ChatGPT’s expanding world of work

OpenAI's economic research team has released a new slice of its OpenAI Signals data that maps work-related ChatGPT messages to 165 standard task categories, called Intermediate Work Activities, across nearly 100 countries. The data runs quarterly from the first quarter of 2025 through the third quarter of 2026, and it is built on O*NET, the occupational framework labor economists already use.

That last part is what matters. Most of what we know about AI and jobs comes from surveys, vendor anecdotes or guesswork. A dataset that lines up real usage with the same categories used in employment statistics gives researchers something they can actually test against wages, hiring and productivity numbers.

We think this is a genuinely useful release, and we want to give OpenAI credit for putting usage data in a form outsiders can work with. The debate over AI and work has been running on vibes for too long. Comparable categories, a time series and country breakdowns are the raw material for better answers.

The headline finding is subtle but worth sitting with. Across 74 countries observed in every quarter, the ten most common activities made up an average of 51.1 percent of classified work messages in Q3 2026, down from 56.0 percent in Q1 2025. Put plainly, people are spreading ChatGPT across more kinds of work, and different countries are leaning into different tasks. OpenAI itself says that could reflect differences in the work people do or differences in access to AI tools. Those are very different stories, and the second one is a warning about who gets left behind.

Now the caveats, and they are real. This covers individual ChatGPT plans only. It excludes enterprise accounts and Codex, which is where a lot of serious workplace use likely lives. It is OpenAI measuring OpenAI, classified by OpenAI's automated systems. And the comparison with the U.S. economy shows how narrow the view is: roughly 43 percent of the employment-weighted task mix for mapped U.S. occupations sits in a catch-all bucket, things like moving materials, operating vehicles and keeping medical records, that did not show up consistently in ChatGPT use. A large share of the economy simply isn't in the chat window.

On privacy, OpenAI says researchers do not read individual conversations, identifying information is stripped before classification, small groups are withheld and statistical noise is added. That is the right design. Still, the bigger trend is a handful of companies becoming the main source of truth about how work is changing. If AI usage data becomes the labor statistic of record, we want independent access and outside replication, not just company dashboards.

I'm watching what researchers outside OpenAI do with this. The value of a release like this is measured by the papers that challenge it, not the charts that accompany it.

OpenAI Signals is the company's ongoing effort to publish evidence on how people use AI; this release is cited as version 3.0. O*NET is a standardized framework that groups occupation-specific tasks into broader activities shared across jobs, which is why it allows comparisons with employment data.