ChatGPT users do other people's jobs, OpenAI finds

By Claire Bennett · 2026-07-30

The number that should unsettle every job description writer

Among occupation-specific ChatGPT messages, 43.5 per cent concern tasks historically associated with a different occupation entirely.verifiedVerified Source: cdn.openai.com^[1] That is not a rounding error or a quirk of methodology. It is nearly half of all the moments when workers with a defined job title turn to an AI tool and ask it for something specific.

OpenAI's Economic Research team published the finding in its Work at the Frontier report on 27 July 2026, drawing on a dataset of more than 800,000 work-related messages from US ChatGPT Business users.^[2] The figure for all work-related messages sits lower, at 16.8 per cent, but the occupation-specific subset is what matters here. It captures the moments where role boundaries are most legible and most clearly being crossed.^[1]

Report co-authors Caroline Chin and Alex Martin Richmond gave the phenomenon a name: "We call this pattern task crossover: work historically associated with one occupation appearing in the AI use of people in another."verifiedVerified Source: cdn.openai.com^[1] Chin and Richmond went further, writing that "AI may change not only how work gets done, but who does it."verifiedVerified Source: cdn.openai.com That second sentence deserves to be circled on every workforce planning document in the country.

Who is crossing over and who is not

The crossover rates by occupation are where the report gets genuinely instructive, because they are not evenly distributed. Customer experience workers recorded the highest crossover rate at 77 per cent of their occupation-specific messages, while engineers recorded the lowest at 28 per cent.^[1] That gap tells a story about the nature of the work itself.

Engineers work inside hard technical constraints. A language model can help write documentation or draft a project brief, but the physics of a system, the architecture of a codebase, the structural tolerance of a component, these do not yield to prompt engineering. A customer experience worker, by contrast, is already expected to be a generalist: managing complaints, synthesising policy, drafting communications, sometimes escalating to legal or HR. AI does not break those boundaries so much as it removes the friction that used to keep people inside them.

The middle of the distribution is revealing. Designers came in at 75 per cent, human resources workers at 69 per cent, legal workers at 56 per cent, and marketers at 53 per cent.^[1] Legal at 56 per cent deserves particular attention. A worker asking ChatGPT to help draft a contract clause, interpret a regulatory update, or prepare a compliance checklist is doing something that was, until recently, the province of a separately credentialled professional. The data says nothing about quality, but at scale, workers are clearly doing it.

The report does not break down crossover by seniority, industry vertical, or company size. A senior marketer experimenting with a legal summary for internal use is a different risk profile from a junior customer service agent drafting policy language that flows into customer-facing documentation. What the data does establish is that the behaviour is widespread and not confined to a particular type of role.

What the methodology actually measured

The dataset covers US ChatGPT Business users, meaning paid enterprise accounts rather than free-tier individual users.^[2] Enterprise users are more likely to be working in structured organisational contexts, with job titles that map reasonably well to occupational classifications. Messages were classified by occupation and by task type, allowing researchers to identify when a message from someone in a given role contained a task historically assigned to a different occupation.

OpenAI's own research team produced the report, so the figures come from the company's own platform data and its own analytical framework. Bushletter could not independently verify the classification methodology from the published report alone, and the dataset is not publicly available. The framing around task crossover reflects OpenAI's institutional perspective, treating the pattern as a positive expansion of worker capability rather than a dilution of professional standards. Neither observation undermines the core finding, but both are worth keeping in mind when the numbers are cited in workforce policy arguments.

The job design problem no one has solved yet

The practical implication is not subtle. If a customer experience worker is doing legal and HR tasks nearly four times out of five when using ChatGPT for occupation-specific purposes, then the job description they were hired against, the training they received, and the performance metrics they are measured against all describe something different from the job they are actually doing.

This is not a future problem. The data comes from current ChatGPT Business users, in current roles, making current requests. The crossover is already happening at the scale of hundreds of thousands of messages, and the question for Australian employers is whether their organisational structures have registered the shift at all.

Australian workplaces carry particular compliance weight in financial services, healthcare, legal and government. Roles in those sectors are often structured around regulatory obligations that assume a credentialled professional is doing the credentialled work. When a worker uses AI to cross into that territory, the accountability question does not disappear. It just becomes harder to trace. Who is responsible when an uncredentialled worker, assisted by a language model, produces legal or compliance output that turns out to be wrong?

The Work at the Frontier report argues that task crossover reflects AI expanding what individual workers can do, not merely substituting for other workers, a distinction that has direct consequences for how organisations should rethink role design and hiring strategy.^[1] That framing is plausible. It is also possible to read the same data as evidence that professional role boundaries are eroding faster than the governance structures around them can adapt.

The honest answer is that the data supports both readings at once. Workers are doing more, and they are doing it across traditional occupational lines. Whether that is expansion or erosion depends almost entirely on what quality controls an organisation has in place, and whether those controls were designed for a world where a customer experience worker might be doing legal analysis before lunch and HR policy drafting before close of business.

Implications for hiring and skills

For recruitment, the task crossover finding complicates the unit of hiring. If the job a person will actually do diverges significantly from the occupation they were hired into, then hiring for occupational credentials becomes a weaker signal of fit. What replaces it is less clear: judgement, communication, critical evaluation of AI output. None of those have standardised credentialling pathways in Australia or anywhere else.

For skills training, the implication is that the most valuable capability may not be depth in a single occupational domain but the ability to operate credibly across several. The Work at the Frontier report does not prescribe a solution. What it does is put a number, 43.5 per cent, on the scale at which the old job map has already stopped describing the territory.

Claire Bennett

Claire Bennett writes about work and workplace culture. She is interested in the gap between how organisations describe themselves and what it feels like to work inside them.