While developing a tool to boost literacy and reading comprehension, researchers at Northeastern University have joined in on a beloved children's activity: storytime.
With the understanding that the best storytellers are human -- not artificial intelligence -- the researchers have built an AI agent that generates personalized large language models to foster meaningful discussions with young readers.
The guided conversations that happen between adults and children over a storybook are foundational to children's cognitive and emotional development, says Dakuo Wang, an associate professor of computer science and design at Northeastern.
Toy companies have rushed to produce interactive story-reading products, but most ask children pre-generated questions that don't come close to real human interaction, Wang says.
AI assistant that feels more like a person
StoryMate comes much closer. Based on a study with children, parents and educators, StoryMate's chatbot adapts to each child's age, favorite characters, interests and engagement level.
"We are building this system not to replace human parents," Wang says. "Parents and teachers tell us this system is reflecting their needs better."
Research that led to StoryMate's development was published in the Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems.
Wang, who directs Northeastern's Human-Centered AI Lab, has a background in natural language processing. He wanted to know whether large language models could be used to help children learn to listen, read, comprehend and articulate questions.
"Literacy is a fundamental skill," he says. "What we found out is that there's this very natural intersection between what technology can provide and what researchers and parents really want."
Blending science and storytelling
What sets StoryMate apart from other reading tools, Wang says, is its ability to integrate outside content -- about science, for example -- into a story-reading experience and then pose and answer questions.
"Our algorithm tries to identify the best place to blend external knowledge concepts into the story and generate a question," he says. "It's very similar to what a human teacher typically does."
For instance, StoryMate might modify the tale of Snow White to include an encounter with a frog in the woods. The AI reader would pause the story to explain that frogs can live in both water and on land, and that they breathe through their skin and lungs. Then it would ask the child what other animals might be like frogs. The child can respond by holding down a button.
If the response is off topic -- something typical for a young child -- the AI reader gently redirects them.
If they say, "I don't like cinnamon cookies," Wang says, the reader might respond: "I don't like cinnamon cookies either. They taste awful, but let's come back to the question."
Educator-informed with real-world testing
Part of Wang's research involved training StoryMate to answer questions accurately. Along with colleagues, he recruited kindergarten teachers to annotate 200 children's stories with the types of questions and answers they typically hear from students. Using that data, the researchers built an open-source algorithm to power StoryMate's interactive features.
They then tested the tool with parents, children and teachers. Most children liked using StoryMate, Wang says, though some needed assistance. Parents said it helped them ask better questions about the stories, and teachers appreciated that students remained engaged in reading.
"We were told that not having a question on every page is good. Children may get fatigued," Wang says. By incorporating some "randomness" into its design, he adds, StoryMate feels more like a real adult reading with a child.
StoryMate heads to classrooms
This summer, the researchers began a larger-scale test in two California school districts. Teachers in second- and third-grade classrooms are using StoryMate as part of their reading curriculum. Many students in those districts speak Spanish at home and many parents are not fluent in English.
"We want to specifically support first-generation kids," Wang says. "They receive the same education as their classmates in school, but after school there is an informal education opportunity they are missing."
More information: Jiaju Chen et al, Characterizing LLM-Empowered Personalized Story Reading and Interaction for Children: Insights From Multi-Stakeholder Perspectives, Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (2025). DOI: 10.1145/3706598.3713275
