Children increasingly encounter AI in learning contexts, but they may not always judge whether its suggestions are accurate or misleading. To explore this, researchers built a game-based collaboration probe that embedded AI suggestions into a biological knowledge reasoning game. They tested three agent interaction styles—partner-like, tool-like, and dual-state—and observed how children verified the AI's input during the task.

The study, which included children aged 8–10, combined interaction logs, post-task ratings, and interviews. Instead of measuring trust as a static attitude, the researchers tracked situated verification behaviors such as acceptance, evidence checking, questioning, revision, and continued AI use. The agent style appeared to shape these pathways: partner-like interaction supported direct questioning, while tool-like interaction was more associated with seeking evidence.

Children generally viewed the AI as useful but fallible, and their questioning behaviors varied in reasoning depth. The findings suggest that the design of AI agents can influence how critically and reflectively children engage with AI suggestions. The authors offer design insights for supporting children's evidence-based use of AI in collaborative tasks, rather than passive acceptance.