
Writing & reflection · By Robert Metzger
Better text. Less confident? What “rewrite this” does to the writer
Half an hour of “rewrite this” can leave you with a better text and less confidence that you could have written it. What research says about editing versus asking for critique, and two prompts I would try instead.
The letter got better. You got less sure you could have written it.
Imagine a job you really want. Your cover letter is finished, but it is not good enough yet. So you paste it into a chat with an AI and type: “Rewrite this.” Then: “More convincing.” “Shorter.” “A bit more professional.”
Half an hour later, the letter is better than your draft. Probably better than anything you would have written alone. And that is exactly the feeling it leaves behind: on your own, you would not have managed this.
In my last essay, “More polite. Less firm?”, I argued that the assistant should have asked before it changed what a message asks for. This time the question runs the other way. What changes when we ask the AI something different?
Confidence down, trust up
Yeon Su Park and colleagues measured this while it happened. In a study presented at CHI 2026, 302 people wrote an argumentative essay with GPT-4.1, in a chat next to their text. Before every new request, they answered two questions: How confident are you that you could complete this writing task on your own? And how much do you trust the AI to help you with it?
On average, confidence fell with every turn, and trust rose with every turn, by almost the same amount. For more than a quarter of the participants, confidence dropped by at least two points on a seven-point scale within a single session. The median writing time was 23 minutes. Only 4 percent ended the session clearly more confident than they started. §4.2
The authors suspect a lopsided feedback loop: the AI keeps producing fluent text, so people get “frequent confirmation of the system’s competence but little confirmation of their own”. §5.1
What interests me most is how people asked. Those whose confidence fell had asked the AI to edit their text more often than those whose confidence held. Those who lost confidence and then regained it had asked more often than the falling group for something else: an assessment of their own writing. The example in the study’s own coding scheme is simply “What do you think?”. People whose confidence fell also ended up with more of the AI’s wording in their final essay, and they felt less like its authors. §4.3–4.4
How often did anyone ask for an assessment? Of 1,410 requests, almost half asked the AI to draft or edit text. Only 7.4 percent asked for a review. §4.1
Editor or author
A study with poems points in the same direction. Jack McGuire and colleagues gave one group a finished poem from an AI system, with tools to edit it as freely as they liked. Another group wrote alone. Ten published poets, who did not know how any poem had been made, rated the poems written alone as clearly more creative than the edited ones.
In a second experiment, a third group chose a theme and then wrote with the AI line by line, taking turns. Their poems were rated about as creative as those written alone. They also reported about as much confidence in their own creativity as the solo writers did. The editors reported clearly less. The authors’ conclusion is about roles: to benefit from AI, people should be co-creators, not editors. Studies 1 and 2
The AI in this study was an older poetry generator, not ChatGPT, and an eight-line poem is not a cover letter. But the pattern matches Park’s. When the AI writes and you edit, you stop writing. You become the editor of your own text.
“But the rewrite is better”
That is the obvious objection, and sometimes it holds. Sebastian Maier and colleagues compared an AI that rewrote people’s ideas for new products with an AI that offered alternatives and then asked questions while people developed the idea themselves. In their first experiment, the rewriting AI showed the larger quality gain over working without AI. In the second, the gains were about equal. But the questioning AI kept the ideas more varied, and people felt more ownership of them. In 18 of 20 conversations the researchers analysed in detail, people simply accepted the AI’s rewrite without working on it further. Results
And “better” is not guaranteed. Zenan Chen and Jason Chan had experts and non-experts write ad copy and measured quality by the clicks the ads received on social media platforms. According to their abstract, using the AI as a sounding board, giving feedback on copy people had written themselves, improved the ads, especially for non-experts. Using it as a ghostwriter brought no significant benefit and even hurt the experts. The authors trace this to an anchoring effect of the AI’s draft. Abstract
So the rewrite is not reliably the better text. And even when it is, the cost lands with the writer.
What I cannot claim
Park’s study is correlational. Maybe people who lose confidence ask for edits, rather than edits making them lose confidence. I have one more objection of my own: the more turns someone takes, the more chances they have to drop by two points, and the more AI text they have available to adopt. The study does not rule out that long sessions simply produce both. All of these studies are single sessions with short tasks.
And asking for critique is more work. In Maier’s study, the questioning mode was clearly more demanding than letting the AI rewrite.
The role I want to keep
In my research, I look at which role an AI takes in creative work, and which role that leaves for us. My hunch from that work: more AI can make a result better and easier to reach without making us any more confident that we could do it ourselves.
I also write stories. In one session, I asked an AI to grill me on a chapter of mine, one question at a time. It pointed out that my main character spent most of the chapter being carried along by events, so the shame I had planned for him felt unearned. I had not seen it. The chapter was mine, the critique came from the AI, and the fix came from me.
That is the difference I care about. A rewrite may give you a better text. Critique can give you a better text and leave you its author.
Two prompts to try
Next time you want help with a text that matters, try one of these before “rewrite this”:
What do you think? Where is my text weak? Don’t rewrite it.
Ask me one question at a time until it is clear what I want to say.
Then revise it yourself. It takes longer. Afterwards, you know that you wrote it.
Over to you
My thesis: we use AI far too often to rewrite and far too rarely to critique. Convince me otherwise.
Of course, asking for critique only helps if the AI is willing to give it. What if “What do you think?” gets you “This is a great letter!”? That is the next essay.
Sources
- Yeon Su Park, Nadia Azzahra Putri Arvi, Seoyoung Kim and Juho Kim (2026). Authorship Drift: How Self-Efficacy and Trust Evolve During LLM-Assisted Writing. CHI ’26. Publication DOI · Open full text, arXiv.
- Jack McGuire, David De Cremer and Tim Van de Cruys (2024). Establishing the importance of co-creation and self-efficacy in creative collaboration with artificial intelligence. Scientific Reports 14, 18525. Open access.
- Sebastian Maier, Manuel Schneider and Stefan Feuerriegel (2026). Partnering with Generative AI: Experimental Evaluation of Human-Led and Model-Led Interaction in Human-AI Co-Creation. CHI ’26. Publication DOI · Open full text, arXiv.
- Zenan Chen and Jason Chan (2024). Large Language Model in Creative Work: The Role of Collaboration Modality and User Expertise. Management Science 70(12), 9101–9117. Publication DOI.