Do educational games work? What the meta-analyses say

Yes, modestly, and only under conditions. Pooled across 57 studies, digital learning games beat ordinary instruction by about a third of a standard deviation (Clark and colleagues 2016, g = 0.33), the gain comes from how the game is designed rather than from it being a game, and single-session play produces almost nothing measurable while play spread over many sessions does. Numbers reported by the companies that sell learning games run two to five times larger than independent trials of the same products. Peabody (peabody.kids) is a free AI-driven adaptive learning platform for kids with about twenty browser games, free through the end of 2026. It has no trial of its own, so this post is about the field's evidence and how Peabody is built to sit honestly inside it.

What a meta-analysis is, in one paragraph

A single study of one game in one classroom can say almost anything. A meta-analysis pools every study that meets a quality bar, converts each result to a common unit, and reports the average with a confidence interval. The unit is a standardized mean difference, written g or d: the gap between the game group and the comparison group, measured in standard deviations of the test. By Cohen's rough convention 0.2 is small, 0.5 is medium and 0.8 is large. In elementary and middle school, measured on real tests, those labels are too generous; a software supplement that moves a standardized test by 0.10 is doing well. The letter k is the number of studies pooled and N is the number of children across them.

The numbers worth knowing

Clark and colleagues' own headline is "the key role of design beyond medium": the medium is not the active ingredient, the design is. The primary paper is at https://doi.org/10.3102/0034654315582065.

The Wouters synthesis (https://eric.ed.gov/?id=EJ1008015) carries the finding that gets quoted least: across 31 studies and 2,216 learners, games were not significantly more motivating than lectures or drill. The assumption that a game is automatically more fun than a lesson is not supported by the pooled data.

Sailer and Homner (https://eric.ed.gov/?id=EJ1245270) looked at game elements bolted onto non-game learning. Their cognitive number is the largest in the table, and they warn in the same paper that novelty effects and publication bias may inflate it.

Kim and colleagues (https://files.eric.ed.gov/fulltext/EJ1323865.pdf) is the one to read if your child is under nine. The effects were biggest for preschoolers, for constrained skills like letter sounds, counting and number facts, and for tests the researchers wrote themselves. On standardized tests and on unconstrained skills like comprehension and problem solving the effects shrank.

What "modest" means

A gap of 0.33 standard deviations moves the average child from the 50th percentile to roughly the 63rd, if everything else is equal. That is worth having. It is not a transformation.

The spread is the other half of the story. A 2026 second-order meta-analysis in Educational Research Review, pooling earlier meta-analyses of math games (https://www.sciencedirect.com/science/article/abs/pii/S1747938X26000552), reports a pooled effect of 0.45 with heterogeneity of 93.9%, which means that almost all the variation between studies is real difference between games and settings, not noise. The authors call the result positive but small given that spread. A parent cannot buy "the average game"; you get one particular game, and the design of that game decides which end of the spread it sits at.

One session does nothing; many sessions do

The most useful single moderator in Clark and colleagues' tables is session count. Games played in a single session came in around 0.08, not significantly different from zero; games played across multiple sessions came in around 0.44. (Those subgroup values were read from a secondary copy of the moderator table and should be checked against the paper before being quoted in print.) Wouters and colleagues found the same shape: games worked better when played over multiple sessions, when supplemented with other instruction, and when played in groups.

The apps pooled by Kim and colleagues averaged about 21 minutes a session. The one randomized trial of Khan Academy Kids, reported by Khan Academy (https://blog.khanacademy.org/khan-academy-kids-improves-pre-literacy-skills-in-preschoolers-research-confirms/), used about 13 minutes a day for ten weeks with 49 four- and five-year-olds. These are the doses that produced effects: short, repeated, spread out.

The practical reading is that a game your child plays for ten minutes on many days should do more than one played for an hour once. That matches the wider spacing literature in spaced practice with AI.

Why the number on the box is bigger than the number in the journal

Every well-known learning product has two numbers: the one its marketing prefers and the one an independent trial found.

The ST Math randomized result is at https://files.eric.ed.gov/fulltext/ED616922.pdf and the Zearn trial at https://edworkingpapers.com/sites/default/files/Efficacy%20of%20Zearn%20Math%20for%20EdWorkingPapers%20v2.pdf. The GraphoGame trial is at https://educationendowmentfoundation.org.uk/projects-and-evaluation/projects/graphogame-rime.

Three things explain the gap.

