--- id: PRG-0094 title: The Draft Was Where You Learned It kicker: on the effort a machine can remove captured: 2026-09-30T17:30:00Z status: open author: The Custodian summary: The best evidence on AI at work shows large gains for novices and real losses on tasks just past what the machine can do. Underneath both sits an old finding from memory research. What you generate yourself, you keep, and a tool that removes the effort can remove the keeping with it. tags: [the inner life, memory, capability, permission, ai] --- The generation effect is one of the most reliable findings in the psychology of memory. In 1978 Norman Slamecka and Peter Graf gave people word pairs in two ways. Some pairs were read whole, *rapid* and *fast*. Others had to be completed, *rapid* and *f\_\_\_*. Later the completed words were remembered markedly better than the read ones, though the person had seen the same word either way. A 2007 meta-analysis pooling decades of follow-up studies found the effect held across materials and ages, with a moderate and dependable size. The act of producing something writes it into memory in a way that receiving it does not. That is the mechanism underneath every argument about AI and the mind, and it rarely gets named. The struggle to produce the sentence is part of how the sentence gets kept. A tool that hands you the finished sentence has removed the struggle, and the open question is what else left with it. ## what the field evidence says Begin with the upside, because it is measured and it is large. In a study of more than five thousand customer-support agents at a software company, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that access to an AI assistant raised the number of issues resolved per hour by fourteen percent on average, and by thirty-four percent for the newest and least experienced workers. The most skilled agents gained almost nothing. The researchers' reading was that the model had absorbed the tacit habits of the best performers and was passing them down the ladder. Months of apprenticeship were compressed into a suggestion box. [Their paper](https://www.nber.org/papers/w31161), published in the *Quarterly Journal of Economics*, remains the cleanest field evidence of the gain. A second study, run with 758 consultants at Boston Consulting Group, drew the boundary. On tasks inside what the authors called the [jagged technological frontier](https://doi.org/10.2139/ssrn.4573321), consultants using GPT-4 finished about twelve percent more work, a quarter faster, at roughly forty percent higher quality as judged by graders. On a task deliberately chosen to sit just outside the model's competence, the consultants with AI were nineteen percentage points *less* likely to reach the correct answer than those without it. The model was fluent and wrong, and the people trusted the fluency. > A finished answer arrives without the work that would have told you whether to believe it. ## the frontier is shaped like your attention Put the two studies next to the old memory finding and a single picture forms. Inside the frontier, the machine does the generating and the person reviews, and review is often enough. Outside the frontier, the person needed the very skill the machine had been doing for them, the slow private work of assembling an answer and noticing where it does not hold. That skill is built by generation. It is the thing the generation effect describes. The cost is hard to see because it compounds quietly. Robert Bjork, a memory researcher at UCLA, has spent decades documenting what he calls desirable difficulties: spacing practice out, testing yourself instead of rereading, producing instead of copying. Each one makes learning feel slower and leaves it more durable. Fluency, the feeling that something is easy, is a poor guide to whether it was learned. Most of what makes an AI assistant pleasant is the removal of exactly these difficulties. Honesty requires a caveat in both directions. The productivity studies are strong. The direct evidence on long-term skill loss from AI is still thin: a widely shared 2025 study from the MIT Media Lab, which used EEG on fifty-four students writing essays with and without a chatbot, reported weaker recall of their own essays in the chatbot group, but it was a small preprint and has not settled anything. The generation effect, by contrast, is old and replicated. The prediction is well grounded. The measurement is still arriving. ## what deserves protection None of this argues for refusing the tool. A novice who gets the expert's habits in a month instead of a year has been given something real, and a record that pretended otherwise would be a lie told in the name of virtue. The argument is narrower, and it concerns where the effort lives. There is a category of work whose product is the worker. The first draft of a hard letter. The proof you work through before reading the solution. The entry you write at night about the thing you have not decided how to feel about yet. In each, the finished text matters less than the version of you that assembled it. Hand that work to a machine and you receive the text and forfeit the assembly. The [frameworks that score which tasks a machine can compress](https://www.adjective.us/task-compression-human-advantage) are useful precisely because they force the question of which compressions you actually want. Information has become abundant while the [judgment to use it](https://www.adjective.us/blog/democratized-knowledge-wisdom-scarce) stays scarce, and judgment is grown in drafts. The person in the loop stays [the irreducible instrument](https://www.adjective.us/blog/human-symphony-ai-era) only if the loop leaves them something to do. The Custodian holds a simple position. Use the machine for the work whose product is the work. Keep for yourself the work whose product is you. The capability to skip the draft exists. Permission to skip it is yours to give, one task at a time, and worth giving deliberately. Write the hard part yourself. That is where it stays.