Industries / Government, education and community / Education / EdTech
Evidence of what works
Testing whether teaching software and devices actually raise learning, and building the rules, reviews and registries that schools use to judge the claims.
7
Milestones achieved, from 1980 to 2025.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1980
Adding up computer studies
James Kulik, Chen-Lin Kulik and Peter Cohen combined the results of 59 independent evaluations of computer-based college teaching by meta-analysis, a method for pooling many studies into one estimate. They found small but significant gains in course achievement and a substantial cut in the time instruction took.
LessonIn 1983 Richard Clark reviewed such meta-analyses and argued that the gains could come from untested differences in teaching method and from novelty, not from the computer itself.
Credit
James A. Kulik, Chen-Lin C. Kulik and Peter A. Cohen
2002
A federal clearinghouse opens
The US Department of Education's Institute of Education Sciences set up the What Works Clearinghouse in 2002 to review studies of education programs, products and practices against fixed standards and report which show effects. Its public data files list the reviewed studies with their ratings, including studies that do not meet the standards.
Credit
Institute of Education Sciences; contractors including Mathematica and Instructional Research Group
2007
Sixteen products, no difference
In a study mandated by Congress, teachers in 132 schools in 33 US districts were randomly assigned, within each school, to use one of 16 reading or math software products or not. Test scores were not significantly higher in classrooms using the software in any of the four groups studied: first-grade reading, fourth-grade reading, sixth-grade math and algebra.
LessonA 2009 follow-up into a second year of use again found no significant difference overall, and only one product, LeapTrack in fourth grade, showed a significant gain.
Credit
Mark Dynarski, Roberto Agodini and colleagues, for the National Center for Education Evaluation and Regional Assistance
2015
Evidence levels written into law
The Every Student Succeeds Act, signed on 10 December 2015, defines "evidence-based" for US schools at four levels: strong evidence from at least one well-designed experiment, moderate evidence from a quasi-experiment (a comparison of groups not assigned by chance), promising evidence from a correlational study with controls for selection bias, and a fourth level for work with a research-based rationale that is still being studied. Activities paid for from the law's school improvement funds must meet one of the first three.
Credit
US Congress
2017
Experiments sorted by type
Four economists reviewed randomized trials and regression discontinuity studies, which compare people just above and just below a cutoff, of education technology in developed countries. Of 29 randomized trials of computer-assisted learning software, 20 found positive effects, 15 of them in math, eight found none and one found a negative effect; programs that only handed out computers raised computer use and skills, with mixed results for learning.
Credit
Maya Escueta, Vincent Quan, Andre Joshua Nickow and Philip Oreopoulos (NBER working paper, 2017; Journal of Economic Literature, 2020)
2023
UNESCO finds little impartial evidence
UNESCO's Global Education Monitoring Report on technology found "little robust evidence on digital technology's added value in education", noted that products change every 36 months on average, and said "a lot of the evidence comes from those trying to sell it". It cited a UK review in which 7% of education technology companies had run randomized trials, and a survey in 17 US states in which 11% of teachers and administrators asked for peer-reviewed evidence before adopting a product.
Credit
Global Education Monitoring Report team, hosted by UNESCO (director Manos Antoninis)
2025
Clearinghouse reviews stop
In February 2025 the US Department of Education canceled dozens of research contracts, among them one to support the outside evaluators who do research for the What Works Clearinghouse. In a June 2025 court filing it said it would restart the clearinghouse's website and reviewer training but not the contracts that produce new reviews.
LessonA clearinghouse run through contracts stopped producing new reviews when the contracts were canceled, though its standards and past ratings stayed online.
Credit
US Department of Education; reported by Chalkbeat and The Hechinger Report
Watch for
AI in clearinghouse reviews
The What Works Clearinghouse's November 2025 review protocol lets its teams use AI tools such as large language models to search the literature, screen studies, extract data and apply the evidence standards, with trained human reviewers still conducting every review. It also pilots a two-reviewer process in place of the usual three, to test whether reviews can go faster "while maintaining quality".
Achieved whenThe clearinghouse or an outside team publishes how often AI-assisted reviews reach the same rating as the standard process.
Claimed by
What Works Clearinghouse, Institute of Education Sciences, November 2025
A testbed for school software
England's Department for Education said in January 2026 that it would spend an additional £23 million to turn its EdTech Impact Testbed pilot into a four-year program that would "generate robust evidence" on edtech products, including AI tools. The Education Endowment Foundation, which funds and runs school trials in England, says the program launched in May 2026 with up to 100 rapid evaluations, to be followed by up to 15 randomized trials or other robust evaluations of the most promising tools.
Achieved whenThe first trial reports are published alongside their protocols, showing which products did and did not raise outcomes.
contractsfinder.service.gov.ukeducationendowmentfoundation.org.uk
Claimed by
Department for Education and Education Endowment Foundation, January and May 2026
Open
01
Rate studies like the clearinghouse
An AI reviewer that reads a published study of an education program and applies the What Works Clearinghouse's standards. It extracts the design, the sample sizes, the attrition (the share of the sample lost before outcomes were measured) and the baseline equivalence (whether the groups started out alike), then returns a rating with its reasons. Test it against ratings the clearinghouse has already published.
