AI made kids better storytellers. But take the AI away, and that advantage just disappears.

A new study, published in Computers in Human Behavior in March 2026, randomly assigned 180 children (average age about 6.5, from North China) to three different ways of getting AI help with storytelling.

Kids were split into three groups: Passive-Active Retelling, Autonomous Storytelling with AI Feedback, and Iterative Storytelling with AI Feedback. The three conditions were designed around the ICAP framework -- a well-established learning-science model that ranks engagement from Passive, to Active, to Constructive, to fully Interactive, with more interactive engagement generally predicting deeper learning. In practice, that meant the groups differed in how much back-and-forth the child actually had with the AI while building a story: retelling and reacting to an existing story on the lower end, versus repeatedly telling, getting feedback, and revising a story on the higher end. The scaffolding techniques themselves were adapted from established methods researchers already use with human adults -- dialogic reading (where an adult asks a child questions about a story to deepen engagement) and elaborative reminiscing (where an adult helps a child build out and enrich a memory or story in detail).

After controlling for baseline performance, age, and interaction duration, children in the Iterative Storytelling condition -- the most back-and-forth, feedback-heavy condition -- outperformed those in the other two groups on all narrative measures during the scaffolding phase. Those measures included story structure (whether a story has a clear beginning, complication, and resolution), story complexity, and internal state terms -- the words a child uses to describe characters' thoughts, feelings, and motivations, a well-established marker of narrative sophistication in child-development research.

Here's where it gets interesting: without continued AI support, no condition differences emerged in transfer to new stories. The advantage that showed up while the AI was actively helping simply didn't carry over once the AI was taken away and kids had to tell a brand-new story completely on their own.

Children in the Iterative condition also reported higher enjoyment and interest than kids in the other two groups -- so even though the skill gain didn't clearly transfer, the experience itself was more engaging for kids getting the richest back-and-forth with the AI.

"Lasting transfer appears to require sufficient practice and interaction time," the researchers concluded. AI-powered feedback works while it's there, but one session of scaffolding isn't enough to build a skill that sticks once the scaffolding is removed.

For parents and educators leaning on AI tutoring tools, the lesson is direct: watch for whether a skill actually transfers to unassisted work, not just whether performance looks good with the AI still in the loop. A single strong session with an AI tutor -- for storytelling, writing, or any other skill -- isn't proof a child has actually internalized anything. The real test is whether they can do it again later, with the AI gone.

Source: https://www.sciencedirect.com/science/article/abs/pii/S0959475226000307 (Computers in Human Behavior, March 2026)