Learning through games: the design cheat sheet
What the meta-analyses say makes a learning game work, where the learning should sit, how rewards and difficulty affect children, and the metrics that can't be gamed.
Learning games work modestly, about a third of a standard deviation over ordinary instruction in the pooled studies, and design decides most of it. The learning should be the game's main action, play should be spread over many short sessions, rewards should be small and informational, and nothing a child owns should ever be taken away.
This is the one-page version of the research on learning through play, for designers, teachers and parents choosing a game. Each row names the study behind it. Peabody (peabody.kids) is a free AI-driven adaptive learning platform for kids where the learning happens inside more than twenty browser games; it is free through the end of 2026, and the last column says how Peabody applies each rule. AI writes the stories; code decides the rest.
The pooled evidence on one screen
| Synthesis | What it pooled | Effect | The design lesson |
|---|---|---|---|
| Clark, Tanner-Smith and Killingsworth 2016 | K–16 digital games, 57 studies | g = 0.33 vs non-game instruction; g = 0.34 for an added design feature | Design matters more than medium (doi) |
| Wouters et al. 2013 | Serious games, 77 comparisons, 5,547 learners | Learning d = 0.29, retention d = 0.36; motivation d = 0.26, not significant | Better with other instruction, many sessions, groups (ERIC) |
| Sailer and Homner 2020 | Gamification (points, badges, story) | Cognitive g = 0.49; motivation 0.36; behaviour 0.25 | Story moves behaviour, not learning (ERIC) |
| Kim et al. 2021 | Apps for preschool to grade 3, 36 studies | +0.31 overall; literacy 0.35, maths 0.29 | Biggest for constrained skills and researcher-made tests (PDF) |
| Tokac, Novak and Thompson 2019 | PreK–12 maths video games, 24 studies | d = 0.255 | "Small but marginally significant" (ERIC) |
| Gui et al. 2023 | STEM games | d = 0.677 | Features added for content beat features added for fun (Springer) |
Read the standardised-test numbers smaller: aligned, researcher-made tests routinely show two to three times the effect of standardised ones.
Twelve design rules
| Rule | Evidence | How Peabody does it |
|---|---|---|
| 1. Many short sessions beat one long one | Clark 2016: single-session games about 0.08, multi-session about 0.44 (subgroup values flagged to verify) | Rounds of about 30 questions, with reviews spread over days and weeks |
| 2. Put the learning in the main action | Habgood and Ainsworth 2011: same game, maths as the attack vs quizzes between levels; better delayed test, 75.7 vs 10.3 minutes of free play (PDF) | Questions sit on things in the world; in Rail Dash the lanes are the answers |
| 3. Stop, don't slow, when a question needs reading | Mayer's modality work; a 2022 eye-tracking follow-up to Habgood (ACM) | Runners come to a full stop at a station |
| 4. Teach before the game needs it | Mayer 2019, pretraining is one of five supported features (Annual Review) | A new item is taught on a card before it is asked |
| 5. Explain, don't just mark | Mayer's coaching; immediate right/wrong alone did not beat a control in the DragonBox RCT (PMC) | A miss comes back three questions later; explanations after a miss are still uneven across games |
| 6. Narration over on-screen text | Mayer's modality principle | When the word is the answer it is said, never printed above the boxes |
| 7. Plain visuals for anything that carries content | Clark 2016: schematic beat realistic | Three big flat answer boxes in every game |
| 8. Aim for mostly right, invisibly | Math Garden targets 75% (ERIC); Wilson 2019's 85% for simple learners | Rounds are about 85% known items |
| 9. Never rubber-band | Hunicke 2005: players notice and resent adjustment that cheats (ACM) | Difficulty changes what is asked, never the world's honesty |
| 10. Let children decline the learning | Ryan, Rigby and Przybylski 2006: autonomy predicts enjoyment (PDF) | Every game has a Just play mode that plays whole |
| 11. Co-op over head-to-head | Clark 2016 via Sailer and Homner: competition alone is weakest | No leaderboards, no comparison with other children |
| 12. Name the idea after the game teaches it | DragonBox's transfer gap (Chan et al. 2023, BJET) | Not yet done: the bridge to the grown-up name is open |
Where the learning sits: four points on a line
| Integration | What it looks like | Strength |
|---|---|---|
| Intrinsic | The skill is the verb: dividing a skeleton's number is the attack (Zombie Division); isolating the box is solving for x (DragonBox) | Strongest on delayed tests and free choice |
| Diegetic gating | The question is on a gate, crate or door, and the world answers; the verb (run, dig, build) is not the skill | Better than a quiz; most of Peabody is here |
| Interrupt quiz | Play pauses, a question appears, play resumes | Lower delayed gains; free play collapses |
| Reward wrapper | Play is the prize for answering elsewhere | Weakest |
A runner's station is an interrupt quiz in the world's clothing. The cheap fix the research report recommends is to make the three answers the three exits, so the answer becomes a way to move. Peabody's Rail Dash and Dip & Dash do; the other runners do not yet.
