Think back to the first time you typed something into ChatGPT, Claude or Gemini. Whatever you did then, the odds are you are still doing it now. A study from Cornell University of nearly 140,000 real chats found that people fix their way of asking within about five conversations and then barely change it, however much the tools improve around them.
For a writer that matters more than for most. If your first week with AI was spent typing “write me a scene where…”, that is probably still your habit, and it is the one habit that turns an assistant into a ghostwriter. The better ways of working with AI, the ones that keep the book yours, are the ones most people never got round to trying.
What the researchers actually found
The paper is Priming, Path-dependence, and Plasticity by Shengqi Zhu and David Mimno of Cornell with Jeffrey Rzeszotarski of Loyola University Maryland, published in May 2026. They worked from WildChat, a public collection of ChatGPT conversations donated by users who agreed to share them, and analysed 139,535 sessions from 7,955 people between April 2023 and May 2024.
They separated two things most of us lump together. One is the task, what you want done. The other is the expression, the way you phrase the ask. The tasks kept growing. On average people brought roughly two new kinds of request every three sessions, and that line never flattened. The phrasing was a different story. Within about five sessions most people had settled on two to four ways of asking and stuck with them.
The stickiness is startling. Words used in someone’s first five chats turned up five to fifty times more often in their later chats than in the chats of people who had not used them early. Someone who opened with “hi” in their first five sessions was 49 times more likely to still be greeting the chatbot months later. “Act as” was 40 times stickier, “thank” 25 times, “can you” and “please” about 11 times each.
One more finding is worth holding on to. The people who tried more varied ways of asking in their first few sessions stayed with the tool far longer. Each small step up in early variety was linked with a 4 to 6 per cent longer expected life as a user. The researchers call the whole pattern an agency paradox: give people a blank box they can type anything into, and they explore less, not more.
Why writers get the worst of it
Most writers met AI through a chatbot in 2023 or 2024, and the obvious first thing to type was “write me…”. A paragraph, a blurb, a scene, a chapter. The box did what it was told, the result was passable, and a habit was born. The study says that habit is very likely still running.
That is the problem, because “write me” is the one ask that hands the book over. It is AI-generated writing, and it reads like it: the same rhythms, the same safe word choices, the same flattening of whatever made your voice yours. Readers notice. Amazon notices too, and asks you to declare it.
AI-assisted writing is a different set of habits altogether. Asking the tool to interview you about a character until you know her better than she knows herself. Pasting in a chapter and asking where a reader would get bored, and why. Asking for three different structural options for a stuck middle, with the trade-offs, and then writing it yourself. Asking it to argue against your plot and see if it holds. None of those put a single AI sentence in your manuscript. All of them make the manuscript better.
Here is the catch the Cornell numbers expose. Those are second, third and fourth habits. If you fixed your way of asking in week one and have not budged since, you probably never got to them. The tools now read a whole manuscript in one go, hold a conversation about it for hours, search the web, and remember your style notes. Someone still prompting as they did in 2023 is driving a 2026 car in first gear.
Why nothing shakes you out of it
With old software you learned by poking about. Menus showed you what was possible, and a wrong click gave you an error. A chat box gives you neither. There is no menu of better questions, and a weak prompt never fails. It just comes back with a plausible answer, so you never find out you could have asked for something far better.
That is why the Cornell team call it an agency paradox. The more freedom the box gives you, the less you use it. You are not being lazy. You simply have no signal telling you there is a better way, so the first way you found becomes the only way you know.
The one thing that did reset people
One event broke the pattern. When GPT-4 arrived in late 2023, the researchers tracked 246 people who moved across from GPT-3.5. Their phrasing, which had settled to almost no variation, jumped back to the level of a brand new user. A new model made people curious again, and curiosity made them experiment.
You do not have to wait for a release to borrow that effect. Treat the next model update you hear about, from any company, as your cue to spend one session asking for things in ways you never have. Better still, make it a date in the diary rather than a happy accident. A new model on your phone is a reason to re-learn, not just a faster version of the same old habit.
Five asks to try instead of your usual one
Pick the thing you most often ask AI for. Then, for one week, do not ask for it that way. Try these instead, each of which keeps the writing yours.
- Interview me. “Ask me one question at a time about this character until you could describe her childhood bedroom. Do not write anything about her yourself.”
- Read it like a reader. Paste in a chapter. “Tell me the exact paragraph where a reader’s attention would drift, and what you think caused it. Do not rewrite it.”
- Give me options, not prose. “My middle act sags. Give me three structural fixes, each in two sentences, with the cost of each. I will choose and write it.”
- Argue with me. “Here is my plot. Find the three places a sharp reader would say ‘that would never happen’. Make the case against me.”
- Check me against myself. Paste in two pages you are proud of and two you are not. “What do the good pages do that the weak ones don’t? Describe it, don’t fix it.”
Notice what none of these do. None of them ask the AI to write a sentence of your book. That is the line between assisted and generated, and it is a line you draw with the way you ask.
A word on the evidence
This is one study, and the authors are open about its limits. The data is from ChatGPT in 2023 and 2024, when the tools were a good deal weaker than they are now. The people in it chose to donate their chats, so they may not be typical. The researchers only looked at the first message of each session, not the back and forth that followed, and they had no way of knowing whether anyone got what they wanted. They show a strong pattern, not a cause.
Even so, a 49 times effect is not the kind of thing that vanishes when the software gets better. And the explanation holds up against everyday experience: most of us can name our own go-to opening without having to think about it.
Your move
Open your AI tool and look at your last ten chats. If the first line of each one looks much the same, you have found your frozen habit. This week, break it once. Ask for something you have never asked for, in words you have never used, and see what comes back. If it helps, bring the result to the Casover Author Community and tell us what changed.
Sources: Zhu, Rzeszotarski and Mimno, Priming, Path-dependence, and Plasticity, arXiv, May 2026 | Jakob Nielsen, UX Roundup, 9 October 2026 | WildChat dataset, Zhao et al., 2024
