School leaders have spent the past two years trying to keep generative AI out of classrooms. A growing body of work suggests the more productive question is how to design tools that help students think rather than do the thinking for them.
At Lehigh University's College of Education, researchers are testing that distinction through a suite of AI agents called StemPal, which includes MathPal for mathematics, StatPal for statistics, and HackPal for programming. Instead of handing over a finished solution, the systems offer hints, explanations, questions, and step-by-step prompts designed to keep students engaged in solving problems themselves. The approach rests on a principle teachers have long understood: struggle is part of learning.
Scaffolding, not shortcuts
"Good teachers do not necessarily answer a student's question by giving away the solution," the research team notes at Lehigh. "They ask another question. They provide a clue. They remind students of something they already know." The AI agents are built to behave the same way: provide enough assistance to keep a student moving without removing the thinking required to get there.
Early research with MathPal suggests students notice the difference from conventional AI tools. High school students who used the system reported that it helped break difficult problems into manageable steps and provided assistance when their teacher was busy with someone else, especially in large classes. A separate effect also emerged: students became better at communicating with AI.
They learned that better questions produced better assistance. That skill - knowing how to formulate queries, evaluate returned information, spot wrong or incomplete answers, and identify what they still don't understand - may be one of the most important things schools can teach in the age of generative AI. Those are not shortcuts around critical thinking. They are forms of it.
Teacher oversight comes first
One of the most worrying visions of educational AI is that intelligent tutoring systems will eventually make teachers unnecessary. The Lehigh project takes the opposite stance. AI can collect information about how students approach problems, what kinds of hints they request, and where they get stuck, but information is not instruction. A teacher must still interpret what is happening, understand the student, and decide when intervention is necessary.
To support that role, the project includes a teacher-facing dashboard that lets educators monitor student interactions with the AI agents during and after class, preserving human oversight. The model pairs well with broader training for educators: an AI for Teachers Learning Path and ongoing discussion of AI for Education can help instructors develop the skills to guide students using these tools.
Early conversations with students align with research on educational scaffolds, strong signals that when AI responds to a student's request for help, the response can be tailored to keep them on track - even if the overall conversation continues to evolve. Teachers and parents still need rules governing academic integrity, and students need to understand when a tool becomes a substitute for work rather than an aid to it.
Why this matters for educators
The practical question for teachers and administrators is no longer whether AI belongs in the classroom, but what kind of AI to adopt. The Lehigh project offers a concrete benchmark: when evaluating tools, ask whether the system gives answers or actually structures students on how to work through a problem. The goal should not be to produce students who get answers faster. A well-designed system, with good teacher oversight, helps students ask better questions and keep thinking when the solution is not obvious.
Your membership also unlocks:
