People bring a wider mix of work to ChatGPT. The top ten tasks fell from 56% to 51%

On 26 October OpenAI released OpenAI Signals data that maps work-related ChatGPT messages to 165 work activities. It covers nearly 100 countries and seven quarters, from the start of 2025 to the third quarter of 2026. Researchers can line it up directly against employment statistics.

OpenAI sorts work messages in the same vocabulary labour economists use

The data is built on O*NET, the standard US framework that groups occupation-specific tasks into broader activities shared across jobs. OpenAI calls them Intermediate Work Activities. Examples include preparing instructional materials, developing marketing content, programming and analysing business data.

The headline finding is a single number. Across the 74 countries OpenAI observes in every quarter, the ten most common activities made up an average of 56.0% of classified work messages in the first quarter of 2025. By the third quarter of 2026 the figure was 51.1%. OpenAI reads this as countries gradually specialising, each in its own direction. The differences, it says, may reflect the kind of work done in each country or uneven access to AI tools.

The release also includes an interactive country chart and a comparison with the US economy. The two distributions differ sharply. About 43% of the employment-weighted task mix in mapped US occupations falls into an “other activities” bucket, including moving materials, operating vehicles and maintaining health records. Those activities did not appear consistently enough in the US ChatGPT sample for OpenAI to report them separately.

The AI and jobs debate gets a quarterly series instead of impressions

Arguments about which occupations AI is changing have so far leaned on surveys, job postings and estimates of which tasks are exposed. A quarterly series in a shared vocabulary lets an economist pick a specific activity and check whether it grows in the same countries where hiring shifts. Anthropic has done something similar since 2025 with the Anthropic Economic Index, which also maps Claude conversations to O*NET tasks. Two leading labs now describe work in the same language, even if their samples and incentives differ.

For companies and governments the falling concentration matters more than the ranking itself. If ChatGPT had stayed mostly with writing and programming, watching a handful of occupations would be enough. The spread into other activities suggests the effects will be harder to find and will surface differently in each country. Anyone planning retraining or internal training now has at least a rough map of what people actually bring to the chat window.

The data sees consumer chat, enterprise deployment is missing

The sample has important limits. It covers only work-related messages on individual ChatGPT plans. Messages from enterprise accounts and Codex usage are excluded, which means the settings where companies wire AI into their processes and where agents work directly with code. The result describes individuals who chose the tool themselves, not organisations that bought it.

Messages are also labelled by an automated classifier after identifying information is removed. OpenAI withholds small groups and adds statistical noise. That protects privacy, but classifier errors flow straight into the shares and nobody outside can check them, because the underlying messages stay with OpenAI. The drop from 56.0% to 51.1% could just as well reflect new users with different habits or changes to the product itself.

Enterprise accounts and outside studies will decide what the data is worth

The next release will show how seriously OpenAI takes its role as a research source. Adding enterprise accounts and Codex would bring back the work the table cannot see today. The second test will be independent studies that link the data to wages or hiring and reach conclusions OpenAI did not write into its own announcement.

Lilith's verdict

OpenAI has opened the shop window where people carry their own work into a chat. The back warehouse with enterprise accounts and Codex stays shut, and that is where it will be decided whose work AI takes and whose it merely speeds up.