← Journal
A person in a teal sweater writing an email on a laptop at a wooden desk beside a sunlit window

More polite. Less firm? The decision inside an AI rewrite

Asking AI to make a message more polite can quietly change what it asks for. Why I think the assistant should have asked, and what research says about suggestions passing on direction, not just words.

You wanted to sound more polite. Did you also want to sound less firm?

Imagine a workshop that did not deliver what its description promised. You want your money back, and you have written one plain sentence:

I’d like a refund.

It feels blunt, so you ask an AI assistant: “Make this more polite.” What comes back reads well:

The workshop did not quite meet my expectations. I hope we can find a solution.

You don’t want to send something that sounds like a machine, so you rewrite it in your own words:

The experience wasn’t what I expected. I’d appreciate your help in finding a way forward.

Every word is now yours. The direction is still the AI’s. And the request for a refund is gone.

The decision inside “more polite”

“Make this more polite” sounds like a request about wording. It is more than that. Inside it sits a decision that only the writer can make: how firmly do I want to ask?

A refund is a specific request. “A solution” invites the other side to offer something else: a voucher, an apology, a discount next time. That might be fine. But it is a different message, and nobody asked whether you wanted to send it.

In my research, I call this the difference between handing over the execution and handing over the judgement. Letting a tool do the execution is fine: finding a friendlier phrase, smoothing a sentence, shortening a paragraph. The judgement about what a message is for should stay with the person sending it. The trouble with “more polite” is that this particular piece of execution contains a judgement. The tool cannot change the tone without also deciding how much of the request survives.

Nishal and colleagues found a similar line among professional writers. They interviewed 20 science journalists about four hypothetical writing tools, each giving the AI a different role. Help with gathering information or getting feedback was broadly welcome. Deciding what goes into the piece was not something they wanted to give away. The authors describe this as keeping decision-making control while handing over execution control. They also found that the exact line differs from person to person and from situation to situation. These were reactions to tool concepts, not observations of everyday use. Still, the pattern fits the refund example: the work was fine to hand over, the decision was not. Discussion, §6.1

Suggestions pass on direction, not just words

Rewriting a suggestion in your own words feels like taking back control. A study by Vitor Welzel and Nicholas Vincent, presented at FAccT 2026, helps explain why that is often not enough.

Forty-seven participants each wrote two short texts about O. Henry’s story After Twenty Years: an analysis of a character’s decision and a creative rewrite of the final scene. For one of the two tasks, they received an AI suggestion. Afterwards, 37 of the 47 said they had used it. Of those 37, 22 said the suggestion gave their text a direction. Only 10 mentioned wording or paraphrasing. And 28 of the 47 said they would have written something different without it. Results, §4.1

So people noticed. Direction was what they took most often, and they knew it. This is not a story about hidden manipulation. It is a story about an easy handover: the direction arrives together with the suggestion, and accepting it costs nothing.

The researchers also compared the texts with the suggestion itself, using texts written without help as a baseline. Assisted texts were closer to the suggestion in vocabulary, in patterns of words and word classes, and in sentiment. Similarity in meaning, measured with text embeddings, was not clearly higher. The pattern came mostly from the analysis task; in the creative task, only sentiment was clearly closer. The authors treat these tests as exploratory, and they are careful to say that similarity alone does not tell us whether the help was good or bad. Results, §4.1, Tables 1–2 · Discussion, §5

One detail matters for the refund example. In the creative task, the suggestion was not a finished text but a plan: a sequence for the scene, ending in resolution and reflection, and the idea of reading the character’s choice through duty and compassion. Appendix A That is a short outline, and it already contains the interpretation the whole piece turns on. Fewer words do not mean fewer decisions.

This picks up the question from my previous essay, “When the words arrive before the meaning”: once a suggestion is on the page, what does it bring besides words?

What the assistant should have done

My thesis is simple: the assistant should have asked.

Not with a questionnaire. With one question, at the moment the rewrite would change what the message asks for:

You asked for a refund. Should I keep that request firm, or are you open to other solutions?

Or it could have pushed back: “A softer version may read as an invitation to negotiate. If you want your money back, I would say so directly.” Either response would have made the decision visible. Both would have left it with you.

A polite version does not have to give anything up:

The workshop did not match its description. I would therefore like to request a refund.

Polite and firm are not opposites. The assistant treated them as if they were, and did so silently.

That silence is the thread I want to follow in the coming posts. A useful writing partner does not only carry out instructions. Sometimes it asks. Sometimes it disagrees.

Three questions before accepting a rewrite

  1. What should this text achieve? Write it down in one sentence before asking for help. “Get my money back” is a purpose you can check a draft against.
  2. What has the suggestion decided for me? Look past the words: emphasis, firmness, what was left out.
  3. Which of those decisions do I want to hand over? Sometimes the honest answer is “all of them”. That should be a choice, not an accident.

Over to you

My thesis: the AI should have asked. Or is that exactly the judgement you would rather hand over?

Sources

  • Vitor H. A. Welzel and Nicholas Vincent (2026). Overreliance in Writing Tasks: Exploring Similarity-Based Measures of AI Influence on Writing and Proposing a Reflective Writing Interface Intervention. FAccT ’26. Publication DOI · Open full text, arXiv v1.
  • Sachita Nishal, Mina Lee, Nicholas Diakopoulos and Jennifer Wortman Vaughan (2026). “Helping Me Versus Doing It for Me”: Designing for Agency in LLM-Infused Writing Tools for Science Journalism. CHI ’26. Publication DOI · Open full text, §6.1.