In a Minecraft-like game, all three groups first explored different states, then increasingly sought situations they could influence.
Researchers compared 26 children aged 8 to 12, 24 adults aged 18 to 25, and reinforcement-learning AI agents playing Crafter, a Minecraft-like game with no stated goal or visible score. Players could explore for up to 20 minutes, while the researchers measured their progress against three information-based objectives: state diversity, information gain and empowerment, or the ability to reach and influence different states.
Human exploration progress was consistently associated with state diversity and empowerment, but not information gain. State diversity increased quickly before reaching a plateau, whereas empowerment rose steadily across play for both people and agents. The researchers also found preliminary evidence that children who spoke more about their goals explored more effectively.
How exploration changed
Adults explored more effectively than the AI agents. Agents trained with intrinsic rewards—rewards designed to encourage exploration rather than complete a specific task—also performed less effectively than agents trained with a reward tailored to the game's structure.
Among people, exploration progress was positively correlated with entropy, a measure related to how varied the states visited were, and with empowerment, which measures how much control an explorer has over possible future states. Information gain was not consistently associated with human exploration progress. Entropy increased rapidly early in play and then leveled off, while empowerment continued to increase, suggesting that novelty may be more informative early and control later.
In an exploratory analysis of recorded speech, children who used more words performed better. In particular, the share of children's utterances that mentioned goals was positively associated with the rate at which they unlocked achievements. This relationship was not found for adults, and the researchers describe the speech findings as preliminary.
Why novelty and control matter
The results suggest that a single exploration objective may not provide equally useful guidance throughout an open-ended task. Measures of novelty could help an AI system explore early, while measures of control may better reflect progress later. The comparison with human behavior also points to entropy and empowerment as potentially useful targets for designing intrinsic rewards, although the tested agents did not yet match adults' exploration.
The findings about children's goal-directed speech offer a possible clue about how people organize exploration, while also suggesting a question for future studies: whether explicitly forming or expressing goals can reliably support exploration in children and AI systems.
Evidence and caveats
The study combined a preregistered, ethics-approved comparison of human players with trained AI agents in the same game. It included 26 analyzed children and 24 adults; speech analyses used recordings from 17 children and 23 adults. The researchers trained agents with different reward objectives and compared them with a random-action baseline.
The sample was small and came from specific age groups and settings, and participants played only one open-ended game for up to 20 minutes. The researchers caution that conclusions about human exploration may not generalize to other populations or environments. The speech analysis was exploratory and needs confirmation in a larger preregistered study. Measures such as entropy and empowerment also depend on how the game's states are represented, which may not fully capture how people understand the environment.
// Source
Proceedings of the National Academy of Sciences · 2026 · DOI: 10.1073/pnas.2612366123
Authors: Aly Lidayan, Yuqing Du, Eliza Kosoy, Maria Rufova, Pieter Abbeel, Alison Gopnik C
Institutions: University of California, Berkeley
