External Representations
External representations — the artifacts humans construct to make spatial structure visible outside the head — are not neutral containers for pre-existing thought. They are active participants in cognition. A timeline forces sequential thinking. A hierarchy forces categorical thinking. A network enables relational thinking that pure text struggles to carry. Tversky's research demonstrates that changing the representation changes what the thinker can think. The affordances of the external tool determine which inferences become easy, which become difficult, and which become nearly impossible. This is why the choice of representation in AI collaboration is not a presentation issue but a cognitive one.
In the AI Story
The tradition runs from Euclidean geometry — whose proofs depend on diagrams that carry information the propositions alone do not — through modern data visualization. Edward Tufte's principle that graphical excellence gives the viewer the greatest number of ideas in the shortest time captures the efficiency gain external representations offer when well-constructed. But Tversky's insight is deeper: the representation does not merely communicate thought faster, it enables thoughts that would not otherwise occur.
This matters for AI collaboration because large language models produce text by default. Text is a sequential representation with specific affordances: it is good for narrative, argument, and stepwise instruction, but it struggles to represent parallel structure, mutual dependency, and spatial relationship without substantial effort. When Claude explains a system architecture in prose, the user must reconstruct the spatial structure mentally from the sequential text — reintroducing precisely the representational mismatch the language interface was supposed to eliminate.
The commonplace book tradition that Ann Blair documents represents a historical answer to the abundance problem: external representations as instruments for thinking through too much information. The AI age inherits this problem at a new scale, and the old tools — structured note-taking, indexing, spatial arrangement — are being rebuilt for a new era.
For builders working with AI, the practical implication is to sketch before prompting. Externalize the spatial model first, then describe it in language. The sketch enforces precision that prose permits to remain vague, and the precision is preserved in the subsequent verbal description.
Origin
Tversky's research on external representations spans the 1990s through the 2010s, with particular attention to how diagrams support reasoning in education, how maps organize spatial knowledge, and how sketches function in design practice. The work synthesizes traditions from cognitive psychology, education research, and information visualization.
Key Ideas
Representations are active. External structures do not merely record thinking; they shape what thinking is possible within them.
Affordances and constraints. Each representation makes certain inferences easy and certain inferences hard — there is no universal best representation, only matches and mismatches with problems.
Sketching as discovery. The hand that draws finds spatial relationships the mind alone cannot conceive. Drawing is not recording but thinking.
Representation-problem fit. The effectiveness of any cognitive tool depends on alignment between the tool's spatial structure and the problem's spatial structure.
Appears in the Orange Pill Cycle
Further reading
- Tversky, Barbara. "Visualizing Thought." Topics in Cognitive Science 3 (2011).
- Tufte, Edward. The Visual Display of Quantitative Information (Graphics Press, 1983).
- Larkin, Jill H. and Herbert A. Simon. "Why a Diagram Is (Sometimes) Worth Ten Thousand Words." Cognitive Science 11 (1987).
- Scaife, Mike and Yvonne Rogers. "External Cognition: How Do Graphical Representations Work?" International Journal of Human-Computer Studies 45 (1996).