Weekly roundup, November 16, 2026 - Stanford's bigger AI evidence review, AI-made lessons that all look alike, and districts pulling back
By Matthew Rhoads, Ed.D.
This roundup covers news and research released between November 8 and November 14, 2026. Each item shows its own publication date, and one paper was written earlier than it was posted. Evidence levels below are provisional: they were applied by hand from our published rules, and none of these items has been through our two-grader review yet. Levels describe study design and who took part. They are not endorsements or official ratings.
The short version
Stanford's SCALE Initiative updated its review of strong causal studies on generative AI in education, and the count grew from 20 to 52 papers. Its main pattern: AI looks better for learning when students use it to understand material than when it does the work for them, and even then results are mixed. A Stanford technical study found that AI-written math problems and essay feedback were much more alike than what teachers write. Two smaller studies looked at how students actually talk to AI tools and how the wording of AI rules changes behavior. Outside AI, a large school-randomized trial in Bangladesh found that giving teachers clear reports on which students and topics needed help raised math and reading scores at low cost. In the news, a teachers union signed a binding student-data agreement with an AI company, and several districts tightened or debated limits on AI.
Research
Stanford's updated review: more causal studies, still mostly not U.S. K-12
The Evidence Base on AI in K-12: A November 2026 Review, Walker, Fesler, Martinez Claeys, Agnew and Loeb, SCALE Initiative at Stanford University, November 10, 2026.
Provisional Level 3: Controlled comparison (a structured review that includes only randomized and quasi-experimental studies; our rules cap reviews here until a person checks the included studies). Labels: Report (not peer reviewed). Most included studies are preprints.
What they did. The team screened 868 papers added to SCALE's research repository that were published between October 2025 and June 2026. 201 were randomized or quasi-experimental, 97 went to human review, and 32 met their standard for strong causal evidence, which is based on the What Works Clearinghouse handbook. Added to the 20 papers in their March 2026 review, the evidence base is now 52 papers. The review summarizes findings in words. It does not pool effect sizes.
What they found. AI appears more likely to support learning when students use it to understand new material than when they use it to produce work, but neither kind of use consistently improves learning. Students often report better experiences with AI, and those experiences often do not translate into stronger learning. Education-specific AI tools did not always beat general-purpose chatbots. On the teacher side, AI drafts helped educators give more and better feedback without taking longer. But in one new trial in Türkiye, giving teachers a broad AI teaching assistant lowered students' motivation and confidence, with no significant change in test scores overall and lower achievement among students of below-median-performing teachers. Five of the new student-facing studies were in U.S. K-12 schools, up from none in March.
Limitations. The authors say the repository is mostly arXiv preprints, that automated screening may miss relevant studies, and that the findings should not be treated as broadly generalizable across U.S. K-12 settings. Most of the new studies are in colleges, and many measure science or computer science outcomes.
What it means in practice. Before buying, ask how the tool will be used: to help students understand and practice, or to produce work for them. If teachers will use AI to make materials, build in a step where they review and adapt what it produces, and check student outcomes, not just time saved.
AI-written math problems and essay feedback were far less varied than teachers' work
LLMs and the Homogenization of Teaching, Tan, Nam, Subramonyam and Demszky (Stanford University), arXiv, posted November 10, 2026. The paper says it is published as a conference paper at COLM 2026.
Technical test (T). No students took part. Our tracker shows the same provisional rules label. It tells us about the technology, not about student learning.
What they did. The team compared AI output with human-written examples in two tasks: 1,477 practice problems from the open Illustrative Mathematics curriculum (grade 6 through Algebra II), and 3,021 feedback comments that 61 experienced U.S. middle and high school English teachers wrote on student essays. They tested frontier and open-weight models with plain prompts, two prompting methods meant to add variety, and extra model training.
What they found. With plain prompts, model outputs were 78% higher on a word-overlap similarity measure (Self-BLEU) and 51% higher on a meaning-based similarity measure (embeddings) than the human-written sets, and they represented a narrower range of instructional forms. The variety methods and extra training added variation mostly through surface features, without broadening that range. The authors conclude that models inherit and amplify the most normative routines of human instruction.
Limitations. It measures the materials, not what students learn from them. The results come from two tasks and the specific models and prompts tested.
What it means in practice. If a team builds a unit from AI-generated problems or feedback, expect sameness. A quick fix is to check a batch against your curriculum's range of task types and rewrite the gaps, rather than asking the model to "be more creative."
How students talked to AI tools was associated with what they did next, more than counts of use would show
From Dialogue to Task: How Cognitive Engagement With Generative Artificial Intelligence Relates to Post-Dialogue Activity and Task Outcomes, Zhuang, Bakharia, Gašević and Khosravi (University of Queensland and Monash University), arXiv, posted November 8, 2026.
Provisional Level 2: Correlational. Labels: Preprint (under review at a journal), Learners: Higher-ed students. Our tracker shows the same provisional rules label.
What they did. The researchers analyzed platform logs from 1,681 students in 40 university courses who used AI helpers built into RiPPLE, a learning platform developed by one of the authors, between June 2025 and June 2026. They coded 829 student-AI conversations by depth of thinking, from passive to interactive, and linked them to what students did next and to the quality of their work.
What they found. Most conversations were passive or active rather than constructive or interactive. When students created study resources, conversations where they wrote their own messages were followed by more returns to the task than template-only requests. Deeper engagement in the conversation was associated with higher peer-rated quality of the finished resource. In practice questions, talking with the AI after a wrong answer was associated with fewer further attempts, but higher accuracy on the next try among students who kept going.
