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

Synthesis: Stanford Evidence Base: AI in K-12 Education — A 2026 systematic review from the Stanford SCALE Initiative analyzing 818 papers on AI in K-12 education. The central finding is that learning outcomes studies provide strong causal evidence for only 20 of 818 papers, and zero high-quality causal studies examine U.S. K-12 student settings. The evidence that exists reveals a consistent pattern — AI improves performance during use but gains frequently fail to persist or transfer, and general-purpose AI tools can actively harm learning outcomes compared to pedagogically designed alternatives.

Stanford SCALE Initiative, AI Hub for Education — Published 2026. Analysis of repository spanning through October 2025.

Key Findings

The Evidence Gap. Of 818 papers in the AI Hub Research Repository, only 20 met What Works Clearinghouse (2025) standards for strong causal inference (RCTs or quasi-experimental designs). Zero high-quality causal studies examine U.S. K-12 student settings; very few exist for U.S. K-12 educators. Most causal research is international, conducted in postsecondary settings, short-term (often single 20-minute sessions), and focused on immediate outcomes. The repository grew from 28 relevant papers in January 2023 to over 800 by October 2025, but methodological rigor has not kept pace with volume.

Immediate Gains, Uncertain Transfer. AI significantly improves performance while students use it — math proofs, programming, economics exams, physics, and argumentative writing all show gains during AI-supported practice. However, effects are mixed or negative when AI is removed. Bastani et al. (2025) found high schoolers using a general-purpose chatbot for math practice performed ~17% worse on closed-book final exams than peers with no AI access, despite higher practice grades. Chen et al. (2025) found Large Language Models (LLMs)-Tutor improved homework scores but did not improve unassisted exam scores. Lehmann et al. (2025) found general-purpose AI for programming increased topics covered but harmed understanding and widened achievement gaps for low-Prior Knowledge students. Kosmyna et al. (2025) found AI essay assistance led to 83% of participants failing to recall a quote from their own essay, versus 11% for non-AI users. This pattern — performance boost during use, learning loss after removal — is the central empirical finding of the review and directly implicates Transfer of Learning as the most critical open question in AI education research.

Easier Doesn't Mean Better. Students consistently report greater enjoyment and reduced cognitive burden when using AI tools. However, reduced effort can undermine deeper learning. Kreijkes et al. (2026) found retention improved only when AI use was paired with traditional strategies like note-taking. Stadler et al. (2024) found general-purpose AI reduced cognitive load but produced lower-quality reasoning and argumentation compared to traditional search. This aligns with Desirable Difficulties research: making practice easier often harms long-term retention and transfer, even when it feels better in the moment.

Pedagogical Design Matters. The most actionable finding: tutoring-specific tools consistently outperform general-purpose chatbots. Bastani et al. found that a tutoring-specific chatbot with pedagogical Guardrails (hints, step-by-step reasoning, refusal to give direct answers) mitigated the exam score drop, while general-purpose GPT Base caused it. This suggests that AI Tutoring effectiveness depends critically on pedagogical design, not just model capability. The review interprets findings through a learning science framework spanning Sociocultural Learning (general-purpose AI may operate outside the ZPD by doing work for students), the expertise reversal effect (novices need guidance, experts need independence), and Metacognition (AI completing tasks reduces opportunities for students to monitor their own understanding).

Educator Evidence. While the student-focused causal evidence is thin, the educator evidence base is even sparser. Very few high-quality studies examine how AI affects teacher practice, workload, or professional development — a gap that is particularly concerning given the rapid push to deploy AI tools in classrooms and the documented risks of AI harming teaching quality.

Tutoring-Specific vs. General-Purpose AI

The single most consistent finding across the review is that pedagogically designed, tutoring-specific AI consistently outperforms general-purpose chatbots (e.g., ChatGPT, Gemini) on durable learning outcomes. The mechanism is preservation of learning work: tutoring-specific tools retain productive struggle via graduated hints, avoid suppressing Metacognition, and calibrate to the learner's zone of proximal development — where general-purpose AI that simply completes tasks short-circuits all three.

Why this happens (learning-science mechanisms): (1) Desirable difficulties — general-purpose AI removes productive struggle; tutoring tools preserve it via graduated hints. (2) Germane load — effective learning requires processing that feels effortful, which general AI short-circuits (cognitive load / offloading). (3) Metacognition suppression — when AI completes reasoning, students lose practice in monitoring their own understanding. (4) Expertise reversal — novices need scaffolding, not answers; general AI gives answers while tutoring AI gives scaffolds.

Caveats: The causal comparison base is tiny (most studies are single-condition AI-access vs. no-access, not head-to-head tutoring vs. general); "tutoring-specific" is not yet a standardized design category; and long-term transfer data (months or years out) is essentially absent.

What this means for practice

• Instructors. Choose tutoring-specific tools with pedagogical guardrails — graduated hints, step-by-step reasoning, refusal to give direct answers — over general-purpose chatbots for practice work, since the general-purpose chatbot in Bastani et al. produced ~17% worse closed-book exam performance while the guarded tutor mitigated that drop.
• Check unassisted performance rather than practice grades: Chen et al. found LLM-Tutor raised homework scores without raising unassisted exam scores, so schedule closed-book checks before treating an improvement as learning.
• Pair AI assistance with established study strategies rather than substituting for them — retention improved only when AI use was combined with traditional strategies such as note-taking (Kreijkes et al., 2026).
• Administrators. Do not justify district-wide rollout with this evidence base: zero high-quality causal studies cover U.S. K-12 student settings, and the 14 causal studies of students are postsecondary or international high school, frequently single 20-minute sessions.
• Researchers. Build delayed post-tests and unassisted outcome measures into every AI-in-classroom study, and target authentic U.S. K-12 settings and educator impact, since the review identifies those designs as missing rather than merely underrepresented.

Limitations

• The review's repository is 87% arXiv preprints and its quality screening pairs an LLM pre-screen with human review against What Works Clearinghouse standards, so the 20 "strong causal" studies it identifies are judged on preprint reporting rather than re-analyzed primary data.
• Search terms were limited to "education" combined with "AI" or "artificial intelligence," and pre-LLM rule-based intelligent tutoring systems were excluded, which bounds what the 818-paper repository can represent.
• The causal studies it does contain are short-term and largely postsecondary or international, so the review cannot speak to year-long use in U.S. K-12 classrooms.
• The saved full text is a truncated extraction; the source itself notes that the educator findings, policy implications, and recommendations were cut off, so those sections are not represented on this page.

Citation

Stanford SCALE Initiative, AI Hub for Education. (2026). The Evidence Base on AI in K-12: A 2026 Review. Stanford University.