Epoch AI 推出 ChatGPT usage explorer，基于5000名美国用户三年聊天记录

ChatGPT grew from 100 million weekly active users within a few months of its launch to more than 900 million in February 2026, according to OpenAI. With so many people now using AI, questions about its effects, both positive and negative, have taken on greater importance. Those questions are hard to answer without a good understanding of how people actually use these tools and how that use has changed over time.

Most behavioral data on AI use is either self-reported, as in our own polls with Ipsos, or is reported by the AI companies themselves, as in OpenAI’s analysis of its usage logs and Anthropic’s Economic Index. Independent analysis of user behavior from logs, covering several years and thousands of users, is much rarer. Today, we are launching the ChatGPT usage explorer to help fill that gap.

The explorer is built on data provided by YouGov. It publishes aggregate metadata from the ChatGPT histories of 5,000 US YouGov panelists who used ChatGPT in 2026 and agreed to share their chat exports. Some user histories trace back to the weeks immediately following ChatGPT’s launch in November 2022, providing a message-level record of how their use evolved over three years.

The explorer is intended as a resource for researchers, policymakers, journalists, and anyone tracking AI use. Below, we describe the data and its limitations, and highlight our initial findings.

Who’s in the dataset

YouGov randomly selected 5,000 ChatGPT users from its US panel and provided Epoch AI with the metadata from their chat histories. To be eligible for selection, people must have (1) opted in to share their ChatGPT histories and (2) used ChatGPT at least once in 2026. Selected people exported their histories between January and August 2026, resulting in about 8.3 million messages across about 660,000 conversations.

This is a retrospective, unbalanced panel dataset: we follow the same people over time (the ‘panelists’), but they contribute different numbers of active months. Among the 4,324 panelists active at least once between December 2022 and December 2025 (the remaining 676 panelists were first active in 2026), we observe activity in an average of 10.7 months, with a median of 9 months. Most (91.2%) are active in at least two months, and 38.2% are active in at least 12 months. 9.6% are active in at least 24 months. Gaps between active months may reflect periods without use or incomplete chat history exports.

What the dataset includes

YouGov removed all message text before sharing user data. We never observed actual transcripts, so the explorer doesn’t classify use cases or topics.

In addition to demographic information on users, the metadata we received includes the following message-level details: the time each message was sent, its character count, which model replied, and which tools that model used.

From this metadata, our ChatGPT usage explorer publishes nine metrics:

• Panelist metrics: active panelists, active panelists by plan, first recorded ChatGPT use, panelists by days active, and messages per active panelist
• Conversation metrics: prompts per conversation, most-used model, tool use, and reasoning model use
• Subgroups (in the downloadable dataset): monthly active panelists and messages per active panelist by gender, age, education, household income, employment status, and plan

A panelist counts as “active” if they sent at least one prompt during the selected timestep (month or week).

Because panelists exported their histories at different times in 2026, data for 2026 months is incomplete. By default, the data we display in the explorer ends in December 2025 to account for this. 2026 data is accessible using the “Show incomplete data” setting in the explorer or within the downloadable dataset. A “Show model/tool releases” setting marks model and tool launch dates on the charts. Every series, monthly and weekly, can be downloaded.

How people in the dataset compare with other ChatGPT users

Participation was opt-in, and the data is unweighted, meaning it has not been adjusted to match the demographics of US adults or of ChatGPT users as a whole.

Additionally, panelists must have used ChatGPT at least once in 2026 to be included, so people who tried ChatGPT in 2022–2025 but stopped before January 2026 aren’t in the data. This means the data is subject to survivorship bias that will tend to push usage rates up over time. The explorer is therefore most informative about changes over time and differences within this specific panel, a self-selected set of more recently active ChatGPT users. It is less informative about broader usage patterns, such as how many Americans use ChatGPT.

Users in our dataset are also older, on average, than ChatGPT’s overall user base. Compared with respondents in our March 2026 survey who reported recently using ChatGPT, this dataset has fewer adults under 30 (14% vs. 29%) and more adults aged 65 and over (18% vs. 10%). Compared with US adults overall, it has more college graduates (55% vs. 40%) and fewer Hispanic adults (8% vs. 19%). An older panel may understate technical and frequent use compared with ChatGPT users overall.

Rates of use in this dataset do correspond closely to OpenAI’s own figures, however. In July 2025, weekly active panelists sent 4.3 prompts per day, compared with about 3.7 implied by OpenAI’s published figures. From June 2024 to June 2025, the panel’s prompt volume grew 5.6×, compared with 5.8× in OpenAI’s data.

The explorer’s methodology page further compares the ChatGPT usage dataset with the Census, Pew, our Ipsos survey, and OpenAI’s paper.

What we’ve learned from the data so far

Usage has intensified among active panelists

The clearest pattern we observe in the data is intensification: month by month, active users increased their usage. On average, they send more messages, are active on more days, and have longer conversations. As discussed above, due to survivorship bias, these trends should be understood as true of the histories of this set of active 2026 users, not of all users during the 2022–2025 period.

The median number of messages per active panelist, counting prompts and replies, rose from 14 in January 2023 to 36 in December 2025. The mean is much higher, at 158 in December 2025, because a small group of heavy users sends a large share of all prompts. The most active 10% of panelists sent 63% of prompts.

Frequent use has also become more common among the panelists. Between December 2023 and December 2025, the low end shrank: the shares of those who used ChatGPT on only 1 day or 2–5 days in the month fell (32.6% to 22.1% and 40.3% to 33.2%, respectively). The top share increased: the 11–20-day share rose from 8.1% to 17.0%, and the share that used it on 21 or more days quadrupled, from 2.6% to 10.5%. The middle ground remained stable, with the 6–10-day share staying close to the original 16.4% at 17.3%. These are shares of each month’s active panelists, so the shift can come from changes in who was using ChatGPT as well as from changes in how often people used it.

The average conversation has grown longer, but the typical conversation hasn’t. For every month except September 2023, the median conversation was comprised of just two prompts, while the mean rose from 3.7 prompts in December 2022 to 5.8 in December 2025. Active panelists also tend to have more conversations each month. The median rose from three conversations a month in December 2023 to five in December 2025.

The charts don’t separate changes in behavior from changes in who is in the data. New users enter the data when they start using ChatGPT, so each month’s active panelists can change. And because every panelist must have used ChatGPT in 2026, the trends describe people who kept using it, not everyone who ever tried it.

The explorer complements our polling work

This new data explorer and our polling on AI usage answer different, but complementary, questions. Our polls are representative of US adults and track trends in population-level AI use. Our latest poll found that near-daily AI use among US adults more than doubled between March and August 2026. The explorer shows what one group of ChatGPT users actually did, message by message, over three years. But both point to AI use becoming more frequent.

Next, we plan to look more closely at the heaviest users and at how individual panelists’ use changes over time.

To learn more, go to the ChatGPT usage explorer for interactive charts and downloads. For more details, there is a documentation page with methodology, definitions, and comparisons with other sources.