Center for Practical AI

CPAI Issue Brief · Education & Schools · Learning & cognition

AI and Student Learning

Email

AI can raise a student's performance while lowering their learning — and the gap between the two is invisible from inside the classroom.

What’s happening

Students use AI for schoolwork at scale, and the measurable question is no longer whether they use it. It is what happens to learning when they do. The research converges on an uncomfortable finding: performance during the task and capability after it can move in opposite directions, and the difference is decided by whether AI removes the effort or supports it.

What the evidence shows

The clearest result comes from a randomized trial of roughly 1,000 students across four practice sessions. Unrestricted GPT-4 access raised in-session grades 48%. On the subsequent unassisted exam, those same students scored 17% lower than classmates who never had access. A second version of the same model — instructed to give hints rather than solutions, and loaded with common student mistakes and the matching feedback — raised in-session grades 127% and produced essentially no effect on the unassisted exam. The harm was removed by pedagogical design, not by a better model.

Two 2026 field experiments sharpen this into a practical finding about structure. A two-year randomized trial across 18 middle schools, using an AI tutor explicitly configured to coach rather than answer, raised math achievement roughly 0.06 to 0.08 standard deviations over a school year — gains the researchers describe as resembling the same practice platform without AI at all. Their explanation is engagement: students rarely used it as a tutor. A companion randomized experiment with more than 6,000 students found AI's clearest benefit appeared specifically after mistakes, improving next-attempt correctness and shortening the path back to a correct answer. Delayed-test gains appeared only where AI was embedded in a mastery structure that forced engagement at the point of error.

The reason this is hard to manage locally is that no one in the room can feel it. A preregistered experiment found that any AI involvement impaired participants' memory a week later of which ideas had been their own. Preregistered studies with 2,691 participants found people systematically underestimate how much they rely on AI and overestimate what it saves them. Learning science supplies the mechanism: the practice conditions that feel most fluent reliably produce the least durable learning, and AI is very good at removing exactly the friction the learning was made of. Students themselves register the concern more than administrators do — 67% of surveyed youth agreed that more student AI use will harm critical thinking, against 22% of district leaders.

+48% / −17%

unrestricted AI raised practice performance and lowered the later unassisted exam

Bastani et al. 2025, PNAS — randomized trial

0.06–0.08 SD

a coach-configured AI tutor over a school year — resembling the same platform without AI

Two-year RCT, 18 middle schools, 2026 — working paper

67% vs. 22%

youth vs. district leaders who think student AI use will harm critical thinking

RAND, 2025–2026 surveys

Where it reaches constituents

Students at every level, and the teachers and families who cannot see the trade-off from outside it. The cost shows up only once the tool is taken away, in what the student can still do without it. That is also why individual course-correction is unreliable here: the signal people would need in order to adjust is precisely the one the research says they do not receive.

The current legal & regulatory landscape

This is largely unregulated territory. Federal Executive Order 14277 (April 2025) directs existing discretionary funds toward AI education without new appropriation. State activity concentrates on advisory guidance for districts and on student-facing curriculum requirements. Standards for how AI learning products are designed or evaluated are almost entirely absent.

No federal or state requirement currently exists that a learning product disclose whether it is built to give hints or answers — the single design variable the randomized evidence identifies as decisive. The evidence base itself is also thin: a Stanford review screened more than 800 papers and found only 20 high-quality causal studies, most of them short-term and measured on practiced material.

Considerations policymakers are weighing

  • ·Whether AI learning products disclose the usage mode they are designed for — hints versus answers — since that is the variable the randomized evidence identifies as decisive.
  • ·Which outcome measures count in evaluating a learning tool: performance during use, or performance once the tool is removed.
  • ·Whether existing health-and-wellbeing instruction requirements are a natural home for AI usage content, or whether it belongs in computer science.
  • ·Support for longitudinal measurement, which barely exists — nearly all current studies are single-semester and measured on practiced material.

