Ask for the edit, not the rewrite

Overview

On 31 July, LinkedIn announced it was deleting its own AI writing button. Hari Srinivasan, chief product officer for the LinkedIn ecosystem, described the change in plain terms: the company is “removing the ‘enhance your post’ feature you see when you write a post or message & replacing with a feature that proofreads your words, but does not change your voice.”

A billion-user platform just decided those are two different products. Most of us still type them as one instruction.

“Make this clearer and more professional” and “fix the grammar” feel like two settings on a single dial — a bit of help, or a lot of help. They are closer to two different operations, and there is now a measurement of what the bigger one costs.

What you’ll do differently: name the defect you actually want fixed, instead of asking for a general improvement.

The content

Start with what was measured, and be precise about what did the measuring. A preprint posted on 2 August 2026 by Ushna Malik and Moiz Sadiq Awan takes an authorship-attribution model — software trained to pick an author out of a lineup by their writing habits — and asks how much of that authorship signal survives an AI rewrite. They call the drop the Idiolect Erasure Rate.

On a corpus of workplace email, the attributor identified the right author about 7 times in 10. After one “make this clearer and more professional” pass, it managed fewer than 2 in 10 — a drop of 52.5 percentage points. Grammar-only editing barely moved the simpler surface-stylometry attributor at all on that same email corpus: 3.7 points, which the authors report as not statistically significant. And on the blog corpus, a plain non-generative grammar corrector removed only 15 points where generative rewriting removed 66.5, so, in the paper’s words, “the effect is specific to generative rewriting rather than any edit.” The pattern replicated on two commercial assistants, at 53.8 and 52.5 points on the email corpus.

The obvious workaround was tested, and this is the sentence worth carrying: “Explicitly instructing the assistant to preserve the author’s voice reduces surface-level erasure but leaves most of the deep authorship signal unrecovered.” The authors’ own heading is blunter — voice-preservation prompts are insufficient.

Now the objection you should be forming, because it is a good one. If this only measures a machine, and I only write for humans, who cares?

Be clear about the answer: nobody has shown that a colleague would notice. The paper says so itself — “IER measures computational attributability, not human recognition. Reduced attribution therefore indicates weaker authorship signals, not necessarily lower recognition by familiar readers” — and it runs no human study at all. The evidence that does exist points away. Nineteen professional writers co-wrote with personalised and non-personalised AI tools in a study published at CSCW 2025; the 30 readers surveyed “could not distinguish AI-assisted work, personalized or not, from writers’ solo-written work.” In six experiments with 4,600 participants, Jakesch, Hancock and Naaman found people simply “were unable to detect self-presentations generated by state-of-the-art AI language models” in professional, hospitality and dating contexts.

So the cost does not arrive as detection. It arrives as suspicion — and suspicion does not need to be correct about any particular message.

That is where the strongest evidence in this file sits, and it is peer-reviewed. Cardon and Coman surveyed 1,100 working professionals for the International Journal of Business Communication, showing them a workplace message — a manager congratulating a team — at four levels of AI assistance. Their finding, in their own words: managers “put their trustworthiness at risk when using medium- to high-levels of AI assistance, as respondents in these conditions begin to question the authorship, confidence, caring, sincerity, and ability of senders.” The University of Florida’s write-up of the study reports the numbers behind that. Rated at low assistance, 83% judged the sender sincere; at high assistance, 40–52%. Professionalism fell from 95% to 69–73%. And it draws the practical line this module runs on: “While low levels of AI help, like grammar or editing, were generally acceptable, higher levels of assistance triggered negative perceptions.”

A separate set of three experiments on freelance-writer profiles found that “perceptions of AI use negatively impacted writing evaluations and hiring outcomes across the board” — the penalty attaches to being suspected, not to being caught.

Two operations, not one setting.

Then hold the finding to its actual size, because the preprint’s title oversells its own result. “Erased” is too strong: after a heavy rewrite the attributor still lands on the right author far above guessing, and pooled across a body of writing, re-identification “saturates at approximately 50%” — half of you survives. The email figure also falls from 52.5 points to 39 once sign-offs and quoted text are stripped out, which means some of what looked like voice was your signature block. The email interval is wide, too: 31 to 71 points. The corpora predate generative AI entirely — early-2000s email and blogs, not modern Teams and Slack. No frontier model was tested. It is a five-page preprint, days old, with no peer review behind it. And the authors decline to call the effect bad: “preserving a recognizable voice may support authenticity and accountability, while reduced attributability may benefit privacy or anonymity.”