Rewards and motivation
| Finding | Number | Source |
|---|---|---|
| Expected, tangible rewards lower free-choice interest | d = −0.40 (for engaging), −0.36 (for finishing), −0.28 (for performance); worse for children | Deci, Koestner and Ryan 1999 (PDF) |
| Positive, informational feedback raises it | d = +0.33 | Same |
| Incentives drive quantity, interest drives quality | ρ = .21–.45, 183 studies | Cerasoli, Nicklin and Ford 2014 (PDF) |
| Points for effort and strategy kept low performers playing | About 15,000 children | O'Rourke et al. 2014, "Brain Points" (ACM) |
| Mindset lectures made players quit | About 25,000 children | O'Rourke et al. 2016 |
| Children value what they own from age 4–6 | 78% showed it on the first trial | PLOS ONE 2014 (paper) |
| Two-minute warnings made screen-time endings worse | Ages 1–5 | Hiniker et al. 2016 (UW) |
The rules that follow: keep pay for right answers small and never a rate a child can compute; reward effort and progress; nothing dies, no lives, no broken streaks; no timed chests or come-back-later bonuses; no random rewards with money value, which the FTC now restricts for under-16s (Genshin order 2025); natural stopping points and no autoplay. Peabody's rulebook, rule by rule, is in child-safe rewards.
Metrics that can't be gamed
Time on site and daily actives reward whatever keeps a child on the screen. The research report proposes measures a designer can't push up by manipulation:
| Metric | Why it is honest |
|---|---|
| Free-choice learning share: minutes in learning mode when Just play is one tap away | The classic measure of intrinsic motivation (Deci; Habgood Study 2) |
| Spaced-check retention: accuracy on retired items days and weeks later | Delayed recall is the learning, not same-session accuracy |
| Rapid-guess rate: answers faster than the question can be read | Gaming the system lowers learning (Baker et al. 2004) |
| Quit-point map: where sessions end | Natural ends vs frustration |
| Out-of-game check: a short paper-style test of the practised facts | Near transfer, the kind games can claim |
Peabody does not yet compute these; they are the plan, not a report.
What games don't do
Video-game training does not raise general cognitive ability against active controls (Sala, Tatlidil and Gobet 2018, LSE). Claim near transfer, the facts and sounds practised, and nothing about brains, attention or memory in general. Peabody has no efficacy study of its own and makes no outcome claim.
Try the integration line yourself: in Deep Dig the word is on the rock you break, and in Rail Dash the lane is the answer. Start free with no sign-up, or see every game.
Questions parents ask
Do learning games actually work?
Modestly. Pooled studies put digital games about a third of a standard deviation ahead of ordinary instruction, with big variation by design. On standardised tests the gains are smaller, about 0.1 for the best-studied products.
What is the difference between gamification and game-based learning?
Gamification adds points, badges and story to ordinary practice; game-based learning puts the practice inside a game. Gamification's story layer moves behaviour (coming back) more than learning; a game where the skill is the main action does better on delayed tests.
Are rewards bad for my child's motivation?
Expected, tangible rewards tied to doing a task can lower a child's own interest, and more so for children than adults. Small, informational feedback ("you can do all the 6s now") does the opposite. Taking things away is the riskiest of all, because children feel losses strongly from a young age.
How long should a learning game session be?
Short and often. Studies that found effects used sessions of about 13 to 30 minutes, and a first-grade classroom study found the longest on-task stretch was about seven minutes. Many short sessions over weeks beat one long one.
Checked against how Peabody works on 16 October 2026. Written with help from AI. Peabody is a reading app, not a reading assessment: if you're worried about your child's reading, talk to their teacher.