Every: writing with AI is harder than you think

Published in Every’s Working Overtime newsletter in April 2026, this piece by Katie Parrott addresses a gap in how AI-assisted writing is discussed. Most coverage focuses either on AI as a replacement for writing skill or on how to prompt more efficiently. Parrott’s argument is different: meaningful AI collaboration requires more rigor and judgment than writing without it, not less.

Context: company, task, scale

Parrott is a writer at Every, a publication that has been public about using AI throughout its editorial process. The piece reflects practice developed over sustained daily use, not a one-time experiment. Every’s editorial workflow involves multiple passes — structural editing, line editing, and a top edit — and the article describes how AI fits into this process at each stage.

The core argument

The perception problem Parrott identifies is that critics “imagine the laziest possible version of AI-assisted writing” — pasting a prompt and publishing the output. Her workflow is structurally opposite. She uses AI to stress-test her own thinking before drafting: structured interviews in which AI asks clarifying questions, negotiated outlines rather than accepting initial structures, and multiple revision passes using specialized personas she has developed (a Hemingway persona for economy, a Hitchcock persona for pacing).

The point is not that these personas produce better prose on their own. They expose weak sections of the draft that a writer can then fix, which requires knowing what good writing looks like in the first place. The article makes an implicit argument that the quality of AI output depends entirely on the writer’s capacity to evaluate it.

Key method

Parrott describes a specific problem that arises in AI-assisted drafts: a kind of machine polish that smooths out the texture of genuine thinking. Her practice involves deliberately “roughing up” AI-improved passages to restore idiomatic phrasing and specific detail. She also applies AI-pattern detection to identify phrases and sentence structures that mark generated text, then rewrites them.

Her test for whether a collaboration is working is whether the published result contains writing worth reading. Speed gained in drafting does not count toward that test.

Who it is useful for

Writers, journalists, and editors who have experimented with AI tools and found the output mediocre, and want to understand what a demanding practice looks like rather than what tools to use. The piece does not name specific platforms. It assumes the reader already has access to AI writing tools and wants to use them without compromising the quality of their work.