Imagination-to-Understanding Ratio
The imagination-to-understanding ratio names the companion problem that Segal's imagination-to-artifact framework does not solve. Alan Kay, the computer scientist whose work at Xerox PARC laid the conceptual foundations of personal computing, has argued throughout the AI moment that the collapse of the production distance must not be mistaken for progress on the comprehension distance. A practitioner can now produce sophisticated outputs — working code, drafted briefs, synthesized analyses — whose causal structure she does not understand and whose failure modes she cannot predict. Gopnik's developmental framework provides the empirical grounding for Kay's concern: the causal model that understanding requires is built through the very cognitive labor that the imagination-to-artifact collapse has eliminated.
In the AI Story
The imagination-to-understanding ratio was a larger problem than the imagination-to-artifact ratio for most of computing history. A programmer who could write assembly understood the machine at the hardware level, because she had to. A programmer writing Python does not need to understand memory management, because the language handles it. A developer using Claude does not need to understand the language at all, because the AI handles the translation. At each step, the imagination-to-artifact ratio has shrunk. At each step, the imagination-to-understanding ratio has grown.
Kay's argument is that the expansion of the understanding ratio is the actual technical and educational problem of computing, and that the AI moment has made the problem worse rather than better. The seductive appearance of capability — the user producing outputs that exceed her understanding — conceals the specific deficit that the blicket principle names: the absence of the causal model that would make the outputs genuinely the user's own.
Gopnik's framework clarifies why this matters developmentally. Understanding, in the theory-theory framework, is not propositional knowledge about how something works. It is a causal model that can be deployed to predict, to intervene, to generalize to novel situations. Such models are built through the specific cognitive labor of constructing and testing hypotheses against the world. If the labor is skipped — if the output arrives without the cognitive construction — then the model has not been built, and what remains is the surface appearance of understanding without its developmental substrate.
The practical implication, which Kay has emphasized in talks and interviews throughout the 2020s, is that AI tools must be designed with the imagination-to-understanding ratio in mind. A tool that produces outputs but also scaffolds the user's causal understanding of why the outputs work is a scaffold. A tool that produces outputs and leaves the user with no causal model is a substitute. The difference is the difference between amplifying a mind and replacing one, and the difference is invisible in any metric that measures only the quality of the outputs.
Origin
The imagination-to-understanding ratio is not a phrase with a single origin but a concept that Alan Kay has articulated in various formulations across decades, most intensively since the emergence of LLMs in the 2020s. Kay's core concern — that computing tools should expand users' understanding, not merely their output — has been a constant of his work since his 1972 Dynabook proposal. The specific juxtaposition with Segal's imagination-to-artifact ratio is part of the ongoing conversation between the two frameworks in the post-2025 AI discourse.
Key Ideas
The companion problem. Collapse of production distance without corresponding work on comprehension distance produces capable users who do not understand what they produce.
Causal models require construction. Understanding is a causal model; causal models are built through cognitive labor; cognitive labor cannot be delegated.
AI widens rather than closes the ratio. Each generation of abstraction has grown the understanding distance; AI has grown it fastest.
Design implication. Tools should scaffold causal understanding alongside output generation, not just produce outputs.
Invisible in output metrics. The ratio cannot be measured by evaluating outputs; it requires testing whether users can reason about why outputs work.
Appears in the Orange Pill Cycle
Further reading
- Kay, A. 'A Personal Computer for Children of All Ages' (Xerox PARC memo, 1972)
- Kay, A. 'The Early History of Smalltalk.' ACM SIGPLAN Notices (1993)
- Gopnik, A. et al. 'Large AI Models Are Cultural and Social Technologies.' Science (2025)