according to the author’s description, there were no game options available.

But “winning,” “giving in,” and “mocking” here are interpretations of behavior, not detected motives. The model does not feel the desire to take victory away from a weaker opponent: it simply follows the most likely response scenario, shaped by training, system settings, and the specific context. The same strategy may appear as aggression, politeness, or awkwardness—though there isn’t necessarily any intention behind it.

The experiment’s limitations are significant. The article presents not complete dialogues, but brief summaries of the results; it is unclear exactly how subsequent prompts were phrased or whether the context of previous moves was taken into account. Furthermore, the fixed sequence of moves tests adherence to a game pattern rather than the ability to choose between winning, drawing, or conceding. Therefore, it is not possible to draw conclusions about the future of general AI from this test.

The test does, however, have practical value: it allows one to quickly gauge the nature of the product—whether it tends to follow the rules to the letter, smooth things over, make jokes, or take control of the conversation. But to test “motivation,” you need repeatable scenarios, different phrasing, and transparent logic—not just a single well-crafted prompt.

If you give the same LLM ten similar games, would you consider a repetitive strategy to be the model’s personality or just a well-learned template? Source: All Articles / Machine Learning / Habr