OpenAI looked at 800,000 work chats and found people quietly doing other people's jobs

OpenAI published the first report in a new research series on Monday, and the headline finding is easy to picture even if you have never opened a research paper. After analysing more than 800,000 work-related ChatGPT messages from US users, the company found that 43.5 percent of occupation-specific messages were about tasks that normally belong to a different job. A marketer troubleshooting a website. A salesperson poking at a customer dataset that would once have gone to an analyst. A small-business owner reviewing a contract. OpenAI calls the pattern task crossover.

The method matters here, because the raw number sounds bigger than it is. OpenAI first stripped out what it calls generic work: writing, summarising, scheduling, the things everybody does regardless of job title. That accounted for 61.5 percent of all work messages. Only the remaining 38.5 percent counted as occupation-specific, and it is within that slice that 43.5 percent crossed a job boundary. Some professions crossed far more than others: 77 percent for customer experience workers, 75 percent for designers, 69 percent for human resources, 56 percent for legal, 53 percent for marketing.

Two tasks travelled furthest into other people’s territory. Financial calculation and technology troubleshooting showed up among the three most common outside tasks in every single one of the other seven occupation groups studied. Designers turned out to be heavy borrowers and light lenders: 35.2 percent of their messages involved somebody else’s work, while design tasks made up only 1.7 percent of everyone else’s. Engineering was closer to the opposite, exporting far more than it imported. Company size mattered too. In workspaces with two to five seats, 18.9 percent of an average user’s messages were outside-occupation tasks; above 100 seats it fell to 16.3 percent. Where there is no specialist down the hall, the person holding the problem keeps it.

Why does this happen at all? A large language model is a generalist by construction: trained on text from every profession at once, it has a shallow but broad competence in most of them, and no knowledge of your company at all. That combination produces exactly the pattern OpenAI measured. Tasks cross over easily when general knowledge is most of the job, such as writing a formula or reading an error message, and stay put when they depend on context nobody wrote down. Worth naming what the study does not show: this is usage data, not outcome data. OpenAI counted what people asked the model to do, not whether the contract review was any good. The company also has a commercial interest in this story, though the method of mapping messages onto the US government’s O*NET occupational database is a reasonable one.

What this means for you: If you are new to AI, the value here is not that AI writes your emails, it is that it lets you attempt the one task a week that used to require booking someone else’s time. Start with the two that travel best, a quick financial calculation or a bit of technical troubleshooting. If you already use AI heavily, treat the numbers as a caution as well as a promise. Doing legal or financial work outside your training is exactly where a confident wrong answer costs the most, and OpenAI’s data measures what people attempted, not whether they got it right. It also covers US users only.

Sources

• How AI is expanding what people do at work, OpenAI
• Work at the Frontier report (PDF), OpenAI
• OpenAI says more workers are using ChatGPT to do other people’s jobs, The Decoder