Intrinsic integration: why the learning must be the game

The best-known experiment on where learning should sit inside a game found that the same game, with the same division content, taught more durably and was chosen about seven times more often by children when the arithmetic was the attack rather than a quiz between levels (Habgood and Ainsworth 2011). The lesson is that a learning game works when the thing you know is the thing you do. Peabody (peabody.kids) is a free AI-driven adaptive learning platform for kids with about twenty browser games, free through the end of 2026. This post explains the experiment, the taxonomy that came from it, where each of Peabody's game mechanics sits on that line, and the honest gap that remains.

Chocolate-covered broccoli

The phrase is the field's own insult for a learning game that is a quiz with a cartoon on top. It is conventionally attributed to Amy Bruckman's 1999 Game Developers Conference talk, "Can educational games be fun?", as cited in a Carnegie Mellon ETC Press paper (https://press.etc.cmu.edu/file/download/1921/242bac6e-a045-4056-8e8a-b2bd990a2f21); whether she coined it is not established. Jake Habgood, who ran the experiment below, later used it in a 2015 talk titled "Blending chocolate, broccoli and skeletons" (https://www.de.ed.ac.uk/node/435) to mean learning content loosely bolted onto play.

The idea is older than the phrase. Malone and Lepper's 1987 taxonomy of intrinsic motivations for learning (https://paperguide.ai/papers/8747b1d0-764e-4259-8ea4-a53c5e07e9f4-making-learning-fun-a-taxonomy-of-intrinsic-motivations-for-learning) lists challenge, curiosity, control and fantasy, and under fantasy it draws the line that matters here. In an endogenous fantasy the fantasy depends on the skill: the word in Hangman is the gallows. In an exogenous fantasy the fantasy is a wrapper: answer a sum and the rocket goes. Malone and Lepper argued that the first kind is where learning and fun reinforce each other, and the second is where they merely take turns.

The Zombie Division experiment

Habgood and Ainsworth, in the Journal of the Learning Sciences (2011; author copy at https://shura.shu.ac.uk/3556/1/Habgood_Ainsworth_final.pdf), built a 3D dungeon game called Zombie Division for seven- to eleven-year-olds and made three versions of it.

In the intrinsic version, skeletons carry numbers on their chests and the player's weapons are divisors. You defeat a skeleton by choosing a weapon that divides its number. Division is the combat.

In the extrinsic version, the combat is identical but the skeletons have no numbers, and the same division questions are asked as quizzes between levels. Same game, same content, same amount of each; only the place where the math lives is different.

A control version had no math at all.

Their definition of intrinsic integration has two parts, and both matter. First, the learning content is delivered through the parts of the game that are the most fun to play. Second, the learning content is embodied in the structure of the game world and the player's interactions with it, so the player uses the knowledge to act rather than to unlock acting.

Study 1 put 58 children in a classroom with playing time strictly matched at two and a quarter hours plus teacher-led reflection. All three groups improved on an immediate test. The difference showed up on the delayed test, where the intrinsic group outscored both the extrinsic and control groups (F(2,48) = 7.49, p < .001, partial eta squared = 0.24), with the intrinsic group scoring 20.66 points above the control group.

Study 2 gave 16 children in an after-school club free choice between the intrinsic and extrinsic versions. They spent 75.7 minutes on the intrinsic version and 10.28 minutes on the extrinsic one (t(15) = 7.38, p < .001), more than seven times longer. This is the result that gets quoted, and it deserves its caveats: one game, one research group, 16 children in the free-choice study, and the authors themselves call for replication. A 2022 ACM follow-up using eye tracking (https://dl.acm.org/doi/pdf/10.1145/3549503) argues the mechanism is attention: when the learning object is the thing you must look at to play, attention goes where the learning is.

The pooled evidence points the same way from a different angle. Gui and colleagues (2023, International Journal of STEM Education; https://link.springer.com/article/10.1186/s40594-023-00424-9) found that game design elements added for content learning were more effective than elements added for the gaming experience.

The taxonomy: four points on a line

DragonBox is worth a sentence because it shows the other edge of intrinsic integration. In a 2023 randomized trial of 1,850 seventh graders (Decker-Woodrow and colleagues, AERA Open; https://pmc.ncbi.nlm.nih.gov/articles/PMC10125888/), DragonBox beat an active control on an algebra test (g of about 0.27 after covariates), but earlier studies found children who enjoyed it more learned less than from a textbook unless the game explicitly bridged from its own symbols to standard notation. A learning game that is fully its own world can leave the child fluent in the world and lost on paper.

