Stanford University’s Research in Action summaries detail an expansion in causal research on generative artificial intelligence in K-12 education, alongside new analyses of how U.S. educators use these tools. The updated review increases the evidence base from 20 to 52 experimental and quasi-experimental studies, adding the first five student-facing AI studies conducted in U.S. K-12 schools.
What Happened
According to Stanford University's Research in Action updates, the November 2026 Evidence Base Review examined randomized controlled trials and quasi-experimental studies published between October 2025 and June 2026 in the AI Hub for Education’s Research Repository. The total number of causal papers increased by 32 since the March 2026 review, with student-facing AI research growing faster than educator-facing research. Five of the new student-facing studies took place in U.S. K-12 settings, up from zero in the March review, though researchers caution that the findings are not broadly generalizable across all U.S. schools.
In two accompanying research briefs, Stanford researchers analyzed activity data from approximately 87,000 U.S. educators in the top 5% of users on the MagicSchool AI platform between May 2023 and November 2025. While the platform's multi-purpose chatbot was the most used tool across all grades, workflows varied by school level. Elementary teachers primarily used tools for communication and student support, whereas middle and high school teachers focused on generating instructional materials. A separate analysis of roughly 41,000 middle and high school power users (educators who logged into MagicSchool on 52 or more distinct days) tracked how subject-area teachers used specific tools such as Lesson Plan and Writing Feedback.
To help school systems evaluate student-facing software, Stanford also released a procurement guide with discussion prompts for district leaders attending AI product demonstrations. Instead of providing a formal evaluation rubric, the guide prompts officials to ask vendors what classroom conditions and human supports are required to keep students engaged. Full texts of the individual Stanford briefs and review were only summarized on the captured index page.
The Bigger Picture
Stanford's district procurement guide draws directly from a working paper on AI tutoring engagement, which found that software access alone rarely drives student use. When elementary students received independent access to an AI literacy platform, nearly half never used it, and the rest logged only two to five minutes per week. Pairing students with in-person human tutors who encouraged engagement raised weekly usage by one to four minutes, a 71% to 80% engagement increase, but did not improve reading achievement scores.
Other studies show that combining AI tools with active human instruction produces stronger academic results, consistent with our previous coverage on the role of classroom relationships. A year-long study of seventh-grade math students found that students who received virtual human tutoring alongside classroom AI tutor use finished 0.36 grade levels ahead of peers using AI alone. Similarly, a three-study investigation in urban, low-income middle schools showed that a hybrid human-AI math tutoring model, which cost about $700 per student per year, increased software usage and math proficiency, with lower-achieving students gaining the most.
Recent causal evaluations of student-facing AI show the limits of unassisted chatbots and the need for software guardrails. In a two-year experiment across 18 Tennessee middle schools, giving remedial math students access to Khan Academy's Khanmigo AI tutor improved test scores by 0.06 to 0.08 standard deviations over a school year, matching typical gains from non-AI Khan Academy practice. Even though 96% of students tried Khanmigo, the median student messaged it on only one-third of active practice days and during just 17% of sessions where they made an error.
Without instructional guardrails, generative AI can reduce skill mastery. A high school math field experiment with nearly 1,000 students found that standard ChatGPT access increased practice scores by 48%, but students scored 17% lower on subsequent exams once the AI was removed. When students instead used a restricted AI tutor programmed with teacher-designed hints rather than direct answers, practice scores rose by 127% and the later learning loss was largely eliminated. In another assessment, a causal study of 378 middle school students in grades 5 through 8 found that learning outcomes on AI-free written exams varied by prior achievement and curiosity. The researchers warned that unguided chatbot rollouts can widen existing achievement gaps.
On the educator side, broader academic literature mirrors the Stanford findings on lesson preparation. A systematic review of 79 empirical K-12 studies concluded that Pre-K through 12th-grade teachers primarily use generative AI as a planning aid for instructional design, lesson planning, assessment creation, and student feedback.