Bridging the Gap: Evaluating AI-powered Math Tutoring to Support Low-income Middle School Students

Eedi is a computer-assisted learning platform that identifies the misconceptions a student is likely to hold from their responses to diagnostic questions and tailors subsequent content to address them. In this trial the treatment version of the platform delivers content as GenAI-powered interactive dialogue, which replaces Eedi's standard static content (instructional videos and fluency practice). Students obtain help in two ways: reactively, when the AI tutor opens an interactive dialogue after an incorrect answer; and on request, when a student asks for help and a live human tutor joins. Schools that do not opt in to the interactive dialogue instead receive Eedi's standard static-content version, with a live human tutor available on request; to date all participating schools have opted in. Treatment teachers also receive Eedi copilot support emails summarizing platform insights and offering teaching tips; these go to all treatment teachers and are not separately randomized.

2026-09-01

2027-06-30

-Overall math achievement as measured by scaled outcome score on Star Maths assessment
-Differential impacts of math achievement by low-income students as measured by scaled outcome score on Star Maths assessment

-Student self-reported beliefs about math and math learning as measured through surveys administered by Eedi
-Teacher experience with adapting to a new platform measured through surveys administered by Eedi

Note that pre-specified subgroups include low-income status, baseline attainment, and year group, each on the STAR Maths scaled score

This study is a cluster randomized controlled trial (RCT) designed to evaluate the impact of Eedi on student mathematics learning over one academic year. The study will be conducted in secondary schools (Key Stage 3) in the UK.

Year-groups (UK equivalent of grade levels in the US) in participating schools have been randomly assigned to either the Treatment group or the Control group. We will communicate the group assignments to the teachers prior to the beginning of the school year.

Not available

Randomization was conducted at the school year-group level (equivalent to US grade level). Each of the 28 schools contributed three year groups (Years 7, 8, and 9), and each school has at least one treated and one control year group. Randomization proceeded in two stages. First, schools were assigned to one of two pools using constrained randomization: a one-treatment pool (one of three year groups treated) and a two-treatment pool (two of three year groups treated), with 14 schools in each pool; one school was pre-specified into the one-treatment pool to accommodate a school-level implementation constraint. We simulated 100,000 candidate assignments, scored each on the standardized differences between pools in school size and share of low-income students, retained the best 20 percent, and selected one assignment at random from that acceptable set. Second, treated year groups were assigned within schools: in one-treatment schools a single treated year group was drawn at random, and these grade counts were mirrored in the two-treatment schools' control assignments. This yields exact grade balance, with 14 treated and 14 control school-year groups at each of Years 7, 8, and 9 (42 treated and 42 control in total). Primary inference will use randomization inference that mirrors this assignment procedure.

year group levels within schools

Yes

84 year groups

15,000

24 year groups treatment (access to Eedi), 24 year groups control (business as usual)

Using our sample size of 84 year groups (with a harmonic mean of 168 pupils per year-group), we estimate power for a two-level clustered design using the online PowerUpR calculator. We vary ICC2 assumptions between 0.1 and 0.2 and also analyze a scenario with attrition of two schools (6 year-groups fewer). We expect 60% of level 1 variance explained by covariates (notably prior attainment) and 30% of level 2 variance explained by covariates. We estimate power for a standardized mean difference (Cohen’s d). Results vary between an MDES of 0.166SD and 0.242SD at 80% power.