OpenAI published a new study by its Economic Research team on September 16, 2026: employees are increasingly using ChatGPT for tasks outside their professional roles, according to the company’s own, independently unverified data. The share rose to 25.9 percent between April and July 2026. A growing part of this is becoming a regular habit in everyday work.
Study analyzes one and a half million work messages
For the study, OpenAI analyzed more than 1.5 million work-related ChatGPT messages from April to July 2026 and tracked around 6,200 employees continuously over the four months. Users self-reported their job titles in their accounts; there was no cross-check against real job profiles or personnel records.
This is the second installment of the “Work at the Frontier” series: the first study, from July 2026, analyzed around 800,000 messages across eight occupational fields and found that about 43.5 percent of occupation-specific requests involved tasks outside a user’s actual responsibilities. The new analysis builds directly on that: instead of a single snapshot, it follows the same employees over several months.
That way, OpenAI aims to move beyond raw usage frequency toward a more solid picture of how work content actually shifts over time. The research feeds into a broader debate about whether generative AI is gradually reshaping existing job profiles without any new job titles or formal role descriptions being created.
Non-expert tasks become a habit for many
The central finding: if someone used a non-expert task in one month, they returned to that exact task the following month in 23.6 percent of cases – compared with just 8.4 percent among comparable employees with no prior use of it. Put differently, roughly three-quarters of non-expert tasks tried once do not reappear a month later and remain a one-off experiment rather than a routine.
As examples of tasks that stuck around especially often, OpenAI cites writing ad copy and customer communication, troubleshooting computer programs, calculating financial figures, and explaining regulations and policies. Specialized activities such as legal research, by contrast, became a lasting habit less often.
OpenAI describes the underlying mechanism as a kind of trial phase: a person tries a non-expert activity once, finds it useful, and then deliberately returns to it in the following months until the task effectively becomes part of their own role. Job profiles can shift gradually this way, well before job titles or contracts are formally updated.
Customer service and design saw the biggest shift
The July predecessor study had already shown, according to an analysis by Built In, which occupations reach beyond their actual role most often: for customer-service staff, this applied to 77 percent of occupation-specific messages, for design roles 75 percent, in HR 69 percent, in legal departments 56 percent, and in marketing 53 percent.
Smaller companies with two to five employees had a non-expert task share of 18.9 percent of all messages, while large companies with more than 100 employees had only 16.3 percent – a sign that small teams are more likely to pitch in outside their lane.
OpenAI itself notes that enterprise customers are not included in the sample, that the eight occupational fields captured do not represent the entire US workforce, and that message counts alone cannot show actual productivity or employment effects. How strongly similar patterns show up outside the US, or in other AI tools such as ChatGPT Work, remains an open question.
What matters now is whether companies reflect this creeping expansion of responsibilities in job descriptions, training, and pay – or whether employees end up taking on non-expert extra work unpaid and unrecognized. OpenAI says further installments of the study series are planned, without naming a date for the next one.
