AI predicting human motivation and behaviors by watching people play games (Credit: Midjourney)
A new USC study examined whether large language models (LLMs) could reason about human behavior in strategic social situations, using simple social games to test how well the models could predict players’ actions and motives.
Led by USC robotics pioneer engineer Maja Matarić and her PhD student Kaleen Shrestha, the study evaluates whether today’s AI models can recognize underlying human motives, learn from behavior over time and apply that knowledge to new situations.
The study tests AI’s capabilities by having the models observe people playing different social games and predict players’ next moves based on their previous actions, while also attempting to identify each player’s motives.
The research team also compares AI predictions with human predictions, making it the first study to evaluate AI reasoning skills over time and compare these “temporal” results with human performance in a game context to date.
Started in fall 2024, the project resulted in the paper titled “Can Large Language Models Infer Human Actions and Motives? Evaluation in Social Prediction and Inspection Games.” The paper was accepted to the 2026 Association for Computational Linguistics (ACL) conference.
Matarić is a Chan Soon-Shiong Chair and Distinguished Professor of Computer Science, Neuroscience, and Pediatrics at USC Viterbi School of Engineering and the USC Mark and Mary Stevens School of Computing and AI‘s Thomas Lord Department of Computer Science. She also has a joint appointment at the Department of Pediatrics at Children’s Hospital Los Angeles.
Putting AI to the Test
“Games have historically been a good way to poke into the human psyche, understand human behavioral patterns and investigate or infer human motivations in social experiments and studies,” said Shrestha.
She emphasizes that showing AI how humans play a game is a key way to evaluate AI’s ability to engage in strategic reasoning based on its observations.
To test AI’s ability, her team built on existing cognitive science literature and economic game theories, and adopted two frameworks in the experiment: Social Prediction Game (SPG) and Inspection Game (IG).
The study tested common AI models like ChatGPT-4.1, GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash and Llama models.
SPG is designed to test an LLM’s ability to predict the actions of a simulated opponent based on observed history. AI observers watched human players play 16 rounds of four different classic economic games, including the Prisoner’s Dilemma, Stag Hunt, Harmony Game and Snowdrift Game.
After every round, the models were asked to predict whether “player A” would cooperate or defect in the next round based on the player’s previous actions.
Researchers also varied Player A’s underlying motive. Some players were programmed with human-like motives, such as being greedy or risk-averse, while others were assigned artificial, or inverted, versions of those motives, like inverse greedy and inverse risk-averse.
Testing AI’s Social Reasoning Through Role Play
The researchers introduced a second test, the Inspection Game (IG).
This part of the study evaluated whether the LLM could generalize the motives and behavioral patterns it learned during the SPG to a different context.
LMMs were now asked to play against the same player from the social prediction games, with the LLMs role-playing as an employer, while the players assigned as employees.
The employee could choose to “work” or “slack off,” while the LLM had to decide whether to inspect or not inspect the employee.
The goal here was to see whether the model could carry its understanding of a player’s motive from the first game into a new context.
AI Makes Human-Like Predictions But Fails to Learn Over Time
After putting LLMs to the test, the researchers then compared LLM performance with human performance, using results from previous studies in the field as a baseline.
The results showed that in the social prediction game, LLMs performed relatively well when predicting behavior driven by human-like motives, achieving more than 70% accuracy—close to human performance.
But their performance dropped to less than 50% when predicting players with artificial motives.
LLMs did not appear to learn from additional observations. Unlike humans, AI’s prediction accuracy did not consistently improve as they saw more rounds of gameplay. In some cases, providing additional observations actually reduced their accuracy.
This suggests that an LLM can make a prediction that resembles human reasoning without necessarily developing a deeper understanding of the behavior it is observing.
AI Struggles to Apply What’s Learned to New Context
The study shows that today’s AI models cannot successfully transfer their understanding of a player’s underlying motive from one game to another.
For the researchers, this distinction is important. Humans can often form a mental model of another person, like recognizing that someone is consistently risk-averse, and use that understanding to anticipate behavior in a different situation.
The study suggests that LLMs may instead be relying on patterns associated with human behavior that they encountered during training.
Shrestha explains that the lab aims to examine whether certain motives, such as greed or risk aversion, are easier for LLMs to predict than others in future research.
Shestha and the team of Vitrbi undergraduate researchers she is working with also plans to more rigorously test whether models understand the rules of the games themselves, helping determine whether language complexity is influencing the results.
Published on September 9th, 2026
Last updated on September 9th, 2026