A 2015 federal law sorts vendor evidence claims into four tiers

The classroom note

When a Title I coordinator is asked to justify buying a reading intervention, or increasingly an AI writing tool, the funding conversation runs through one piece of federal law. The Every Student Succeeds Act, a statute signed on 10 December 2015, replaced No Child Left Behind and wrote a definition of ‘evidence-based’ into the Elementary and Secondary Education Act. That definition, at section 8101(21)(A), does not say yes or no to a product. It sorts evidence into four tiers, and grant programmes, including much of Title I improvement spending, can require a district to show which tier its purchase meets before public money changes hands.

What the evidence says

The tiers are precise about study design, not about subject matter. Strong evidence requires at least one well-designed, well-implemented experimental study, ordinarily a randomised controlled trial. Moderate evidence allows a quasi-experimental study instead. Promising evidence permits a correlational study that statistically controls for selection bias. A fourth tier, evidence that demonstrates a rationale, needs only a logic model tied to research plus an ongoing effort to study effects; no study result is required at all. The Department's own guidance, revised on 28 September 2023, ties the top two tiers to whether a study meets What Works Clearinghouse standards, and the Clearinghouse's own page adds numeric floors: a sample of at least 350 students across at least two sites for the top tiers. Nothing in the statute requires the intervention to be a curriculum rather than software; an AI product is eligible for any tier its evidence actually supports.

The implementation question

The practical difficulty for a district evaluating an AI tool is that the tiers were built for interventions with a research history: reading programmes, tutoring models, class-size changes. A generative AI product launched in the last two years is unlikely to have a randomised or quasi-experimental study behind it at all, which leaves the bottom tier, demonstrating a rationale, as the honest answer. That tier asks only for a plausible logic model and a promise to keep studying the product; a vendor can meet it with a slide deck. A district that asks which tier a product meets and accepts ‘rationale’ as sufficient has not screened out much, because almost any product can clear that bar by describing how it is supposed to work.

What holds and what fails

The tier system holds where a study already exists and can be checked against WWC standards: sample size, comparison group, statistical significance are all verifiable facts, not claims. It fails as a filter when no study exists yet, which describes most generative AI products sold into schools this year. This is an editorial reading, not a statutory one: procurement staff who want the law's evidence requirement to do real work should ask for the tier, the sample, and the underlying study, and treat a ‘rationale’ answer as an admission that evidence is still pending, not as evidence itself.

• Which tier, by number and design, does the vendor claim, and can they name the study?
• Does the cited study match the population and setting the school actually serves?
• If the honest tier is demonstrating a rationale, what ongoing evaluation has the vendor committed to, and who will see the results?

Evidence tiers were written to sort mature interventions, not to certify novelty; asking where a product sits inside that structure is still the most concrete question a school can put to an AI vendor.