On September 16, 2026, the economic research team at OpenAI published a study covering more than 1.5 million work-related messages sent to ChatGPT by registered users in the United States. The findings suggest that some employees regularly return to tasks traditionally associated with other professions. This shift may broaden the content of jobs even before their titles change, but the study measures neither the productivity achieved, nor the quality of the results, nor the effects on employment and wages.
A study of how ChatGPT is actually used at work
The Work at the Frontier report analyzes more than 1.5 million messages classified as work-related, sent between April and July 2026 from individual ChatGPT accounts. The participants also had a ChatGPT Business account, which allowed OpenAI to link their messages to an occupation or department declared at sign-up.[1]
This methodological detail matters. The messages do not come from companies’ ChatGPT Business workspaces, and the study does not observe all of the work performed. OpenAI says it excluded users who had opted out of having their data used for training, automatically removed personal information before classification, and prevented researchers from reading individual conversations.[1]
The lab then compares the requests with the activities listed in O*NET, the US database of occupations and tasks. A request is considered “cross-functional” when it corresponds more closely to another occupational family than to the occupation declared by the user. The study thus seeks to determine whether AI merely enables one-off experimentation or whether certain new activities become part of everyday routines.
Tasks outside one’s occupation that become recurring
Among roughly 6,200 people tracked continuously over the four months, the share of messages corresponding to cross-functional tasks they had already used rose from 13.1% in April to 25.9% in July. The indicator nearly doubled, but the way it is built mechanically favors an increase: each new month expands the number of tasks that can qualify as recurring. OpenAI points this out in the report.[1]
A second comparison nonetheless strengthens the hypothesis that these practices are taking hold. When a user had turned to ChatGPT for a cross-functional task in a given month, they returned to that same category the following month in 23.6% of cases, compared with 8.4% among comparable users who had not used it before. That is a gap of 15.2 percentage points.[1]
Not all task expansions are equal
Return rates vary widely depending on the nature of the work. Customer interactions reach 54%, advertising or promotional copywriting 44%, and the creation of marketing materials 37%. By contrast, legal research is repeated in only 10% of cases, while explaining financial information and presenting business information reach 15%.[1]
This difference may reflect how often tasks naturally occur, but also the level of perceived risk. Producing a first draft of a sales message is relatively easy to check. Interpreting a legal rule or explaining financial information requires expertise, validation and more clearly defined accountability. A low repetition rate therefore does not mean that AI is incapable of helping, but that its integration depends on the context and on the consequences of an error.
Does AI really let people borrow expertise?
OpenAI observes that cross-functional requests are slightly shorter, by about twenty characters on average. They less often ask for explanations, advice or a detailed format, but they provide more examples and context and more frequently request verification. The authors put forward the idea of “borrowed expertise”: the user brings a problem and contextual elements, then relies on the model to apply knowledge from another field.[1]
This interpretation calls for caution. A well-worded output does not prove that a skill has been acquired. Users can obtain a deliverable without understanding the method, the assumptions or the limitations. In the short term, this reduces friction between departments. In the longer term, organizations must distinguish assisted autonomy, which relies on the ability to check the work, from uncontrolled delegation that could weaken learning.
The International Labour Organization estimates that one in four workers worldwide holds a job with some degree of exposure to generative AI. It nevertheless considers the transformation of tasks more likely than the complete disappearance of most jobs, notably because human intervention remains necessary.[2] OpenAI’s study offers a behavioral perspective consistent with this view, without directly measuring employment.
Broader jobs, but not necessarily less specialized
Broadening tasks can change the division of labor. A project manager may prepare marketing materials, an analyst may produce a first legal summary, or a developer may structure a customer communication. The job title stays the same, while the mix of activities gradually evolves.
This versatility does not mean specialists will disappear. When the stakes become complex, regulated or irreversible, their judgment remains central. AI may shift the point at which they step in: specialists no longer carry out every step, but define the rules, check the assumptions, handle exceptions and take responsibility for final validation.
For companies, then, the issue is not just access to ChatGPT. It is about work design: which tasks can be broadened, which require authorization, where human review should sit, and how new responsibilities should be recognized? A silent expansion of scope can create value, but also lead to overload or an informal transfer of responsibility.
Skills to develop beyond prompt mastery
- Frame a problem and provide sufficient context without exposing confidential information.
- Assess the quality of a response, check sources and spot weak assumptions.
- Understand the limits of one’s own area of expertise and know when to call on a specialist.
- Document the use of AI, human corrections and the final decision to ensure traceability.
- Collaborate across functions so that broader tasks do not turn into software-assisted individual silos.
In the European Union, Article 4 of the Artificial Intelligence Act has, since February 2, 2025, required providers and deployers to take measures to ensure a sufficient level of AI literacy among the people who use AI on their behalf. The expected level depends on the knowledge, experience and context involved, as well as on the people concerned.[3] Training therefore cannot be limited to a few prompt-writing techniques.
Privacy, governance and social dialogue
The study itself raises a data governance question. OpenAI states that researchers did not read the conversations, that personal information was removed automatically, and that the analysis was based on labels and pseudonymized identifiers. These safeguards reduce exposure, but they do not eliminate questions about transparency, legal basis, retention periods and individuals’ rights when work-related usage is analyzed.[1][4]
Within organizations, risks also arise when employees paste customer data, contracts, HR information or trade secrets into an unauthorized service. A usage policy should specify the approved tools, the prohibited categories of data, the oversight procedures and the responsibilities. The CNIL recommends, in particular, defining the purpose of the processing, the roles of the parties involved, the legal basis and, where necessary, an impact assessment.[4]
Finally, the redefinition of tasks must be discussed with employees and their representatives. It can increase autonomy and reduce certain dependencies, but it can also raise expectations without any change in time, recognition or pay. The ILO stresses social dialogue so that productivity gains go hand in hand with better working conditions.[2]
What the study cannot yet establish
- The sample is not representative of the entire US workforce, let alone the global labor market.
- The study observes sampled messages and may miss repeated uses found in conversations that were not selected.
- Automatically classifying a task against O*NET does not capture the full reality of a job, which depends on the sector and the organization.
- No direct measure links the observed usage to quality, productivity, errors, wages or employment.
- OpenAI is studying the use of its own product, with no announced independent validation of the full set of results.
Toward jobs that are reshaped before they are renamed
The main takeaway is not that ChatGPT replaces one profession with another. It is that the boundaries between functions can become more porous. Some simple, verifiable tasks are absorbed by employees who were not traditionally responsible for them, while high-stakes activities remain more closely tied to specialized expertise.
To assess the real scope of this shift, it will be necessary to track the quality of outputs, incidents, workload, career mobility and the recognition of skills. If AI broadens jobs, organizations will also need to broaden training, governance and accountability mechanisms. The question is therefore not only what employees can now accomplish with a model, but under what conditions these new tasks become lasting professional progress.
Further reading
To explore the transformation of jobs, professional uses of AI and the skills needed to adopt them in more depth, check out these analyses from the aivancity blog.
Sources
[1] OpenAI Economic Research. (2026). Work at the Frontier: How workers are unlocking new ways of working. View the report
[2] International Labour Organization (ILO). (2025). Generative AI and Jobs: A 2025 Update. View the report
[3] European Commission. (2026). AI literacy, Questions and Answers. View the source
[4] CNIL. (2024). Recommandations sur l’application du RGPD au développement des systèmes d’intelligence artificielle. View the recommendations
[5] OpenAI. (September 16, 2026). How workers are unlocking new ways of working. View the source