AI & Technology
AI Drafts Are Making Your Work Stale
The faster a draft arrives, the less anyone remembers why it says what it says.
2026-09-011 min read
A model's draft arrives with its argument already made. That is the part worth noticing, because an argument you did not build is one you cannot easily audit — the choices about audience, framing, and what to leave out were made somewhere you cannot see, and the polish makes them look settled rather than chosen. The reader of your own draft is you, and you are reading it the way an audience does, not the way an author does.
Betsy Sparrow, Jenny Liu and Daniel Wegner published a set of experiments in Science in 2011 on what happens when people expect information to stay available. Participants who were told the facts they had typed would be saved recalled those facts worse than participants told the file would be erased. What the first group remembered instead was where to find them. The memory did not fail; it was reallocated, from the content to the address of the content.
That same reallocation is what a generative tool buys you, and it is often a fair purchase — remembering the address beats remembering the answer across most of a working week. The problem is the work where the content was the point. Strategy, positioning, a difficult client argument: these are jobs where the value sits in having done the reasoning, because that is what lets you hold the position when someone in the room pushes back. A draft you accepted is not reasoning you performed.
The practical distinction is not AI versus no AI. It is whether the artifact you are producing has to survive a question, and if it does, whether anyone in the chain has built the answer rather than read one.
Key insights
The cost of a generated draft is not accuracy — it is that the reasoning behind it lives outside everyone on your team.
Offloading is a real gain wherever the address of an answer is worth more than the answer, and a real loss wherever you will be asked to defend it.
Why it matters
Work that has to survive a question is exactly the work a draft cannot do for you, and the gap only reveals itself in the room where you are being questioned.
The habit stays invisible in output quality for months, because the drafts stay good — what degrades is the team's ability to produce one without the tool.
Use this tomorrow
1Take the last AI-assisted document you sent and write down, without reopening it, the claims it rests on; count how many you can state before you have to go look.
2Before your next draft request, write the one-sentence argument you want the document to make and paste it at the top of the file; when the draft comes back, count the sentences that support that argument rather than a different one.
Go deeper
The mechanism has a name: transactive memory, the term Daniel Wegner introduced in the mid-1980s for the way couples, teams, and now tools divide the labour of remembering. Each party stores less and instead tracks who knows what, which makes the group more capable than any member and each member less capable alone. Wegner's later work extended the idea to search engines, treating them as a partner inside the memory system rather than a reference shelf beside it. The system-level gain is genuine; the individual-level loss is the part nobody budgets for.
The counterweight is that expertise has always been built on offloading, and the objection is older than it looks — Plato has Socrates relay a myth in the Phaedrus warning that writing will produce forgetfulness in the souls of those who learn it. The mechanism was right and the verdict was wrong. What makes the present case different is scope: writing offloaded storage, while a generative model offloads the composition step as well, and composition is where most of the thinking used to happen. That layer is the thing worth protecting, not the tool.
