Two Thirds of ChatGPT Writing Requests Edit the User's Own Text
In September 2025 a team of economists and OpenAI researchers published How People Use ChatGPT as an NBER working paper. It is the largest look so far at what people actually type into a chatbot, built by running a privacy preserving automated classifier over a representative sample of consumer conversations.
The headline numbers got the attention. By July 2025, 700 million people were sending 18 billion messages a week, around 10% of the world's adult population. But the finding I keep thinking about as a product designer is smaller and buried in the writing section.
About two thirds of all Writing messages ask ChatGPT to modify text the user provided: editing it, critiquing it, translating it. Only about a third ask it to create something from scratch.
Writing is not a niche use, either. The paper says it is the most common use case at work, accounting for 40% of work related messages on average in June 2025.
Most AI features are built for the other third
Think about the AI writing features that have been added to products in the last two years. Almost all of them start from a blank state. A text box that says "Describe what you want." A button that says "Generate." A magic wand on an empty document.
That is the third of the use. The paper says the larger share starts somewhere else. The person already has words. A draft email, a paragraph that does not sound right, a message in a language they do not read well, a report they need to shorten. They bring their own material and want help with it.
From the outside, a "Generate" button looks like the obvious AI feature, because it demos well. You type one line and a page appears. In real use, people mostly do not want a page to appear. They want their page to get better without losing what made it theirs.
Asking beats doing
The paper also introduces a simple way to classify what people want from a message. Asking is looking for information or advice to make a decision. Doing is asking the model to produce an output or perform a task. Expressing is neither, just sharing views or feelings.
Across all messages, about 49% were Asking, 40% Doing and 11% Expressing. Asking also grew faster than Doing over the year, and the authors report that Asking messages are consistently rated higher in quality, both by a classifier that measures satisfaction and by direct user feedback.
Work is different. About 56% of work related messages are Doing, and nearly three quarters of those are writing tasks. So at work, people want output. But the output they want is mostly writing, and the writing they want is mostly a change to something they already wrote.
I read that as a pretty clear brief. People use these tools to think and to improve their own work. They use them much less to replace the work entirely.
What I would design differently
Start from the selection, not the empty box. The most natural entry point for an AI writing feature is text the user has already selected, with a short set of actions that match what the paper found people do: tighten it, change the tone, check it, translate it. A blank prompt can still exist, but it should not be the front door.
Show the edit as an edit. If the model is changing someone's text, the user needs to see what changed, not receive a new version and compare it by eye. Tracked changes have existed in word processors for decades for a reason. A rewrite that hides its changes makes the person read the whole thing again, and that is where trust leaks out.
Keep the voice. When someone brings their own text, the thing they are most afraid of losing is how they sound. The default behavior for a rewrite should be the smallest change that solves the problem, with a bigger rewrite available on request. Most AI writing tools do the reverse.
Treat critique as its own feature. The paper groups critiquing with editing, and I think it deserves its own surface. A lot of people do not want their text rewritten at all. They want to know what is weak about it and fix it themselves. That is feedback on the text, and very few products design for it.
Make translation a first class action. I live in Tokyo and switch between English, Spanish and Japanese most days. Translating my own writing is one of the most ordinary things I ask these tools to do, and the paper lists it among the main ways people modify their writing. Yet in many products it is buried in a menu, if it exists at all.
The small numbers are worth reading too
Two more figures from the paper are useful for anyone deciding what to build.
Only 4.2% of ChatGPT messages were about computer programming. If you only follow developer conversations, you would assume coding is the main use. On consumer ChatGPT, it is a small slice. The paper itself notes that other assistants show much higher coding shares, so the product and its audience shape the mix.
And only 1.9% of messages were about relationships and personal reflection. Companionship is a real use, and I have built in that space, but on a general assistant it is a small one. If you are building for it, you are building for a specific audience that chose that product for that reason.
Non work use also grew faster than work use, from 53% of messages in June 2024 to more than 70% a year later. Whatever we design for the office, most of the time these tools are in someone's personal life.
Where this leaves product teams
The pattern in this data is that people come to AI with something already in hand. A question they are trying to answer, or a piece of writing they are trying to improve. They want help with their thing.
The products that respect that will feel like tools. The ones that keep starting from a blank box and a generate button will keep demoing well and getting used less than their teams expected. If you are designing an AI writing feature right now, start by asking what the user already has, and how the model can make it better without taking it away from them. What the model can produce from nothing is the less interesting question.