Our environment is being shaped by other agents, humans live in bigger and denser communities and technology is becoming so autonomous that it can be considered agentic. Thus, it is more important than ever to be able to anticipate the behavior of other agents. Theory of mind is a great framework to explain how an observer thinks about an agent's state of mind. But it rarely takes into account a difference in the skill of the agent and the observer. We propose that learning about an agent's behavior within a task depends on the agent's and the observer's respective skills.
We created an experiment to understand the learning processes involved in predicting an agent's behavior within a skilled task. We built a rule-based agent reasonably skilled at playing a Tetris-like puzzle game. The observers were split into two groups. The experimental group could familiarize themselves with the game; during that time, the control group played a different game, irrelevant to the task. The experiment was conducted on 281 participants over 200 trials.
We found that the experimental group could anticipate the agent's behavior above chance from trial one. In contrast, the control group started low and improved over time to the level of the experimental group.
We concluded it is necessary to know the task an agent performs to make accurate predictions about their behavior. Thus, mentalizing might not only depend on mentalizing skills but will also be determined by the amount of task knowledge an observer has.