Limitations. Students chose when and how to use the AI, so these are associations, not causes. Work quality and practice scores were measured while students had AI help, so the study does not show gains in independent learning. Some comparison groups were small.
What it means in practice. Counting logins or messages says little. If you evaluate an AI tool, look at what students type and what they do afterward.
The wording of AI rules may shape whether college students disclose their use
Wording Matters: Understanding How GenAI Policy Rules Shape Teachers' and Students' Workflows and Perceptions, Mehreen, Yuan, Nithyanand and Hoque (University of Iowa), arXiv, posted November 8, 2026.
Provisional Level 1: Descriptive or exploratory. Labels: Preprint, Learners: Higher-ed students and instructors, Outcome: Perception or attitude. Our tracker shows the same provisional rules label.
What they did. The team analyzed 694 AI policy and guidance documents from 101 U.S. universities, identified nine recurring rules, and interviewed 20 students and 20 instructors about them.
What they found. Rules that frame AI use mainly as possible dishonesty could lead students to hide their use out of fear of scrutiny. Rules that make students accountable for the accuracy and ethics of AI output could encourage them to take ownership and be more open. Students described uncertainty about acceptable use, fear of penalties and stigma. Instructors described extra approval work and no reliable way to verify compliance.
Limitations. It is an interview study of 40 people, recruited through outreach and listservs, and the universities sampled are highly ranked national universities. It describes perceptions, not measured behavior.
What it means in practice. If your school is writing AI rules, test the wording with a few students before adopting it. Rules about owning and checking AI output may get you more honest disclosure than rules about cheating alone.
Research beyond AI
Clear reports on which students and topics needed help raised math and reading scores
Enhancing Teacher Productivity with Information on Students and Skills: Experimental Evidence from Bangladesh, Djaker, Ganimian, Angrist and Sabarwal, EdWorkingPapers (Annenberg Institute at Brown University), November 2026. The paper's title page is dated September 16, 2026.
Provisional Level 3: Controlled comparison. Labels: Preprint (working paper, not peer reviewed), Learners: K-12 students (grade 6) and their math teachers, Outcome: Learning or achievement. The paper does not mention a preregistration.
What they did. Researchers randomly assigned 312 secondary schools in three divisions of Bangladesh to a program or to business as usual, during the 2022 school year. In program schools, grade-6 students took two short math tests, and their teachers got reports flagging the topics and students that needed support, plus an answer key explaining common errors.
What they found. At the start, teachers could not estimate their students' scores or rankings well, and they were about twice as likely to interact with high achievers as with low achievers. On independent end-of-year tests, the program raised math achievement by 0.23 standard deviations and reading by 0.17 standard deviations, with similar gains across the achievement range. Student self-reports of three social-emotional skills rose by 0.10 to 0.14 standard deviations. The authors estimate the cost at $5.06 per student.
Limitations. It is a working paper. About 77% of students who took part at baseline also took part at the end, and dropout rates were nearly identical in both groups. Program students started somewhat lower in math. Teachers in this setting had little usable information about their students' performance, so effects could be smaller where teachers already get regular reports.
What it means in practice. Data only helps when it tells a teacher what to do next. A short report that names the topics and students to revisit may matter more than another dashboard, whether or not AI is involved.
In the news
A teachers union and Anthropic sign a binding student-data agreement
AFT, UFT and Anthropic Sign Binding AI Safety and Privacy Agreement for Schools, American Federation of Teachers, November 10, 2026.
Commentary (C). An agreement announcement, not a study.
The AFT and the United Federation of Teachers said Anthropic signed their AI Safety and Privacy Standard for Schools, which they describe as legally binding and as applying automatically to Anthropic's existing agreements with districts using Claude for Teachers. The standard says student and educator information may be used only for authorized educational purposes and never sold or used for targeted advertising, that schools control how long data is kept and when it is deleted, that educators remain responsible for instructional decisions, and that schools and families get plain-language information about what the products collect and where they fall short. AFT President Randi Weingarten said the union still wants the standard written into law.
What it means in practice. Districts can ask any AI vendor whether it will sign terms like these, and compare them with their current contracts. The release does not describe how compliance will be checked.
Districts tighten, debate and push back on school AI
Seattle schools dial back tech, Axios Seattle, November 9, 2026. South Portland schools vote down AI moratorium, Portland Press Herald, November 11, 2026. Protests in Chicago After Google Warns Schools About Disabling AI Feature, Chalkbeat via The 74, November 14, 2026.
Commentary (C). News reports, not studies.
Seattle Public Schools announced it will block Gemini and other unapproved AI tools, including Google Search AI summaries, for K-12 students, while keeping MagicSchool AI approved for students and staff. In Maine, the South Portland school board voted down a proposed pause on generative AI for kindergarten through grade 8 and is waiting for its AI steering committee's recommendations in January. In Chicago, about 30 parents, educators and students protested outside Google's office. Chalkbeat reports that Google sparked an uproar this month when it said schools that ban AI won't be able to use its search, and that Chicago Public Schools said it is exploring ways to suppress the AI overview at the top of Google searches, or will look at other search options.
What it means in practice. Districts are making different choices. Whatever yours decides, write down which AI features are on for students, who approved them, and why, so the decision can be revisited as the evidence above grows.
How we rated these
Every item links to its original source. Levels are provisional and were applied by hand from our published rules. They will be updated if our two-grader review changes them. Read our method.
Get Navigating Education free
You're subscribing to Navigating Education, Dr. Matt Rhoads' free weekly newsletter on the science of learning, AI and practical teaching. It goes out Tuesdays. Each issue includes an Evidence Signal section.