Borrowing expertise rather than learning it
OpenAI published the second report in its Work at the Frontier series this week, analysing more than 1.5 million work-related ChatGPT messages sent between April and July 2026. The first report documented task crossover: workers using AI for activities historically associated with a different occupation. This one asks what happens after the first time.
The first finding is about how people prompt. When the task falls outside their own occupation, workers write shorter prompts on average. They are less likely to ask for explanations, how-to guidance, a specific response format or advice — and more likely to supply examples or background, and to ask the model to check or verify something.
OpenAI’s reading is that people are borrowing expertise rather than acquiring it. Instead of asking the model to teach them a new field, they bring a document or an example from a colleague and ask it to apply someone else’s knowledge to their problem. These are all figures OpenAI reports about its own product, from its own logs.
Some of it sticks
The more consequential number concerns recurrence. Among roughly 6,200 workers observed consistently from April through July, previously used cross-occupation tasks grew from 13.1% of occupation-specific AI activity in April to 25.9% in July.
A second analysis puts that in context. Among sampled matched one-month follow-up observations, workers returned to a cross-occupation task they had used the previous month 23.6% of the time, against 8.4% among comparable workers with no observed use of it in the previous month.
OpenAI notes a mechanical caveat in its own footnotes: as workers try new tasks in any given month, the pool of tasks that can be counted as recurring grows. The trend is therefore partly an artefact of accumulation, and the paper does not attempt to strip that out.
Not every borrowed task survives
Recurrence varies sharply by what the task is. Workers came back the following month to discussing goods or services with customers 54% of the time, to advertising or promotional writing 44% of the time, and to creating marketing materials 37% of the time. Explaining financial information came back at about 15%. The average next-month return rate across all cross-occupation tasks was 18.5%.
The pattern is not hard to read. Customer-facing writing is repetitive, low-stakes and easy to check. Explaining financial information is none of those. Where the cost of a wrong answer is high, people try once and go back to a colleague who knows.
Method, and its limits
Occupations were identified from the role or department information users gave during ChatGPT Business onboarding. Training-disabled data and messages without usable classifications were excluded. OpenAI says researchers never read individual messages and that all analysis was done on aggregated, anonymised data.
The population is therefore ChatGPT Business users who filled in their department, which is not the workforce. And the conclusion OpenAI draws — that jobs may broaden while their titles stay the same — is an interpretation of four months of usage logs, not a measurement of anyone’s job description. It is a plausible reading. It is not yet evidence of a change in how work is organised.
