AI in K–12 Education: The Good, the Bad, and the Guardrails to Consider
This post was originally published on the Institute of Education Sciences (IES) Regional Educational Laboratory (REL) Northeast & Islands blog.
As artificial intelligence (AI) permeates everyday life from workplaces to homes to classrooms, concerns about its rapidly evolving role in K–12 education are also growing. Prompted by a sense of urgency and uncertainty regarding AI, the Governing Board for REL Northeast & Islands—a panel of state, district, and school leaders from each of the region’s nine states and jurisdictions charged with overseeing its programmatic work—turned to REL’s Ask an Expert (AAE) service. The Board requested a timely update on the state of the research on AI and student outcomes to help inform concrete guidance for teacher and student use.
In response, REL Northeast & Islands drew on its expertise to gather and synthesize relevant research. To share their findings, the team hosted a meeting with Governing Board members to discuss current needs, challenges, and promising directions related to current AI tools and student learning. In this blog, we share emerging trends from the review of the research as well as implications and guardrails to consider for practitioners and education leaders.
What the Evidence Shows So Far
While the evidence on AI in education is expanding, with new studies being published daily, strong causal studies examining how current AI tools promote or hinder student learning are still extremely limited. AI seems to be advancing faster than research can evaluate it. So how can education leaders figure out what is helpful and what is harmful when it comes to AI?
Artificial intelligence (AI):
Computer systems that can perform tasks associated with human intelligence, such as recognizing patterns, generating content, making predictions, or adapting responses based on data.
The AAE team searched IES’s What Works Clearinghouse, the most trusted repository of scientific evidence for education. Although it did not identify any research studies on AI and student outcomes, it did locate a 2026 comprehensive review of existing research on AI in K–12 education. The review found only 20 rigorous education research studies that produced causal evidence about AI’s impacts. It also found that most research on AI has been conducted in postsecondary settings, with causal studies more common in high school than in middle or elementary school settings. Even so, some interesting patterns are beginning to appear—patterns that educators may recognize from research on earlier educational technologies.
The emerging evidence on AI tool use and student learning suggests three general patterns:
- Teacher-mediated and AI-augmented tutoring may promote student learning.
- Student-facing AI tools show mixed effects on education outcomes.
- General-purpose AI use may hinder students’ learning, particularly if it reduces students’ cognitive effort or replaces student thinking.
We’ll examine each of these findings in more detail below.
Teacher-Mediated and AI-Augmented Tutoring May Promote Student Learning
AI works best when it supports—but does not replace—educators. The most promising research on AI and student outcomes to date comes from studies of teacher-mediated and AI-augmented tools—AI tools that are either used by or alongside a teacher or tutor. A helpful way to think about these tools is that they keep a human (or in this case a teacher) in the loop. In other words, a teacher remains actively involved by reviewing, guiding, or approving what the AI does. These AI tools proved helpful in improving student learning when they provided teachers and tutors with insights or tailored instruction for individual students, such as through AI-generated diagnostic reports that identified students’ specific learning patterns. Other AI tools have assisted tutors in real time by suggesting instructional strategies or guiding in-lesson support which was associated with greater student mastery of topics.
Research also points to the benefits of human-supervised AI tutoring. In this approach, students use an AI model that has been fine-tuned for pedagogy, while expert tutors provide direct supervision. So, while students use AI chatbots, there is tight human-in-the-loop oversight. These students performed as well as or better than those working with human tutors alone.
Student-Facing AI Tools Show Mixed Effects on Education Outcomes
It’s important to remember that not all AI is created equal. When students use AI for support during the learning process, the effects on student performance have been mixed. Simply giving students access to an AI tool, for example, did not improve exam scores but did improve homework performance. Similarly, when students used tutoring-specific AI chatbots that provided hints without giving away answers, they performed as well as students using traditional study methods on an exam—but not better. AI also appeared to be more effective when it was paired with traditional learning strategies, such as note-taking, than when used on its own.
AI that is specifically designed to support student thinking and provide feedback or hints instead of answers, is associated with better student outcomes. Tutoring-specific tools that provide suggestions and step-by-step reasoning may help students to internalize that logic to foster deeper learning. These Socratic-style AI chatbots ask users probing questions rather than simply telling students the right answers.
General-Purpose AI Use May Hinder Students’ Learning
Unsurprisingly, when AI does the thinking, student learning suffers. Students who used general-purpose AI chatbots, such as ChatGPT or Claude, while learning or studying, were more likely to have lower exam performance, shallower learning processes, reduced brain activity, and weaker recall even if the students felt that the tools were helpful. In fact, students using a traditional search engine for a research paper had higher-quality reasoning and argumentation than students using a general-purpose AI chatbot.
Together, these findings suggest that how AI tools are designed and used is an important factor in effectiveness. AI is more likely to support learning when it is purposefully designed, instructionally integrated, and used to support—not substitute—teacher expertise and student thinking.
Guardrails to Consider
While AI tools have distinct capabilities and risks, prior research on educational technology initiatives such as massive open online courses (MOOCs) and technology rollouts or computer-assisted instruction suggests that simply providing access to technology is not enough to improve student outcomes. Success requires thoughtful design and implementation that integrates technology into instructional practice rather than replacing it. Similarly, the evidence on AI does not support unmanaged adoption; it supports thoughtful use with clear guardrails. Here are a few to consider:
- Use AI to support, not replace, student thinking. General-purpose AI chatbots may harm learning, especially when they do the information processing and problem-solving necessary for students to engage in independent learning. One guardrail to consider is avoiding the use of AI to replace the productive struggle that is essential for deeper thinking.
- Center human relationships in teaching. Emerging evidence suggests that students may perceive AI-mediated feedback as less caring and supportive than feedback that comes from teachers. While AI might help both students and teachers, nothing replaces human connection.
- Vet tools with learning, privacy, and access in mind. There is a gap in the literature about how AI tools could benefit students who lack access to private tutoring or extracurricular supports. However, students’ ability to benefit from these tools may also depend on more than just their availability. Until stronger evidence around AI use in education is available, the same caveats we apply to using technology in the classroom should also apply to AI so that we don’t create disparities in learning opportunities, risk student privacy, or shortcut learning.
- Use AI interactions as a means of teaching critical thinking skills. Research indicates that AI companions are designed to be particularly engaging through “sycophancy,” meaning a tendency for feedback to match users’ beliefs and provide validation, rather than challenging their thinking. Having students identify and review multiple sources is a valuable skill across the board, not just for AI use. Also, having AI users include a phrase in their prompt like “don’t just , agree with me” can prevent these sycophantic responses.
Decision-makers should be cautious about how AI tools are designed and consider whether the underlying design choices have guardrails in place that are intended to preserve deep learning and support student reasoning.
In addition to these guardrails, leaders can:
- Create or refine local guidance. Formal policies might have to be updated regularly as AI evolves, but thoughtfully designed guidance will provide educators with appropriate, scaffolded support.
- Invest in educator capacity. Provide examples, templates, and shared learning opportunities. Regular professional development and a chance for teachers to engage with AI themselves is critical.
- Clarify goals for using AI. Like educational technology in general, AI needs to have a clearly defined purpose in the classroom.
Recommended Resources:
- National Guidance: Artificial Intelligence and the Future of Teaching and Learning.
- Blog Post Summarizing State Guidance: AI Pilot Program in K–12
