The Evidence Base on AI in K-12: A November 2026 Review

Synthesis: Stanford's AI Hub for Education has doubled its causal evidence base in eight months. The March 2026 review found 20 studies that met What Works Clearinghouse standards among 818 papers; the November 2026 update adds 32 more — 28 student-facing and four educator-facing — for a total of 52, drawn from a repository that now holds 1,686 papers. Five of the new student studies were conducted in U.S. K-12 settings, against none in March. The update does not overturn the March conclusions so much as split them: using AI to produce work still predicts worse unaided performance, using it to explain and guide can predict better, and neither pattern is consistent. The most actionable new evidence is about design and delivery — adaptive sequencing, required mastery, and guidance on how to use the tool — and the most sobering is a randomized study in which giving teachers a broad AI teaching assistant reduced student motivation and confidence.

Stanford SCALE Initiative, AI Hub for Education — March 2026 review and November 2026 update. The update covers papers published October 2025 to June 2026.

What changed

The March review screened 818 repository papers through October 2025 and found 20 strong causal studies, 14 of them student-facing and 8 educator-facing (papers studying both are counted twice). The November update examined the 868 papers added through June 2026: 201 were RCTs or quasi-experimental designs, 97 reached human review, and 32 met the standard. That brings the total to 52 causal papers, 42 student-facing and 12 educator-facing.

Three shifts matter more than the totals:

• U.S. K-12 evidence exists now. Five new student-facing studies were run in U.S. K-12 settings, against none in March, and one new educator-facing study joined them. The authors are explicit that this is a start, not a basis for generalizing across U.S. schools.
• The subject mix moved to science and computing. Papers studying science and technology outcomes grew from 3 to 16, and social-emotional outcomes from 3 to 11, while math and literacy did not grow comparably.
• The evidence is still mostly postsecondary. Among the new student papers, 15 were U.S.-based and 13 international, but most were in college settings; U.S. student papers grew from 3 to 18 overall.

Key Findings: students

1. Using AI to produce work still looks harmful; using it to learn looks better. A Brazilian undergraduate study found AI sped up presentation completion while lowering performance on a surprise conceptual assessment. Canadian beginners who generated complete coding solutions performed well on the task but understood little of the solution. In an adaptive math platform used in U.S. K-12 and postsecondary, students after ChatGPT's release spent less time on text-based problems that could be pasted into a chatbot, did worse on those problems in proctored assessments, and the gap was relative to graph-based problems that are harder to hand over (Barcaui, 2025; Chen et al., 2026; Rismanchian et al., 2026).
2. Learning support is mixed, not reliably positive. Immediate gains appeared for self-directed LLM study, Socratic tutors, multimodal platforms, adaptive practice and personalized support messages, but several well-designed studies found no benefit over static materials or standard practice platforms. Gains that persisted three days to a week were documented in U.S. studies; a U.S. K-12 study found no significant overall improvement on a delayed assessment (Anand & Idan, 2026; Heldreth et al., 2026; Oreopoulos et al., 2026; Rotter et al., 2026; Singh et al., 2026).
3. General-purpose chatbots can help when students use them to learn. U.S. college students who used general-purpose chatbots to study unfamiliar material performed better on later unaided assessments, the opposite of the Bastani and Kosmyna findings in the March review. The authors' candidate explanation is use: explaining concepts versus generating answers. Students with stronger prompting, filtering and verifying skills gained more (Anand & Idan, 2026; Contractor & Reyes, 2026).
4. Positive experience is not evidence of learning. Students reported better experience, lower workload or more useful feedback without corresponding outcome gains in several studies, while rating AI feedback as clearer and more motivating than educator feedback that produced equivalent learning (Gardella et al., 2026; Henze et al., 2026; Meyer et al., 2026; Taneja et al., 2026; Zhao et al., 2026).
5. Design and delivery show promise, but which features matter is unresolved. Adaptive sequencing raised unassisted certification scores by 0.15 SD among 770 Taiwanese high school students, and adding an AI tutor to a math platform helped only when students had to answer three questions correctly in a row before advancing (6,997 U.S. students in grades 6–8). Scaffolding by timing gave inconsistent results: delayed feedback helped code quality in one study and did nothing for writing in another (Chung et al., 2026; Meyer et al., 2026; Morris & Maes, 2026; Oreopoulos et al., 2026).
6. Pedagogical structure is not automatically better. A U.S. study found immediate answers with explanations beat layered hints and encouragement, and a German study found no difference between a general chatbot and an educational tool with added strategy prompts. The content and relevance of the added structure may matter more than its presence (Cao et al., 2026; Fütterer et al., 2026).
7. Motivation, trust and well-being are affected. Providing teachers in Türkiye with an AI teaching-support tool reduced students' intrinsic motivation and confidence even though students were not told the tool was in use. Requiring U.S. students to answer independently before consulting AI lowered their rating of its credibility. Well-being apps improved resilience and social connection over six weeks but not depression, anxiety or stress (Cachia et al., 2025; Islam et al., 2026; Sungu et al., 2026).

