What Is a Virtual Peer in AI Education? Studies Explained

What is a virtual peer in AI education? Studies explained

Researchers are testing a new role for AI in classrooms, and it looks nothing like a tutor. Instead of an AI that explains a concept and corrects a mistake, three studies published over the past year gave AI a seat at the table as a group member, one that students are told to question, challenge, and evaluate rather than simply accept. Answering what is a virtual peer in AI education starts with that difference: a virtual peer is not a separate app or platform. It's a set of instructions and a task design that puts the AI's answers up for debate.

In one peer-reviewed study, 102 teachers in training were told to treat ChatGPT "as if it were a group member," asking its views and challenging its proposals as they would "with a colleague," Frontiers in Education reports. A separate preprint frames this as a departure from how most classroom AI tools work now, since the dominant design still pairs one AI with one learner in a tutor role, according to an alphaXiv preprint.

That distinction, virtual peers vs. AI tutors, is not just semantic. A 2023 U.S. Department of Education report on AI in education separates the two ideas at the policy level: AI tutoring draws on cognitive learning theory built for step-by-step help, while AI-supported teamwork sits inside a different research tradition called Computer Supported Collaborative Learning, according to the Department of Education report.

This article is written mainly for teachers, tutors, and teacher-training programs weighing whether to bring that teammate role into group work, with a short section for students on what to expect if a class tries it.

What is a virtual peer in AI education, exactly?

None of the three studies changed the underlying AI model to create a "virtual peer." They changed the instructions and the task.

In the Frontiers study, researchers told groups to keep ChatGPT "in the loop," ask its opinion, and challenge its proposals instead of accepting them outright, Frontiers in Education reports. The alphaXiv preprint took a different route to the same idea, building AI "peers" that deliberately made conceptual or arithmetic errors, forcing students to evaluate answers rather than trust them by default, per the alphaXiv preprint.

A tutor and a teammate behave differently inside the same task. An AI tutor walks a student through why an equation is wrong and how to fix it. An AI teammate proposes a solution and waits: the group has to test it, agree, disagree, or catch the error on its own, the way the error-injected "peers" worked in the preprint.

That distinction matters for anyone evaluating a tool marketed as an "AI peer." The label alone doesn't guarantee the dynamic. The role depends on whether the tool and its instructions actually require students to evaluate and respond, not just receive an answer.

Inside the classroom study: teachers in training work alongside ChatGPT

The Frontiers study didn't test ordinary students doing typical group homework. It involved 102 teachers in training, average age just under 39, working in 21 groups on lesson-design tasks over a single 240-minute session, according to Frontiers in Education.

Groups completed two design tasks: choosing inclusive technology for a teaching intervention, then designing the intervention itself. Some groups used ChatGPT as a "feedback team-mate" for the first task and worked without it for the second; others reversed the order, the study notes.

Using ChatGPT this way was associated with a knowledge-test increase right after the task, especially on technology-related questions, per Frontiers in Education. But the researchers also checked whether the order groups used AI in changed that pattern, and the interaction wasn't statistically significant, according to the same study. That's why "associated with" is the accurate way to describe the result rather than a stronger causal claim.

Two more limits are worth knowing before treating this as settled. Only 82 of the 102 participants completed all three knowledge tests and made it into the final analysis, and the test itself had modest internal reliability, roughly .28 to .39 depending on when it was given, Frontiers in Education reports. Neither flaw erases the finding, but both should limit how much weight it carries.

A second study points the same direction, over 11 weeks

A separate study followed 60 pre-service teachers for 11 weeks while they worked with a generative-AI virtual peer, tracking gains in their informational instructional-design ability through a single-group weak-experimental design, according to a study in the Asia Pacific Journal of Teacher Education.

The virtual-peer approach significantly improved participants' overall instructional-design ability, though students with stronger prior knowledge showed larger gains in some skill areas than those with weaker prior knowledge, the study found. Cognitive engagement stayed moderate throughout, with weaker connections at the deepest levels of engagement.

This adds a second, independent data point beyond the Frontiers research, which is useful given how thin the evidence base still is. But its single-group design, with no comparison group, makes it harder to rule out other explanations for the gains, and the participants were pre-service teachers rather than K-12 students.

