# Recall-to-judgment shift

The claim, from Jeremy Osborn's CACM piece [[programming-differently-difficult]]: AI coding assistants remove the retrieval half of programming and leave the evaluation half, so the hard question changes from "how do I write this?" to "does this actually make sense?" The total cognitive burden doesn't fall — it moves from long-term-memory retrieval to reasoning.

## Why relocation and not reduction

Programming was historically bounded by memory. Decades of studies — Brooks, Pennington, Soloway, and Siegmund's fMRI work — establish that the job runs on working memory plus long-term recall of syntax, idioms and patterns, and that the mental models involved are expensive to build and fragile under context switching.

An assistant absorbs the retrieval side. You no longer need precise recall of an API's shape to use it. What it cannot absorb is deciding whether the thing it produced is right, because that requires a model of the system it's being inserted into. So the work that remains is the work that was always harder: architectural reasoning, impact analysis, and judging semantic correctness in code that is already syntactically valid.

Two second-order effects make the accounting worse than it first looks. Generation gets faster while validation gets slower and more frequent, so time shifts rather than vanishing. And the evaluative capacity being drawn on is built by exactly the hands-on work now being skipped — see [[skill-atrophy-supervision-paradox]].

## The accessibility paradox

The most useful consequence: **the barrier to producing code falls while the barrier to producing good code rises.**

Both halves are real, which is why arguments about whether AI democratizes programming keep failing to resolve. People who would have washed out on syntax and error-handling idioms now have an entry path — that's genuine. But the skill that now gates quality is judgment, and judgment is harder to teach and slower to acquire than recall ever was. There's no equivalent of drilling flashcards for "is this architecture appropriate."

The evidence Osborn cites points the same way: Shihab et al. found students finishing brownfield tasks faster with Copilot while reporting they didn't understand why the suggestions worked, and Alanazi's meta-analysis found efficiency gains with only small, unstable gains in learning. Faster completion, no clear gain in the capacity the completion is supposed to be building.

## The medicine analogy

Diagnostic instruments lowered the burden of memorizing obscure clinical detail and raised the stakes on interpretation and error detection. Expertise moved from recall to reasoning. Nobody argues that made medicine easy — it made it differently difficult, and it changed what medical training had to produce.

## Where it shows up elsewhere

This concept is the cognitive-science floor under several practice-level arguments in the wiki that were reached by other routes:

- [[control-the-ideas-not-the-code]] and [[vibe-engineering]] both prescribe abandoning line-level review to spend the attention on design. That only works if judgment about design is the scarce resource, which is exactly this claim.
- [[reviewing-ai-code]] takes the same premise to the opposite conclusion: if validation is now the bottleneck, and [[code-review-throughput-limits]] cap validation, then output is capped too.
- [[llm-as-average-democratizer]] is the same shift seen from the output side — when plausible average work becomes free, discriminating good from plausible is what's left.
- [[programming-as-theory-building]] is Naur's much earlier version: the program is a theory in the programmers' heads, and only the artifact transfers.

## Limits of the framing

It assumes judgment is separable from recall, and the evidence for that is thin. The Soloway result cited in its own support cuts the other way — programmers did better when a construct matched an internal plan retrieved from long-term memory, which suggests judgment partly *is* accumulated recall. If so, a generation that never builds the recall may not develop the judgment that's supposed to replace it, and the accessibility paradox gets sharper rather than resolving.

## Cross-references

- [[programming-differently-difficult]] — the source
- [[skill-atrophy-supervision-paradox]] — what erodes the judgment this depends on
- [[measuring-ai-coding-productivity]] — why the efficiency-vs-learning split resists measurement
- [[recall-to-judgment]] is the individual-cognition twin of the organizational shift in [[agent-principal-agent-problem]]