The briefCredit · unclaimed02
Check the evidence claims
A public table of the evidence claims that widely used school software makes on its makers' websites, such as "ESSA Tier 2" or "proven to raise scores", each matched to the study it cites and checked against the law's evidence levels by an AI model and two human reviewers. Each maker sees the draft finding and can reply before it is published.
The briefCredit · unclaimed03
Registered plans versus reports
A trial-by-trial comparison of what education technology trials said they would measure before they started and what their final reports measured. Sources are study registries such as the Registry of Efficacy and Effectiveness Studies (REES), run by the Society for Research on Educational Effectiveness, and the protocols and analysis plans that the Education Endowment Foundation posts for its trials. Flag every primary outcome that was dropped, added or changed without explanation.
The briefCredit · unclaimed
Open 01 · the brief
Rate studies like the clearinghouse
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An AI reviewer that reads a published study of an education program and applies the What Works Clearinghouse's standards. It extracts the design, the sample sizes, the attrition (the share of the sample lost before outcomes were measured) and the baseline equivalence (whether the groups started out alike), then returns a rating with its reasons. Test it against ratings the clearinghouse has already published.
Why it is still open
The clearinghouse's November 2025 protocol allows AI to help apply its standards but reports no agreement figures, and Stanford's 2026 review of research on AI in schools used a language model only as a first screen before researchers reviewed the papers. The nearest public project found, a June 2026 hobby pipeline on GitHub that sorts studies into ESSA levels, reports results on six test studies. In October 2026 the Institute of Education Sciences asked the market for a commercial replacement for the clearinghouse's online review tool.
Built with what exists
The clearinghouse's public study files, which include studies that meet and do not meet its standards, with a February 2025 copy kept on GitHub; full texts from ERIC, the federal library of education research; an AI model such as Claude or Gemini given the clearinghouse's Version 5.0 handbook; Gemini Decision 1 to answer the handbook's fixed-answer questions for each study in one pass (see the Note "A thousand questions, one pass").
First step
Download the clearinghouse's study file, pick 30 rated studies whose full texts are free on ERIC, and compare one model's ratings with the published ones.
Done looks like
On 300 studies the clearinghouse has rated, spread evenly across its rating categories, the AI's rating matches the published rating in at least 85% of cases, with every disagreement listed and the code and prompts published.
Open 02 · the brief
Check the evidence claims
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of the evidence claims that widely used school software makes on its makers' websites, such as "ESSA Tier 2" or "proven to raise scores", each matched to the study it cites and checked against the law's evidence levels by an AI model and two human reviewers. Each maker sees the draft finding and can reply before it is published.
Why it is still open
The checks found report totals or cover few products. Instructure, which sells evidence services, and the nonprofit InnovateEDU judged 150 commonly used classroom tools in March 2026 and found 21% had evidence above ESSA's lowest level. A LearnPlatform review commissioned by the Jacobs Foundation found that 21% of 53 K-12 companies in the foundation's portfolio had a study eligible for an ESSA level. England's EdTech Evidence Board pilot reviewed the evidence portfolios of 12 products and found much of it was internal surveys and correlational studies, with four of the six experimental studies submitted over nine years old. This research found no public, product-by-product comparison of makers' claims with the studies behind them.
govtech.comjacobsfoundation.orgassets.publishing.service.gov.uk
Built with what exists
A web agent to collect claims from product websites (see the Note "Deleting 80% to start over"); ERIC and journal pages for the cited studies; the statute's definition of "evidence-based" and the clearinghouse handbook; an AI model to draft each check; a shared spreadsheet for the two reviewers' decisions.
First step
Collect the evidence pages of 20 early-reading products and check their cited studies by hand to settle the categories.
Done looks like
Claims from 100 widely used K-12 products checked, each linked to its cited study and the evidence level that study supports, with the two reviewers agreeing on at least 90% of products before discussion, published with makers' replies and checked again a year later.
Open 03 · the brief
Registered plans versus reports
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A trial-by-trial comparison of what education technology trials said they would measure before they started and what their final reports measured. Sources are study registries such as the Registry of Efficacy and Effectiveness Studies (REES), run by the Society for Research on Educational Effectiveness, and the protocols and analysis plans that the Education Endowment Foundation posts for its trials. Flag every primary outcome that was dropped, added or changed without explanation.
Why it is still open
In education, missing outcomes have been studied by comparing dissertations with the journal articles made from them. Pigott and colleagues found in 2013 that nonsignificant outcomes were 30% more likely than significant ones to be left out of the published version, and a 2026 replication in special education found significant outcomes 1.48 times as likely to be published. Comparisons of registered plans with reports are established in medicine, where a Cochrane review gathered 16 such studies and a 2026 pipeline screens clinical-trial registrations against publications automatically. This research found none for education technology trials.
Built with what exists
REES, whose entries are public and, by its own guide, searchable and exportable (guide); EEF project pages, which post the original and amended evaluation protocols and the statistical analysis plan, as for its trial of Lexia reading software (example); an AI model that compares long documents (see the Note "Seven more flagged"); the clinical-trial pipeline's matching method as a model.
First step
Take five EEF trials of digital products, line up each protocol's primary outcome against its final report by hand, and record the kinds of change found.
Done looks like
Every EEF-funded trial of a digital product that has both a published protocol and a final report, plus 50 REES-registered edtech studies, compared; each flag confirmed by two people; the share of trials with an unexplained change to a primary outcome published with the full list.