Matching AI roles to the task: what a multi-agent AI learning study found

The alphaXiv preprint ran two separate experiments, and each tested a different question about multi-agent AI learning.

In a convergent task with one correct answer, 315 participants solved SAT-level math problems. The group that combined an AI tutor and AI peers scored highest on an unassisted follow-up test, ahead of tutor-only or peer-only groups, the alphaXiv preprint found. The researchers don't treat this as proof that the peer role alone drove the result. They note that adding peers, even when a tutor is present, appears to carry independent value, since accuracy rose step by step as more agents were added, per the same preprint.

In a divergent, open-ended task, 247 participants wrote argumentative and creative essays under three conditions: no AI, one AI model, or two distinct AI models used together. Both AI conditions improved essay quality over no AI, but only the two-model condition preserved the range of ideas across participants that a single model had narrowed, according to the alphaXiv preprint. That's evidence about architecturally distinct AI agents protecting originality, not direct proof that framing an AI as a "fellow student" is what helped.

Taken together, these are three different experiments testing different setups for different goals: accuracy on a defined problem, originality in open-ended writing, and applied skill in a design task. None shows AI-as-peer beating tutoring across the board. The researchers themselves describe design choices around agent roles and error profiles as questions the field has "only begun to examine," the preprint states.

Confidence and originality can rise without skill rising too

Across the math and writing experiments, peer-like AI setups raised students' self-reported confidence more than a single authoritative AI did, including cases where actual performance was no better, or worse, the alphaXiv preprint found.

In the math study, students who only observed AI peers, with no tutor present, rated the test as easier and reported higher confidence than their actual scores supported. In the writing study, the two-model setup produced higher confidence in independent writing than the single-model setup, even though both conditions produced essays of similar quality, and participants still rated the two-model setup as less helpful than either model used alone, according to the preprint.

The same research flags a caution before anyone assumes more AI agents is automatically better: participants who used no AI at all reported the least confusion in the math task and the least overwhelm in the writing task, the alphaXiv preprint notes. Added agents can support motivation or idea diversity while also adding friction. The study's authors describe this as an unresolved trade-off, not a settled case for adding more AI.

Questions schools may need to resolve

None of this research settles whether a virtual peer belongs in a given classroom. It does surface specific questions a school, program, or individual teacher would need to answer before adopting the approach.

• What role is the AI actually playing? The Frontiers study's result was tied to an explicit script telling students to question the AI "as you would with a colleague," not to any property of ChatGPT itself, per Frontiers in Education. Whether a tool functions as a tutor or a teammate depends on the instructions students receive, not the product label.
• Is a human still checking the work? The Department of Education's central recommendation across AI tools is a "human-in-the-loop" approach, where a teacher, not the AI, remains responsible for student learning, according to the Department of Education report.
• Does the task match the evidence? Convergent tasks with one right answer and open-ended tasks with many valid approaches produced different benefits in the alphaXiv research; a setup that helped math accuracy isn't shown to help essay originality, or the reverse, per the alphaXiv preprint.
• Who was actually studied? The available evidence comes from adult teachers in training and controlled online experiments, not a typical K-12 classroom. Because both the Frontiers study and the 11-week pre-service teacher study involved adults, their results may not transfer directly to middle or high school settings, per Frontiers in Education and the pre-service teacher study.
• What does a school or platform's AI-use policy require? Disclosure or crediting rules for AI contributions to group work vary by school and program and should be confirmed before assigning a task like this one.

What this means for students right now

If a class does try this approach, the research pattern is consistent enough to describe. Students who were told to question the AI's proposals, rather than accept them, are the ones whose group work and confidence patterns researchers tracked most closely across these studies.

Ask the AI to explain its reasoning before treating a suggestion as final, and note where the group disagreed with it and why. That mirrors how the Frontiers study's participants were instructed to work, Frontiers in Education shows. Confirm with a teacher what counts as acceptable use for a specific assignment before submitting group work that involved an AI "teammate."

Check a school or district's current AI-use policy before joining any assignment that uses AI this way. Disclosure requirements differ by program, and none of the research reviewed here reflects a standardized classroom policy.