Skip to content
Home » The Headlines Said K-12 AI Research. The Evidence Mostly Wasn’t.

The Headlines Said K-12 AI Research. The Evidence Mostly Wasn’t.

97% of Stanford’s AI-in-Education Papers Missed the Cut — LXD Research
AI & Evidence · November 2026

When Stanford’s SCALE Initiative released its November 2026 review of the evidence on AI in K-12, the coverage arrived within days. Chalkbeat, GovTech, and a long tail of newsletters and blogs summarized the findings, and most of them led with the same idea: when students let AI do the thinking, they learn less.

We read the appendix before we read the headlines, and what we found there changes how we read the coverage. Most of the evidence behind those headlines did not come from U.S. K-12 classrooms. This post looks at who was actually studied, why that matters for children learning to read, and who is reviewing the 97 percent of papers that never made Stanford’s cut.

What did the Stanford review actually study?

The SCALE team searched a repository of 1,686 papers and identified 52 with strong causal evidence, about 3 percent. That is up from 20 in their March review, which is real progress. The report itself is careful. It says the findings should not be treated as broadly generalizable across U.S. K-12 settings, and that much of the new research comes from postsecondary students and focuses on science and computer science rather than literacy or math.

Papers in repository 1,686 Through June 2026
Strong causal evidence 52 Up from 20 in the March 2026 review
Share that qualified 3.1% 52 of 1,686 papers. A strict filter, by design
Literacy papers 10 Of the 52 causal papers, per Ozernov-Palchik’s IDA slide

Who was actually in the studies behind the headlines?

The coverage was less careful than the report. Most of it carried the K-12 label from the report’s title straight into the headlines. But the report shows that most of the new papers are in postsecondary settings rather than K-12, and among the 28 new student-facing studies, only five were conducted in U.S. K-12 settings. The rest involved college students, adults recruited online, or students in other countries, including Germany, Taiwan, Türkiye, and Sierra Leone.

Ola Ozernov-Palchik, founder and director of Boston University’s Evidence-Based AI in Learning (EVAL) Industry Collaborative, laid this out in a slide at her presentation at the International Dyslexia Association (IDA) conference on November 12, 2026. Follow the 52 causal papers down to the classrooms where most of our readers work:

Does cognitive offloading apply to children learning to read?

That matters for the finding that got the most attention. Cognitive offloading, where students rely on AI to do their thinking and then perform worse on their own, is real in the studies that tested it. But look at who was tested.

Business undergraduates 85 Brazilian students who prepared a presentation with or without ChatGPT
Programming students 24 Canadian undergraduates learning to code
Novice programmers 22 U.S. university students working with Copilot

These are adults. None of this describes a seven-year-old learning to read, or a sixth grader learning fractions, and none of it describes what U.S. edtech companies are building for young children.

What this means: The headline finding comes mostly from adults doing tasks where AI can simply do the work. It tells us little yet about how children who are learning to read, write, and do math respond to AI built for them.

Natural and adaptive does not mean effective

When we built Lexia Core5, we wanted it to respond to every student, but the technology had limits. That is the experience many product teams remember, and it is a long way from a college student pasting a prompt into a chatbot.

Then
One actress in a sound studio

We could scaffold, but not truly adapt. All the feedback was one actress in a sound studio, with the same scripts over and over.

Now, with AI
Infinite combinations

When a student makes an error, a product can bring that item back two days later, or two weeks later, to support retrieval practice. The feedback sounds natural instead of pre-scripted.

But natural and adaptive does not mean effective. Product developers will say they are moving too quickly to study, and we understand that, because they are building while they scale. That is exactly why documentation matters more. Release notes can include the user research behind each change and why the team thinks it will help, and then an independent researcher can follow up to see whether the change improved outcomes.

The loop the field needs: user input, documentation, follow-up research, and then leader decisions.

Where are the U.S. K-12 literacy studies?

Here is the part we could not stop thinking about. Of the 52 causal papers, only two were U.S. K-12 literacy studies, and they are our evaluations of Coursemojo, an AI-supported reading and writing program. Both used quasi-experimental designs with sixth graders, one with 2,045 students across 12 schools and one with 2,771 students across 13 schools.

U.S. K-12 literacy studies 2 of 52 Both are LXD Research evaluations of Coursemojo
Tennessee TCAP, ELA +8 Points higher, about, for Coursemojo students (2,045 sixth graders, 12 schools)
Texas STAAR, reading +10.5 Points higher for Coursemojo students (2,771 sixth graders, 13 schools)

We are proud of that work, and we want to be honest about the mixed NWEA MAP result in the second study, because that is how real evidence looks. But we do not want to be the answer to the question of where the literacy evidence is. Two studies from one research group is a starting point, not a field.

