OpenAI Signals: ChatGPT work data vs Japanese surveys

On October 26, 2026, OpenAI’s economic research team published “ChatGPT’s expanding world of work”. It is a dataset of how ChatGPT is used for work: work messages from about 100 countries, split into 165 types of work activity and aggregated by quarter. You can browse it as charts or download it as CSV, and it is meant for researchers to use in their own analysis.

I wanted to know whether it could be compared with how people use AI at work in Japan, so I read the data and the documentation on how it was built. The short answer: Signals is data that counts messages, while the Japanese surveys are data from asking people and companies. Both describe a “breakdown of work,” but in different units.

What was published#

This release is part of OpenAI’s ongoing usage-data publication under the name “Signals.” It is version 3.0, and the documentation (README) lists 24 CSV files. The file for the breakdown of work is share_of_work_related_messages_by_onet_iwa_country_quarter.csv. For each country and quarter, it gives the share of work messages that fall into each work activity. The period runs from Q1 2025 through Q3 2026.

The data is licensed under Creative Commons Attribution 4.0 (CC BY 4.0), and the documentation shows how to cite it.

“Work activities” follow the O*NET taxonomy#

The work activities come from O*NET, the occupational information database maintained by the US Department of Labor, specifically its Intermediate Work Activities (IWAs). IWAs group the detailed tasks of each occupation into activities shared across occupations. The O*NET list has 332 of them, at a level of detail like “Program computer systems or production equipment” and “Resolve computer problems.”

A language model decides which IWA each message belongs to. The classification prompt in the README asks it to pick the single best match out of the 332, or return “unknown” when it cannot tell. As an example, the message “What’s the difference between Python and JavaScript? Which is better for a beginner?” is classified as “Evaluate characteristics, usefulness, or performance of products or technologies.”

165 of them are published. IWAs with few messages are grouped into “Other IWA” for privacy and to keep the numbers stable. Crossing country, quarter and work activity produces many small cells, so this file has an extra condition: a cell appears only if it has at least 1,000 effective messages per quarter.

The two findings OpenAI reported#

The official post gives numbers for two findings.

The first is that usage is gradually spreading out. It covers the 74 countries with data in every quarter. Averaging, across countries, the share of each country’s classified work messages taken by its top 10 work activities, the figure fell from 56.0% in Q1 2025 to 51.1% in Q3 2026. OpenAI describes this as each country’s work usage becoming more specialized, and specialized in different directions from country to country. It names possible reasons, such as differences in what work each country does and in access to AI tools, but does not settle on one.

The second is a comparison with the work actually done in the US. When the mix of work activities is weighted by employment across US occupations, about 43% falls into “Other IWA.” These are activities like moving loads, driving vehicles and managing medical records, which did not reach enough messages in ChatGPT. OpenAI says this 43% is the main reason ChatGPT usage looks so different from the mix of work in the US economy.

Limits to keep in mind#

Here are the main limits, from the Signals data page and the README.

Item	Details
Plans covered	Free, Go, Plus, Pro (accounts mainly managed by individuals)
Not included	Messages from Enterprise and other business plans, Codex
Users covered	Accounts self-reported as 18 or older. Excludes deleted accounts and users who opted out of training
Sample	1 million messages each month, July 2024 to September 2026
Classification	Automatic estimates by a language model, not self-reports by users
Privacy	Personally identifiable information is removed before classification, and researchers do not read the messages. Differential privacy noise is added, and small cells are suppressed

The biggest effect comes from leaving out business plans and Codex. OpenAI itself writes that work and technical uses are likely undercounted. If a company buys ChatGPT Enterprise for its employees, that usage is not in this data. Business usage is published separately as “Enterprise Signals.”

That the classification is an estimate also matters for reading the numbers. Both “is this a work message” and “which work activity” are judgments a model made from the conversation. The README asks users to treat such columns as “inferred behavioral classifications, not self-reports.” It also notes that finer cuts, such as topic shares by state, rest on fewer messages and vary more. The file that crosses country and work activity is a similarly fine cut.

How it differs from Japanese surveys#

Two sources are often cited for how AI is used at work in Japan: the Ministry of Internal Affairs and Communications’ 2026 White Paper on Information and Communications (Japanese), and the survey of corporate trends in generative AI (Japanese) by Teikoku Databank, a Japanese credit research firm. Put the three side by side and they count different things.

Survey	Method and coverage	How the breakdown of work is counted
OpenAI Signals	Real messages from consumer ChatGPT, classified automatically	Share of work messages. Split by O*NET IWA (165 types)
White Paper (individuals)	Survey of individuals, including people who do not use AI	Share of people who answered, for example, “use regularly” for each purpose. Split into 11 purposes such as understanding information and writing or proofreading documents
Teikoku Databank	Survey of companies. 10,312 valid responses	Share of companies choosing a “main task.” Split into items such as writing, summarizing and proofreading text, and gathering information

In the White Paper’s survey of individuals, 58.8% of people in Japan had used some generative AI service. The figure for the US and Germany was 75.6%. In the question on purposes of AI use, the share answering “use regularly” was highest for understanding information at 33.7%, followed by writing or proofreading documents at 20.1%, translation at 13.7%, and programming or data processing at 4.9%. These purposes cover both work and personal life. According to Nikkei (Japanese), among people who had used it, 29.9% in Japan said they used it “almost every day,” against 47.9% in the US.

In Teikoku Databank’s March 2026 survey, 34.5% of companies were using generative AI in their work. Among those, the main task was writing, summarizing and proofreading text for 45.1%, gathering information for 21.8%, generating ideas during planning for 11.0%, and programming support such as code generation for 5.9%. In the White Paper’s survey of companies, about 70% said they used it to help write meeting minutes and emails.

The Japanese surveys point to usage weighted toward text and gathering information. To check the same thing in Signals, you would filter the CSV by the country code JP and look at Japan’s top work activities and its “top 10 share,” the measure that fell on average across the 74 countries. If Japan’s top 10 share is above the average, you can read that as Japan concentrating on a few uses.

But in a survey’s “share of people,” a light user and a daily user each count as one person. In Signals’ “share of messages,” the uses of heavy users weigh more. Work done outside ChatGPT’s consumer plans, such as a company-provided business plan or Copilot, is not in Signals either. It seems best to read it as covering work messages sent from consumer ChatGPT in Japan. For usage inside your own company, it seems better to measure it in your own admin console, like the admin center report covered in my post on Fujitsu’s Copilot adoption rate.

Who the data is useful for#

It is easiest to use for researchers studying AI and labor. Because it is aligned with the O*NET taxonomy, it can be joined with employment and wage statistics by occupation. You can also track differences between countries and changes from quarter to quarter.

For people driving AI adoption at a company, it is material for comparing your own usage with individual use worldwide. If you map your internal usage logs to IWAs, you can line them up in the same taxonomy. Keep in mind, though, that what you are comparing against is people using personal accounts for work, not how other companies use AI as organizations.