In 1961, at Stanford, children watched an adult hit an inflatable doll called Bobo. Then they were left in a room with the same doll. Many of them did exactly what they had seen, sometimes with startling accuracy. Albert Bandura's experiment is among the best known in psychology, and it says something every parent suspects: children don't learn only from what we tell them. They learn from what they watch.

In this article we'll use Bandura's idea, but turn it around. The child won't watch a model. The child will be the model.

What Bandura said

Before Bandura, much of the psychology of learning held that people learn by doing something themselves and getting a reward or a punishment. Bandura showed that we can learn whole behaviours just by watching someone else, without trying and without a reward. He called it observational learning.

Later he added a second big idea: self-efficacy, the belief that you can manage something. In his view it grows most from your own successes, then from seeing someone similar to you succeed, and only after that from encouraging words.

The twist: learning by teaching someone else

There's one direction in which Bandura's idea gets even stronger. When the child is the model rather than the observer, they have to organise what they know in order to show it to someone else.

In 2009 a Stanford team led by Catherine Chase and Daniel Schwartz studied students teaching a virtual character. The students put in more effort when teaching the character than when learning the same material for themselves. The researchers called it the protégé effect. A child who takes care that someone else understands ends up understanding better themselves.

The engineering side: the rubber duck

Programmers have an odd habit, described in "The Pragmatic Programmer": when you can't find a bug, explain your code line by line to a rubber duck. Often, halfway through the explanation, you see the mistake yourself. The duck does nothing. The explaining does the work.

The same thing happens when a senior explains to a junior how the system works. Usually the senior learns something new about their own code. In this game your child's favourite character is the rubber duck, only a little more scatterbrained.

The game: "Teach Dino"

The learning goal is sorting and explaining: where animals live, on land, in water or in the air. The character here is called Dino, but it can be anyone your child loves.

1. Watch. Dino sorts three animals and thinks aloud: "A fish has no legs, it has fins. I'll put it in the water." This is Bandura's model. The child sees not just the answer, but how you get there.
2. Teach Dino. Now Dino sorts on his own and sometimes gets it wrong in a funny way: the fish in the tree, the bird under water. Your child moves the animal to the right place. Dino asks "Why?" and your child explains aloud.
3. Praise Dino. When Dino gets it right, your child taps "Well done, Dino!". There is praise here, but it goes from the child to the character, not the other way round.
4. Dino's test. At the end new animals arrive and Dino makes fewer and fewer mistakes. Your child sees that their explanations helped.

The most interesting animals are the ones that don't fit in one place: a frog, a penguin, a duck, a turtle. When Dino hesitates, he asks: "Can it go in both places?" There's no single right answer here, and it's the best conversation in the whole game.

We should be honest about one thing. Dino doesn't really learn. The game simply makes fewer mistakes each round. The learning happens in the child, while they explain.

The prompt

Build a small game as a single self-contained HTML file, based on
Albert Bandura's observational learning and the "protégé effect":
the child teaches a character who makes mistakes. Language: English.

Topic: where animals live. Three zones: land, water, air.
16 animals as emoji, including a few that live in more than one
place: frog, penguin, duck, turtle.
Character: [the child's favourite], called "Dino" below.

Three parts:
1. "Watch": Dino sorts 3 animals and thinks aloud (text and
   speechSynthesis), for example: "A fish has no legs, it has fins.
   I'll put it in the water."
2. "Teach Dino": Dino places one animal at a time in a zone. About
   a third of the time he gets it wrong in a funny way (a fish in a
   tree, a bird under water). The child can move the animal. After
   a move Dino asks "Why?", the child explains aloud, and an
   "I see!" button continues. When Dino is right, the child can tap
   "Well done, Dino!".
3. "Dino's test": 6 new animals. Dino makes fewer mistakes each
   round. For an animal that fits two places, Dino asks "Can it go
   in both places?" and accepts both.

Rules: the game never grades the child. No points, stars, timer,
"right" or "wrong" aimed at the child. Dino's mistakes are funny,
not sad. Nothing is stored after the game is closed.

Design: calm colours, large touch targets, works on a phone,
no ads and no external requests.

The uncomfortable part

The Bobo doll experiment has critics, and they're right about some things. Bobo was made to be hit, and a lab isn't a living room. But the main finding holds up: children copy the models around them, and not only what we'd like them to copy.

And the most important model in a child's life isn't a character in a game. It's you. A 2015 study by Erin Maloney and colleagues found that when parents who are anxious about maths often help with homework, their children learn less maths over the year and become anxious about it themselves. A child doesn't just hear the explanation. They hear the sigh before it.

So the most useful part of this game might not be Dino at all. It's your child seeing you get something wrong, laugh, and fix it. "Oops, I put the fish in the tree. Let's try again." That's a model no app can provide.

Bandura isn't one of the six in the book from the AI Learning Game Kit, but the kit does the same job for the part after the prompt: from a working game to an icon on the phone.

Children don't do what we tell them. They do what they see us do. Including when we get it wrong.