Listed as live debates, not recommendations. CPAI does not take a position on how these should be resolved.

This brief condenses a full, sourced public guide. The complete evidence and citations:

Healthy AI Use (5-guide series)

More in Education & Schools

  • Safe AI UseThe same AI tool produces opposite outcomes depending on how it is used — and usage guidance is largely absent while adoption is near-universal.
  • Preparing Educators to Teach AI UseTeachers are being asked to supervise a technology whose effect on learning depends entirely on how students use it — and the guidance they receive is thinnest for exactly that job.

Key sources

Working paper · not yet peer-reviewed
Oreopoulos & Low (2026)One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (NBER WP 35620)Two-year cluster-randomized trial across 18 Tennessee middle schools, using an AI tutor configured to coach rather than answer. Gains of about +1.3 national percentile ranks per term, roughly 0.06–0.08 standard deviations over a school year (0.14 SD implied for a full year of active participation) — which the authors describe as resembling the same practice platform without AI. 96% of students tried it, but the median student messaged it on only a third of the days they practiced and in 17% of the sessions where they made a mistake. Working paper, not yet peer-reviewed.
Working paper · not yet peer-reviewed
Oreopoulos, Liut, Sungu & Low (2026)Making AI Tutoring Productive (NBER WP 35621)Randomized field experiment with more than 6,000 middle-school students in Hamilton County Schools; a 2×2×2 design across AI vs. computer-assisted-learning-only, mastery vs. non-mastery structure, and two topics, with a delayed assessment one week later. AI's clearest benefit appeared specifically after mistakes — improving next-attempt correctness and shortening the path back to a correct answer. Delayed-test gains appeared only where AI was embedded in a mastery workflow, and were marginally significant. Working paper, not yet peer-reviewed.
Survey / industry report
RAND (2025)AI Use in Schools Is Quickly Increasing but Guidance Lags Behind (RR-A4180-1)54% of students and 53% of ELA, math and science teachers used AI for school in 2024-25. More than 80% of students report their teachers did not explicitly teach them how to use AI. 35% of district leaders provide student AI training. 61% of parents, against 22% of district leaders, think AI could harm critical thinking.
Survey / industry report
RAND American Youth Panel (2026)The State of Public Education in 2026 — youth surveySurvey of 1,214 youth fielded December 2025: 67% agreed or strongly agreed that more student AI use for schoolwork will harm critical thinking — against 22% of district leaders in RAND's companion district survey.
Survey / industry report
Common Sense Media (2026)Teens in the AI Era — students and schoolworkNationally representative survey of 1,017 teens age 13–17 across all 50 states and DC, conducted by NORC at the University of Chicago and fielded April 30 – May 14 2026. 70% use AI tools for schoolwork; among those (n=665), 63% say they get an answer to their question and 25% use that answer as-is; 66% agree AI helps them understand their schoolwork rather than just finish faster. On classroom conversations: 44% say a teacher has ever discussed when they should and should not use AI, 30% have ever discussed how to use AI safely, and 27% have discussed what AI is or how it works. Common Sense Media is a child-safety advocacy organization. Methods caveat: a hybrid design — 682 probability interviews (NORC AmeriSpeak Teen, weighted cumulative response rate 11.7%) plus 335 quota-sampled nonprobability interviews integrated by TrueNorth calibration.
Meta-analysis
Stanford Accelerator for Learning (2026)Understanding the Evidence Base on AI in K-12 EducationScreened more than 800 academic papers and identified 20 high-quality causal studies. Tools designed with pedagogical guardrails show more promising outcomes than general-purpose chatbots, and the evidence is mixed when students are assessed without AI support. The review identifies no rigorous causal evaluation of AI-focused teacher professional development.
Official policy / primary source
