Every: how Katie Parrott learned to stop accepting AI output on autopilot
Katie Parrott, a staff writer at Every, opens this essay with an anecdote: she sent a client a finished article, then discovered she had already completed the same assignment four weeks earlier and forgotten it. AI had been involved in both drafts, and somewhere between the tool’s confident output and her own review, the question of whether she had seen the material before never surfaced.
The essay is about why that happens and how to work differently.
What AI autopilot is
Parrott draws on research about fluency bias — the finding that text that reads smoothly is judged as more likely to be true — and on studies of automation in aviation, where operators tend to accept system output without checking it against their own judgment. Applied to AI-assisted writing, the problem isn’t that the model thinks for you; it’s that the output’s polish makes you feel you already thought. Reviewing becomes a feeling, not an act.
Three practices
The first is to write your own rough position before looking at what the AI produced — even a few bullet points. This gives you something concrete to compare against rather than starting from the model’s framing.
The second is to build a gap between drafting and reviewing: writing one day and reading the next, or reading output in a different context from the one in which it was generated. The goal is to interrupt the fluency effect by making the text feel less familiar.
The third is to write one sentence justifying each AI recommendation or paragraph — explaining why it is right for this specific client, argument, and reader. If you can’t write that sentence, look at the passage again.
Who it is useful for
Writers and editors who use AI as a drafting assistant and want a practical account of where the judgment risks actually live. The piece doesn’t argue against AI use; it argues for keeping the judgment work for yourself while still benefiting from the speed. Parrott’s framing is worth reading alongside any organization’s AI style guide, because the risk she describes is behavioral, not policy-level.