Here is the number that stopped me. By July 2025, ChatGPT had more than 700 million weekly active users sending around 18 billion messages a week. That works out to roughly 10% of the world’s adult population.
A tool that did not exist in any form three years earlier was touching one in ten adults on the planet every week.
The user count is the figure the headlines reached for, and I understand why. It is the easy one to picture. But the more interesting thing buried in the data is not how many people are using ChatGPT. It is what they are using it for.
Before I go further: I am not a researcher, an economist, or a data scientist. This is one curious person reading a single working paper and thinking out loud about it. The paper below is a non-peer-reviewed study of a particular sample over a particular window, not a settled law about how everyone uses anything, so hold the numbers as findings rather than gospel.
Most of it isn’t work
The big study here is NBER Working Paper, “How People Use ChatGPT,” released in September 2025 by a team including David Deming of Harvard and OpenAI’s economic research group. They built a privacy-preserving pipeline that classified a representative sample of consumer conversations without any human reading the actual messages, covering roughly the period from May 2024 to June 2025.
The central finding is a shift, not a snapshot. The authors write: “We find steady growth in work-related messages but even faster growth in non-work-related messages, which have grown from 53% to more than 70% of all usage.” Over the study window, the share of messages that had nothing to do with anyone’s job climbed past seven in ten.
I find that quietly remarkable. The story we were told about this technology is often a workplace story. Productivity, automation, the office. And the office part is real. But the thing people actually do most, at least when you watch a representative slice of them, is bring ChatGPT into the rest of their lives.
What ‘non-work’ actually means
This is where I had to correct my own assumption. “Non-work” sounds like it might mean entertainment, or messing about, or some vague digital leisure. It mostly doesn’t.
The team also sorted messages by what the person wanted the model to do. They report that about 49% of messages are Asking, 40% are Doing, and 11% are Expressing. “Asking messages” seek information or advice to support a decision, while “Doing messages” request that ChatGPT complete a task and produce a deliverable.
So the largest single bucket is people asking for information and advice to help them decide something. Not generating reports. Deciding. What to cook, how to read a contract, whether a symptom is worth a doctor’s visit, how to phrase a difficult message to a relative. The everyday admin of being a person, handed to a machine that answers instantly and never sighs.
That, to me, is the surprise inside the surprise. The non-work majority isn’t people goofing off. It is people quietly outsourcing the small decisions that perhaps used to mean a phone call to a friend, a forum thread, or an hour of squinting at search results.
The companionship story the numbers don’t support
There is a popular version of all this where people are falling in love with chatbots and treating them as confidants. It makes for a good headline. The data does not back it up. The paper found that only 1.9% of ChatGPT messages are on the topic of Relationships and Personal Reflection. One classifier on one sample is not the final word, but a sliver this small is hard to square with the lonely-confidant narrative that keeps getting recycled. I have wandered into that sliver myself, if I am honest. I have asked ChatGPT for an unflattering, honest read on me, the kind of mirror you would normally only get from a friend brave enough to risk the friendship. It is a strange and slightly humbling exercise. But it is exactly that, an exercise I tried once or twice, not a habit. That matches the data better than the headlines do, and I think the headlines are wrong.
What I notice in my own use
When I look at my own days, the pattern in the paper holds. I use AI for some of the gathering and the thinking-aloud, the parts where I am asking rather than creating. It is good at fetching, summarizing, laying options on the table. The Asking bucket is where most of my use lives too.
And that is the part I keep turning over. The small decisions I used to make by calling someone, or sitting with for an afternoon, are the ones I now hand off fastest. I am not sure yet whether that is a gain or a quiet loss.