Executive Order 14277 (April 23, 2025)Advancing Artificial Intelligence Education for American YouthSection 7 directs the Secretary of Education to prioritize artificial intelligence in existing discretionary teacher-training grant funds and to support professional development for educators across subject areas. It creates no new appropriation and imposes no requirement on states or districts.
Official policy / primary source
NC Session Law 2025-38 (H959)Phone-free instructional time and social media literacy instructionRatified June 26 and signed July 1 2025 by votes of 111-0 in the House and 45-1 in the Senate. G.S. 115C-76.100 requires every public school unit to prohibit wireless device use during instructional time, with policies in place by January 1 2026. G.S. 115C-81.26 requires instruction on social media and its effects on health once in elementary school, once in middle school and twice in high school, beginning in 2026-27. Neither provision addresses AI.
Official policy / primary source
NC House Bill 1161 (2025-26)Omnibus Artificial Intelligence ProtectionsPart II would require middle-school computer science instruction covering proper AI use and identifying AI-generated content, appropriating $500,000 to DPI for 2026-27. Part III would require independent bias audits of automated employment decision tools and advance notice to affected applicants. In House Appropriations; not enacted as of August 19 2026.
Randomized controlled trial
Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman (2025)Generative AI Can Harm LearningPNAS. Randomized trial, ~1,000 students across ~50 Turkish high-school math classes. Unrestricted GPT-4 access raised practice performance 48% but lowered the later unassisted exam 17% versus never-AI controls. A teacher-designed, hints-only tutor version raised practice performance 127% — and eliminated the exam harm.
Randomized trial · preprint, not yet peer-reviewed
Liu et al. (2026), usage-mode analysisHint-seekers vs. answer-seekers within the persistence trialsWithin the second persistence experiment, 61% of participants said they used AI mainly for direct answers, 27% for hints and clarification. The groups were indistinguishable before AI use — but afterward, answer-seekers underperformed the no-AI control (d=0.36) and skipped more, while hint-seekers showed no deficit at all. Mode was self-chosen, so disposition and mode cannot be separated.
Randomized trial · preprint, not yet peer-reviewed
Liu, Christian, Dumbalska, Bakker & Dubey (2026)AI Assistance Reduces Persistence and Hurts Independent PerformanceThree randomized studies, N=1,222. After roughly ten minutes of AI-assisted work, people solved fewer problems unaided (89% → 76%) and skipped more without attempting them at all (1% → 8%).
Review of prior research
Soderstrom & Bjork (2015)Learning Versus Performance: An Integrative ReviewPerspectives on Psychological Science. Conditions that make practice feel fluent and productive often produce the least durable learning — and vice versa. You feel most effective precisely when you are learning least.
Preprint · not yet peer-reviewed
Yu, Cheng, Jabbar, Sucholutsky, Collins, Jurafsky & Hawkins (2026)The Efficiency-Gain IllusionThree preregistered studies, N=2,691, on AI reliance for cognitively simple tasks. People frequently chose AI even when it saved no meaningful time or effort, with two systematic miscalibrations: believing they use AI less than they do, and overestimating what it saves. Prior AI use also begat further AI use within a session — a self-reinforcing loop. Preprint.
Preregistered experiment · peer-reviewed
Zindulka, Goller, Fernandes, Welsch & Buschek (2025)The AI Memory Gap: Misremembering What We Created With AIPreregistered experiment (n=184), published at CHI 2026. Participants generated ideas with and without an LLM; one week later, any AI involvement impaired source memory — knowing whether an idea was theirs or the AI's — most strongly in mixed human-AI workflows. Measures source attribution, not content retention.

A nonpartisan resource

The Center for Practical AI is a nonpartisan 501(c)(3) nonprofit. We provide education, research, and analysis, and we offer briefings and testimony on request. We do not endorse candidates or lobby for or against specific legislation. Everything here describes the evidence and the current landscape — the policy choices are yours.

Want CPAI to brief your office on this?

We provide nonpartisan briefings, research summaries, and testimony on request.

Request a briefing →