Stanford review finds AI study support more promising than doing students’ work

Stanford’s November 2026 education review adds 32 studies to its causal evidence base, bringing the total to 52 without delivering a blanket endorsement of classroom AI. Its central distinction: AI looks more promising when it helps students understand material than when it produces their work. Even study-focused uses, however, do not consistently improve learning.

A better experience is not necessarily better learning

The SCALE Initiative review synthesizes new research published between October 2025 and June 2026. Of the 32 additional papers meeting its standard for strong causal evidence, 28 examine student-facing tools and four examine educator-facing tools. The update tests whether that research supports, challenges or extends the authors’ March review.

For students, handing academic tasks to AI can reduce learning. Results are more favorable, though still mixed, when students explicitly use it to study. Students also often describe more positive learning experiences with AI, but that satisfaction frequently does not translate into stronger learning outcomes. Some structured support improves well-being or social confidence without consistently improving emotional health.

Nor does an education-specific label guarantee better results. Educational AI tools do not always outperform general-purpose chatbots. Interactive formats, adaptive sequencing and guidance on using a tool show promise. But research on when to provide AI support has produced inconsistent results, leaving the authors unable to identify a clear winning design.

Teacher assistance has its own dividing line

Targeted teacher tasks offer some encouraging findings. Educators given AI-generated drafts or suggestions supplied more and higher-quality student feedback without spending more time. Earlier research also found reductions in lesson-planning time without lower-quality output. Yet one new study found that a teacher-support tool reduced students’ motivation and confidence.

Experience does not settle who benefits, either. Earlier research found that targeted, real-time guidance particularly helped less experienced and lower-rated tutors. A new study instead found that a broader teaching-support chatbot reduced achievement among students of lower-performing teachers. The authors suggest that differences in the tools’ outputs—and the expertise needed to evaluate and adapt them—may help explain the contrast.

Stronger causal evidence, limited classroom reach

Despite the review’s K–12 focus, most of the newly added student-facing research took place in postsecondary settings. Five new student-facing papers studied U.S. K–12 settings, compared with none in the March review. The authors caution against treating findings across different populations, tools, implementation conditions and outcomes as broadly transferable to U.S. schools.

The review’s methodology also distinguishes a large research repository from a smaller body of evidence designed to assess cause and effect. Its selection process narrowed the new material in three stages:

• Of 868 added papers, 201 used randomized controlled trials or quasi-experimental designs to estimate an intervention’s effects.
• An initial language-model screen helped prioritize papers; 97 advanced to human review for relevance and causal rigor.
• Thirty-two met the strong-causal-evidence standard. At least two human researchers review each causal paper discussed in the key findings.

That screening does not make the evidence complete or uniformly peer-reviewed. The repository draws primarily from arXiv preprints, which have not undergone peer review. The authors acknowledge that automated screening can miss relevant causal studies before human review, while the search can overlook journal papers or research distributed elsewhere.