Literacy is the foundation for everything else a student does in school. Districts are adopting AI reading and writing tools right now, and many are making those decisions on headlines about college students. The children who stand to gain or lose the most are the ones still learning to read, and they are the ones the evidence barely covers.

There is also a structural problem. Products are shipping faster than studies can be designed, and many developers say they are moving too quickly to study. But the speed is the reason evidence matters more, not less. When the state of Arizona told reading companies that they needed top-tier research or could no longer sell in the state, it signaled where the market is heading. Companies that wait until a request for proposals arrives will find they cannot build evidence in a month.

Who is looking at the other 97 percent?

The Stanford review applies a strict filter, and it has to. Causal claims need that rigor. But a strict filter only speaks to the top slice. Of the 1,686 papers in the repository, 52 met the bar, which is about 3 percent. The same pattern shows up in product-level evidence. Evidence for ESSA, which is K-12 only and applies review standards similar to the What Works Clearinghouse, lists roughly 170 products across reading, math, and science. The EdTech Index, a K-12-focused product list, includes more than 2,400.

Evidence for ESSA listings (reading, math, science)~170
Products on the EdTech Index2,473

Product counts are as of November 18, 2026. The Evidence for ESSA figure adds its reading (111), math (54), and science (5) listings.

That leaves most of the field outside the strictest reviews. Someone needs to look at what sits below the bar: the logic models, the correlational studies, the smaller and retrospective designs. They answer different questions, like what dosage a district needs and how a product gets implemented, and they are exactly the groundwork a later trial depends on. That is where graded validation fits, from Evidence for ESSA, Digital Promise, EduEvidence, and LXD Research badges. The strict filter protects against overclaiming. The tiers keep the rest of the evidence from being invisible.

What should the path to strong evidence look like?

The fix is not to demand a randomized trial for every product on day one. It is to build evidence in a sequence that makes each step more useful than the last.

StepWhat it tells a district
ESSA Tier IV (Demonstrates a Rationale)Why the product should work, what implementation should look like, and what outcomes to expect.
ESSA Tier IIIHow usage relates to outcomes in real classrooms, with statistical controls, and what dosage a district needs to see similar results.
ESSA Tiers II and IWhether outcomes differ from a comparison group, matched or randomized, once the company is confident in its measures and implementation.

Skipping the foundation is the most common way companies waste money. Jumping to a top-tier study without understanding implementation factors, or the dosage needed to see gains on the assessments that will pick them up, is a real risk.

Get the early decisions right. Design sets the ceiling, and execution decides whether a study reaches it. Studies involve dozens, if not hundreds, of decisions, and early ones like recruitment can keep a study from earning the rating it should. Getting them wrong can mean an ineligible study and paying for the same study twice.

Make it independent. Vendors have an obvious interest in favorable results. Independent researchers, along with third-party review through programs like Evidence for ESSA, Digital Promise, and EduEvidence, give districts and states a check that does not depend on the company’s own account.

Document as you go. For AI products that change constantly, release notes can capture the user research behind each change and why the team expects it to help, so an independent researcher can follow up on whether outcomes improved.

What should district leaders ask about any study?

When a vendor makes a claim, look at the study behind it, at Tier III or above. Who was in it? How many students? Was there a comparison group, or was it compared to a norm sample? What assessments measured the impact? How long did it last? And how closely do those students resemble the ones you serve every day?

A badge is a floor, not a ranking: A badge tells you a study cleared a minimum bar. It does not tell you how much stronger it is. A Tier III rating could reflect a correlational study, a randomized trial that was too small, or something in between. The only way to know is to read the study the researchers wrote.

For vendors, outcomes-based contracting is one place where this question becomes concrete. Our primers cover what outcomes-based contracting is and how to get ready for it.

What should readers do with this?

If you write about this research, or read coverage of it, check who was studied before you apply a finding to children. If you lead a district, ask vendors for studies with students like yours. If you build products for K-12, start building independent evidence now, because the companies and investors who fund it will be the ones districts trust. The evidence is growing, but it is still in its infancy.

If you are building an AI product for K-12 and want to know where to start, LXD Research can help you sequence your evidence, from a research-based logic model through ESSA-aligned studies. Our published studies show what that looks like in practice.

Source: Walker, S., Fesler, L., Martinez Claeys, J.P., Agnew, C., & Loeb, S. (2026). The Evidence Base on AI in K-12: A November 2026 Review. SCALE Initiative at Stanford University. Download the report (PDF) or cite it by DOI. The funnel figures (52 causal, 10 literacy, 6 K-12, 2 U.S.) come from a slide in Ola Ozernov-Palchik’s presentation at the International Dyslexia Association (IDA) conference on November 12, 2026. Learn more about the EVAL Collaborative.


Not Sure What Level of Evidence Your Product Has?

Get a research-path recommendation for your product, timeline, and procurement goals.

Book a Free Conversation See How ESSA Certification Works

Or email us at info@lxdresearch.com