OpenAI says ChatGPT users are doing other people's jobs

OpenAI has published a new study built entirely from its own ChatGPT conversation logs, arguing that the tool is quietly reshuffling who does which job tasks inside companies. The finding rests on how OpenAI itself defined an occupation and a task, a method the report describes only in outline. That distinction matters because the resulting numbers are being offered as evidence for a claim OpenAI has made before: that AI reorganizes jobs from the inside, ahead of any change to a title or a job posting.

The company reviewed more than 800,000 messages sent by U.S. ChatGPT users. Before counting anything as crossover, it set aside activities common to nearly every job, a bucket it labeled generic and filled with scheduling, basic summarizing, and general writing help. What remained, OpenAI calls task crossover: work historically tied to one occupation showing up in the ChatGPT use of someone in another. By that method, 16.8% of all work-related messages, and 43.5% of messages tied to a specific occupation, fell outside the sender’s own declared job category.

The crossover was uneven across roles. Designers asked for outside-occupation help in 75% of their occupation-specific messages, human resources workers in 69%, legal workers in 56%, and marketers in 53%, while customer experience workers led the group at 77%. Two particular activities kept resurfacing everywhere: people running numbers and people fixing technical problems. OpenAI found each of those ranked in the top three outside-occupation tasks inside every one of the seven other job groups it tracked, which makes crunching financial figures and troubleshooting technology the most portable skills in the entire dataset.

The direction of the flow depended on the role. Marketers sent 24.3% of their own messages toward work tied to a different occupation, but the comparison flips in reverse: marketing-style tasks turned up in 8.9% of messages from people outside marketing, the largest outbound share OpenAI recorded for any occupation. Design worked the opposite way. Designers imported outside work constantly, with 35.2% of their messages covering tasks tied to another job, yet almost none of that flowed back out; only 1.7% of messages from non-designers touched on design work at all.

Workspace size shaped the pattern too. Small teams running two to five ChatGPT seats posted the highest crossover rate among typical users, 18.9%, against 16.3% at companies with more than 100 seats, a gap that fits OpenAI’s theory that small-company employees absorb tasks a bigger employer would route to a specialist. That relationship broke down among the heaviest ChatGPT users, a detail OpenAI reports without much explanation beyond a guess that heavy users have already settled into fixed routines.

None of this comes from a survey. OpenAI does not say how it determined each user’s occupation: whether from account setup fields, a classifier applied to the messages themselves, or some mix of both. It also does not disclose the time window the 800,000 messages span, or whether the sample was drawn randomly or filtered toward business users in a way that could skew the results. The line between a generic task and a crossed-over one is likewise OpenAI’s own analytical judgment call, not an external standard anyone outside the company can check.

That is the pattern worth naming plainly: a company measured the effects of its own product using a taxonomy it built, on logs only it can see, without outside audit of either the classification method or the sampling. The underlying observation, that ChatGPT lets people take on tasks outside their formal role, may well be accurate. But the 43.5% figure describes what OpenAI’s own classifier found in OpenAI’s own data, not an independently verified account of how work is actually being reorganized across the economy.

For operators, the specific percentage matters less than the direction. Task boundaries are loosening fastest in customer experience, design, and HR, and running financial calculations alongside fixing technical problems are the two skills spreading into every other function OpenAI tracked. Teams staffing those roles this quarter should test whether a generalist with ChatGPT access already covers work they assumed required a dedicated specialist hire.

OpenAI published this analysis, titled “Work at the Frontier: How AI is Expanding What People Do at Work,” on its website; the version reviewed carried no dated byline.