Where Peabody's games sit

Peabody's games come in three kinds, and each sits at a known point on the line.

Slow games put the question on a thing and let the child set the pace. In Deep Dig the layers of rock are the phonics sequence and a word is on the rock you are breaking; in Potion Shop you read the orders; in Stone Age the science question is on the thing you are building. These are diegetic gating: the question is spatially and causally inside the world, the world responds to the answer, and the verb (dig, brew, build) is not itself the content.

Propelled games are runners. Peabody Rush, Dino Dash, Temple Escape and Rail Dash run on their own, and the rule for all of them is stop, don't slow: the run comes to a full stop at a station, the question is asked, the child answers, and then Go. The full stop is deliberate. Mayer's work on modality and attention, and the 2022 eye-tracking follow-up to Zombie Division, both say that splitting attention between a moving world and a question is bad for both; a runner that keeps moving while a child reads is asking a six-year-old to do two things at once. So the stop is the right call. It is also, honestly, a step below full integration: a station is an interrupt quiz in the world's clothing. The run is the fun part and the station is where the run is not happening.

Two of the runners already do better. In Rail Dash the lanes are the three answers, so picking the answer and steering are the same motion. In Dip & Dash the answers are doors in the stage. The next step for every runner is to make the three boxes the three exits from the station, so that the answer is a movement verb and the chosen exit changes the next stretch of world. That is the second half of Habgood's definition, and it is not built yet.

Fast casual games treat learning as a bonus and live near the reward-wrapper end by design, which is why every game is also measured by how much children choose to play it.

Three rules apply to all of them. Three boxes, no typing: every answer in every game is three big tap boxes, so the question is the same shape whether it sits on a crate or at a station. Just play: every game has a question-free mode, the game must play whole in it, and it pays from the world while learning modes pay more. Ryan, Rigby and Przybylski (2006, Motivation and Emotion; https://selfdeterminationtheory.org/SDT/documents/2006_RyanRigbyPrzybylski_MandE.pdf) found that in-game competence and autonomy predicted enjoyment and preference; a visible way to decline the question is autonomy support, and which mode a child picks when both are one tap away is the cleanest measure of whether the learning is the game. Heard, not shown: when the word is the answer, Peabody says it and never prints it above the boxes that carry it, because matching letters to letters is not reading, and "starts with" questions show the picture, not the word. This is Mayer's modality principle (narration rather than on-screen text) applied to a reading game.

Under all of it, AI writes the stories; code decides the rest. The question on the crate comes from the practice brain, ordinary code whose rules are written up in how Peabody picks the next question, and a question answered right is never asked again, so the crate always carries something worth answering.

What this doesn't show

• Habgood's free-choice result rests on 16 children and one game. The direction is well supported by theory and by Gui and colleagues' pooled finding, but the size of the effect is not settled.
• The taxonomy is a line of reasoning, not a measured scale. Nobody has run the four positions head to head in one game.
• Full intrinsic integration can produce a transfer gap, as DragonBox shows, unless the game names what the grown-ups call the thing. Peabody's stories practice reading concepts without naming them, and the bridge to the standard name is still a to-do.
• Peabody has not been tested in a trial. Whether its diegetic gating beats a quiz for its own players is unmeasured; the number to watch is the share of play time spent in learning modes when Just play is one tap away.

What a parent can do

Play one of the games with your child and ask one question: is the thing they have to know the thing they have to do? In Deep Dig the word is on the rock; in Rail Dash the lane is the answer. Then watch what happens when they find the Just play switch. A child who turns the questions off and never comes back has told you something true about that game, and Peabody counts that too. The games are on the games page, and the free library is next door when the running is done.

Questions parents ask

What is intrinsic integration in plain words?

The learning is the move. In Zombie Division you divide to attack. In a quiz game you attack, then stop to divide, then attack again. Habgood and Ainsworth's experiment found the first taught more durably and was chosen seven times more often when children could pick.

Isn't a quiz between levels still practice?

Yes, and in Study 1 the quiz group did improve on the immediate test. The difference appeared on the delayed test and in free choice. A quiz between levels teaches; it just does not make children want to come back to it.

Why do Peabody's runners stop completely instead of slowing down?

Because a child cannot read a question and dodge a train at the same time. The full stop protects the reading. The honest cost is that the station is a quiz in the world's clothing, and the planned fix is to make the three answers the three exits.

What is Just play for?

It is the control condition built into every game. A child who wants the run without the questions can have it, the game plays whole, and the share of time children spend in learning modes anyway is the measure Peabody trusts most.