Which is the honest place to land. This is not a rule against rewriting. It is a reason to notice that you have been buying the large size by default.

Try it

Five minutes, on a real message you’re about to send. Works anywhere there’s a rewrite box — Copilot in Outlook, Gemini in Gmail, your workspace assistant, plain chat.

1. Say the defect out loud before you type anything. Not “improve this” — what’s actually wrong with it? Too long. Three typos. The second paragraph buries the ask. That sentence is rude. You almost always know.
2. Ask for that, and fence off the rest. Try: “Fix spelling, grammar and punctuation only. Do not change my word choice, sentence structure or tone. Return the text with the changes listed underneath.” The list is the useful half — it shows you what you actually do wrong, which the rewrite never does.
3. If you do want the bigger version, ask for it as advice, not as output. “Tell me the three things that make this unclear. Don’t rewrite it.” Then fix them yourself.
4. Where you take the whole rewrite, retype the final pass. Not superstition — it’s the one step that puts your habits back in.

Where this breaks. Plenty of writing shouldn’t sound like anyone: a compliance summary, a standard status update, a policy paragraph. Ask for the full rewrite there and don’t feel clever about it. Nor is “keep my voice” a fix — that was the instruction the paper tested and found insufficient, so if the request matters, make it about what to change rather than what to protect. And be careful with the strong version of this: the measured penalty is what people think when they suspect heavy AI use, and no study here shows your team can tell. Treating a colleague’s polished email as evidence of anything is exactly the reasoning this module doesn’t support.

The point isn’t purity. It’s that “make it better” is a much larger instruction than the one you usually mean.

Additional reading

• The Assistant Erased You — Malik & Awan, 2 August 2026. The source of the 52.5-point email figure, the light-versus-heavy gap, and the tested-and-failed voice-preservation prompt. Read the limitations: five pages, no peer review, corpora that predate generative AI, and an explicit statement that the metric is not about human recognition.
• Professionalism and Trustworthiness in AI-Assisted Workplace Writing — Cardon & Coman, International Journal of Business Communication, July 2025. Peer-reviewed, 1,100 professionals, a constructed workplace email at four assistance levels in a two-by-four design. The strongest evidence in this module, and the only peer-reviewed one measuring how colleagues judge a sender. The University of Florida summary (6 August 2025) carries the sincerity and professionalism percentages if the journal is paywalled for you — note those figures come from the release, not the abstract.
• Hari Srinivasan’s post — LinkedIn’s chief product officer for the ecosystem, 31 July 2026. Worth reading past the “seems like AI slop” button for the line most coverage skipped: “AI and slop are not the same thing; many people refine thoughts with AI.”
• “It was 80% me, 20% AI” — Hwang et al., preprint November 2024, peer-reviewed as CSCW 2025. Nineteen professional writers, 30 readers, and the finding that cuts against the alarming reading of this module. Note the scope: creative writing, not workplace email.
• Human heuristics for AI-generated language are flawed — Jakesch, Hancock & Naaman, PNAS 2023. Six experiments, 4,600 participants, and the reason to be sceptical of anyone claiming they can spot it.
• Generative AI and Perceptual Harms — Kadoma, Metaxa & Naaman, preprint October 2024, peer-reviewed as CHI 2025. Three experiments establishing that the cost of suspected AI use lands whether or not the suspicion is right.

Editor’s note

Obvious objection acknowledged, it’s important for me to state clearly that I believe the context matters most if you’re choosing to use AI to deliver a piece of work. An expression in a forum where your opinion is the driving force behind the work’s creation (like a LinkedIn post, or indeed this Editor’s Note) carries an expectation that the words are yours. An edited article synthesising a research sweep is naturally a more forgivable work of AI authorship. The purpose of this module is to turn your attention to the judgement that underpins a decision to use AI more or less heavily in the production of your work. I predict that this will be a source of increasing scrutiny and anxiety in the future, and this site reflects a personal attempt at improving my own judgement in that domain.