1. Which test. Kim and colleagues found that researcher-made, in-game-aligned tests routinely show two to three times the effect of standardized tests. A company measures what its game teaches; a school measures what the state asks.
2. Which comparison. Sitzmann (2011, Personnel Psychology; https://doi.org/10.1111/j.1744-6570.2011.01190.x), pooling 65 studies of simulation games, found that when the comparison instruction was itself active, learners in the games learned less than the comparison group, and reported strong evidence of publication bias on top.
3. Who counted. A quasi-experiment chosen and funded by the vendor sits at ESSA's moderate tier; a randomized trial run by a third party sits at the strong tier. Zearn and ST Math are the only widely used K-5 math products with strong-tier standardized-test effects, and those effects are +0.07 to +0.11. That is what "it works" looks like when someone else counts.

What a parent should look for

1. Short, repeated sessions. Ten minutes should be a natural unit, with a stopping point the child can see.
2. The learning inside the play, not between levels. A quiz that interrupts a game is the weakest arrangement in the taxonomy; a question that is itself the move is the strongest. A companion post on intrinsic integration goes through the experiment behind that.
3. A precise skill named. "Letter sounds for the first eight phonics steps" or "addition facts to 20" is a claim you can check at the kitchen table. "Boosts math" is not.
4. Which test the claim rests on. Ask whether the number comes from the product's own items or from something a teacher would recognize.
5. A place for you. Adult engagement is the strongest moderator in the phonics-game evidence; a product that gives you something to talk about with your child is using that lever.
6. Claims no bigger than the evidence. The 2021 complaint against Prodigy to the FTC (https://fairplayforkids.org/feb-19-2021-advocates-to-ftc-prodigy-math-game-preys-on-kids-and-families/) turned on unsubstantiated efficacy claims alongside aggressive upselling. A company that says "we have not run a trial" is telling you something a company that says "studies show" may not.

What this doesn't show

• The averages hide games that did harm as well as games that did well. Heterogeneity above 90% means the pooled number is a poor predictor of any one game.
• Almost all of these studies ran in schools with a teacher present. Evidence on unsupervised home play is thin.
• Games do not improve general thinking. Simons and colleagues (2016, Psychological Science in the Public Interest; https://journals.sagepub.com/doi/abs/10.1177/1529100616661983) reviewed the brain-training literature and found little evidence of real-world benefit; what a game teaches is the thing it practices.
• Nothing here is evidence about Peabody, which has not been tested in a trial.

Where Peabody stands

Peabody takes three positions that follow from the evidence.

It has no trial and says so. There is no efficacy page and no effect size anywhere on the site.

Its rules are published. The engine that chooses every question, called the practice brain, is ordinary code, and its constants are written up in how Peabody picks the next question: decks that are about 85% things the child already knows, at most two new items a deck, reviews at growing gaps, a question answered right never asked again. AI writes the stories; code decides the rest. A test suite called the learning invariants runs before any game can change, and it checks that every game asks the brain, that every answer is three boxes, and that Just play is the only question-free mode.

It measures play time and acts on it. Games are ranked by how much children actually play them, a new game arrives most days, and games that go unplayed are retired to a Trash tab. There is no thumbs-down button and no leaderboard; the measure is what children choose when every game is one tap away.

What a parent sees in the free grown-up account is kept in the same honest register: which facts are known, which are being learned and which were missed, with "fast in games" kept separate from "known in reading". Or skip the account and start free on any browser.

Learn more elsewhere

• Clark, Tanner-Smith and Killingsworth, 2016: the digital-games meta-analysis, free to read in full.
• Stanford SCALE: the evidence base on AI in K-12: a 2026 review: of more than 800 papers, about 20 were strong causal studies, and hint-giving tutors did better than answer-giving chatbots.
• Dunlosky: strengthening the student toolbox: ten study habits ranked; practice tests and spacing win, highlighting and rereading don't.

For about seventy more, checked and sorted by who they're for, see where to learn about AI.

Questions parents ask

Is a game better than a worksheet?

On the pooled evidence, a well-designed game is a little better than ordinary instruction for the specific skill it practices, and a badly designed one is worse. Clark and colleagues' value-added result says one good design feature is worth as much as the medium itself, so the honest answer is that it depends on the game far more than on the format.

How long should my child play?

The studies that found effects used short sessions, roughly 13 to 21 minutes, on many days. Multiple short sessions were the strongest moderator in both Clark's and Wouters' syntheses. Treat one long sitting as play rather than practice.

Do points, badges and stories help?

Sailer and Homner found that game elements improved what was learned (g = 0.49) but that the motivational and behavioral effects were less stable, and that a story frame moved behavior without moving learning. Points and stories are good at bringing a child back; they do not do the teaching.

Has Peabody been tested?

No. Peabody has no published efficacy study and makes no outcome claims. What it publishes instead is its rules, and what it measures is which games children choose to play.