Key Findings: educators

1. Time savings are real and so are unintended effects. The earlier English trial found lesson-planning time fell without detected quality loss. The new Türkiye trial found no overall test-score effect but lower student motivation and confidence, and reduced achievement among students of below-median-performing teachers. Teachers used the tool mainly to generate materials they accepted with minimal revision (Roy et al., 2024; Sungu et al., 2026).
2. AI increases the quantity and quality of educator feedback. Teaching assistants given AI drafts provided nearly twice as much feedback, covering more rubric criteria, without spending more time, and their students' revisions scored higher without post-test gains (Lu et al., 2026; Mahinpei et al., 2026).
3. AI-generated student feedback can match educator feedback. Multimodal AI feedback produced similar learning gains to prewritten educator comments and was rated clearer and more motivating; German secondary results showed no significant difference in revisions or later writing (Meyer et al., 2026; Zhao et al., 2026).
4. Progress summaries change behavior more than achievement. K-12 students in the U.S. were more likely to start and complete assignments when teachers received AI summaries, but class-level assignment scores did not improve (Lim et al., 2026).
5. Educator expertise moderates the benefit. Real-time expert-like suggestions helped less experienced tutors most in the earlier U.S. study; a broad teaching-support tool harmed students of lower-performing teachers in the new one. Tool output quality and the expertise needed to evaluate it are the authors' candidates (Wang et al., 2025; Sungu et al., 2026).

Where the March review held, and where it did not

Held: performance gains during use are common and durability is not; a general-purpose chatbot can harm unaided performance; the educator evidence base is thin.

Qualified rather than confirmed: the March claim that tutoring-specific tools consistently outperform general-purpose chatbots. The new comparisons point both ways — the tool that beat general-purpose AI in Bastani et al. withheld answers and carried teacher guidance, while a U.S. tool that revealed answers immediately outperformed layered hints, and a German educational tool with strategy prompts matched a plain chatbot. The update's conclusion is narrower and better supported: effects depend on the substance and instructional relevance of what the tool adds, not on tool category.

Also unresolved: whether AI narrows or widens inequity. The update says it provides little new evidence, and that benefits varied by prior performance in the Sierra Leone trial (larger for higher-baseline students) and by teacher quality in Türkiye.

What this means for practice

• Instructors: distinguish tasks where AI explains from tasks where it produces. The new evidence supports AI for study and feedback on the student's own work, and warns against AI that completes the artifact. Check unaided performance, not practice grades.
• Administrators: evaluate the specific tool in your setting against the outcome you want. Time saved and satisfaction are the easiest things to measure and the least informative; the Türkiye trial is the case where both looked acceptable and student motivation fell.
• Researchers: the field now has U.S. K-12 studies, five of them, and still no long-term follow-up. Delayed, unassisted outcomes and educator-facing designs remain the gaps worth targeting.
• Software developers: adaptive sequencing, mastery gates and guidance on tool use have the best design evidence in this update; adding prompts or a pedagogical wrapper without changing the instructional content has not separated from a general chatbot.

Limitations

The repository is primarily arXiv preprints, screened by an LLM and then by at least two human reviewers using a codebook based on What Works Clearinghouse standards, so inclusion is judged on preprint reporting rather than re-analysis. The authors note that LLM pre-screening can drop a relevant study before human review. Generalization is limited by design: the new studies vary in population, setting, tool and implementation, and the authors warn against reading any single result as what will happen in another school. Most of the new evidence is postsecondary and short-term, and nothing in the update establishes whether gains persist over months or years.

Citation

Walker, S., Fesler, L., Martinez Claeys, JP., Agnew, C., & Loeb, S. (2026). The Evidence Base on AI in K-12: A November 2026 Review. SCALE Initiative, Stanford University. https://doi.org/10.26300/80wn-ca81