Transfer of learning

Why does something learned thoroughly in one setting so often fail to help in another?

Almost every argument for a subject rests on transfer. Latin trains the mind. Chess teaches strategy. Mathematics teaches rigour, which you will then apply to your taxes and your arguments. The claim is always the same shape: what you practise here strengthens something general that then works everywhere.

It is a comfortable claim, and it is mostly false. So the interesting question is not whether transfer happens — it plainly does, or nobody could read a new sentence — but why it stops so abruptly at the edge of the setting where it was learned.

Begin by asking what would have to be true for the comfortable claim to hold. It requires that learning deposits something detachable: a capacity for reasoning, or attention, or problem-solving, that sits apart from the material it was built on and can be pointed at anything. Is that what learning looks like from the inside?

Consider what you actually remember from a lesson. Not a naked principle, but a principle wearing clothes: the worked example, the notation used, the kind of problem it appeared in, the room, the teacher's phrasing. Memory encodes an event, and an event includes its circumstances. Godden and Baddeley had divers learn word lists on land or underwater, and recall was measurably better when the setting matched. If even the water around you gets bound into the trace, why would the far more relevant clothing — the notation, the problem type — fall away?

Now the consequence. If knowledge is stored with its context, then whether it transfers depends on how much of that context the new situation reproduces. That predicts a gradient, not a switch: near transfer, to situations that look and feel like the training, should work reasonably; far transfer, to situations sharing only an abstract structure, should mostly fail. And that is what the record shows. Thorndike and Woodworth tested the formal-discipline theory in 1901 and found training on estimating the areas of rectangles barely helped with other shapes; Thorndike's later work found no general mental improvement from studying Latin. A century on, the same result keeps arriving in new clothes: Owen and colleagues trained thousands of people online and found gains on the trained tasks and no spread to untrained ones, and meta-analyses of working-memory training and of chess and music instruction find near transfer and near-nothing far transfer.

But there is a second cause, and it bites harder. Suppose the structure really is shared — does the learner see it? Gick and Holyoak gave people a story about a general splitting his army to converge on a fortress, then the medical problem of destroying a tumour without burning the tissue in the path of the beam. The two are the same problem. Very few solved it spontaneously; told there was a hint in the story, most did. The knowledge was present and unretrieved, because retrieval is cued by surface features, and the surfaces did not match.

Which raises the question of what a surface even is. Chi, Feltovich and Glaser asked physics novices and experts to sort problems into groups. Novices grouped by apparatus — pulleys with pulleys, inclined planes with inclined planes. Experts grouped by principle — all the conservation-of-energy problems together, whatever the picture showed. Nobody taught the experts to see it that way; the categories are a residue of having solved hundreds of problems. So the abstraction that makes transfer possible is not a shortcut around domain knowledge. It is what a large amount of domain knowledge eventually turns into.

Does that mean people are simply bad at abstraction? No — and this is the part that should unsettle the schooling story rather than the learner. Brazilian children working as street vendors reliably computed prices in the market and then failed formally equivalent problems written out on paper. The competence was real. It just did not recognise the classroom as its home.

Two refinements hold the picture together. First, "far transfer is rare" is a claim about spontaneous transfer. Structure can be taught directly — give learners two analogues and ask them to say what the pair has in common, and abstraction improves markedly. The pessimism is about the by-product, not about what deliberate instruction can achieve. Second, near transfer being easy is not a consolation prize: most of what we want from education is near, and the honest move is to define the target situations and practise things that resemble them, rather than to hope a general faculty will grow.

Where I have simplified: some researchers argue the standard tests are too strict, measuring whether learners can immediately solve a new problem rather than whether they are better prepared to learn how — Bransford and Schwartz's "preparation for future learning" — and by that measure the picture is less bleak. That reframing is contested. What is not seriously contested is the negative result on which the popular claims depend: no one has shown that practising a narrow cognitive task makes people generally smarter.

• Why interleaving and varied practice improve transfer while making practice feel worse.
• Whether teaching by comparing two analogues can be built into ordinary curricula.

Every term the collection defines is gathered in the glossary.