The impact of generative AI on employment may not first manifest as the sudden disappearance of a certain occupation. On September 16, OpenAI announced an economic study that analyzed over 1.5 million work-related ChatGPT messages from April to July 2026, observing how employees utilized AI to handle tasks that were originally part of other occupations. The study found that some cross-occupational activities would evolve from occasional attempts to become repetitive tasks, and while the content of the jobs might expand, the job titles remained unchanged for the time being.
The study inferred occupations based on the job or department information provided by ChatGPT Business users in their onboarding settings, and then summarized and anonymously categorized the messages. OpenAI stated that researchers did not read individual user messages, so data from disabled training and messages that could not be effectively categorized were excluded. This provided a large-scale behavioral sample, but it also came with limitations: the samples were from people who used ChatGPT Business, and thus could not represent all workers; there might also be errors in the self-reported job titles and message categorization.
The so-called "cross-disciplinary tasks" refer to employees using AI to complete activities that are traditionally associated with another profession. For example, an operations staff member might write marketing materials, or a technical personnel might prepare customer-facing instructions. The focus of the research is not on calculating how much work AI has replaced, but rather on whether employees repeatedly engage in these new tasks. This perspective shifts the discussion from the total number of positions to the combination of tasks.
From 13.1% to 25.9%, the increase reflects reutilization, but it does not represent the proportion of all employees.
The study continuously observed approximately 6,200 employees. In their AI activities related to specific occupations, the proportion of cross-occupational tasks previously used increased from 13.1% in April to 25.9% in July. This indicates that some of these new activities are gradually becoming part of the daily work routine. However, this figure represents the "observed proportion of AI activities," and it does not mean that 25.9% of the positions involve cross-disciplinary tasks, nor does it imply that employees spend a quarter of their time engaging in AI activities.
Another follow-up study a month later showed that the probability of employees using a certain cross-functional task again the following month was 23.6% after they had used it the previous month; in contrast, when employees had not used that task before, the usage rate the next month was only 8.4%. The difference between the two groups is consistent with path dependence: after a successful first attempt, people are more likely to continue using it. However, observational data cannot prove that AI alone causes this change; team arrangements, project cycles, and management requirements also affect reusability.
There is a significant difference in the stickiness of different tasks. For cross-professional tasks involving discussions with clients about products or services, the follow-up rate the next month is 54%; for advertising or promotional copywriting, it is 44%; and for creating marketing materials, it is 37%. Explaining financial information has a follow-up rate of only about 15%. The average follow-up rate for all cross-professional tasks is 18.5%. Tasks that are high-risk, require professional responsibility, or are under strict regulation may have an even lower reuse rate, but studies have not proven that this is the sole reason.
The way of providing guidance has also changed. When employees are handling tasks that go beyond their regular responsibilities, the guidance given is on average shorter, with fewer requests for explanations, operational instructions, specified formats, or suggestions; at the same time, examples and background information are provided more frequently, and there are also more instances where employees are asked to AI check or verify information. OpenAI summarizes one possible interpretation as "borrowing professional knowledge": employees come with specific questions and materials, rather than the system having to teach them a new profession.
This approach may both improve efficiency and conceal gaps in capabilities. Just because a user can provide background information does not mean they have the necessary professional knowledge to judge the answers correctly. Marketers using AI to interpret financial data, or engineers drafting legal terms, may produce results that seem smooth but are actually unreliable. If an organization encourages cross-disciplinary tasks, it is necessary to also specify which outputs must be reviewed by experts from the relevant fields.
Expanding a position does not equate to promotion; companies also need to redefine responsibilities, training, and rewards.
As employees continuously take on new tasks, job descriptions may fall behind reality. When an employee handles more marketing, analysis, or customer communication through the use of AI, they do not automatically gain new decision-making powers, training, or compensation. If a company only regards AI as a tool to increase output, the expansion of job responsibilities can become a form of hidden pressure. True job design should clearly define new responsibilities, quality standards, and pathways for advancement.
Managers cannot judge success solely based on usage volume. Frequent use may indicate that a tool is valuable, but it could also signify that processes need to be reworked repeatedly. Better indicators include task completion time, error rate, manual review costs, customer outcomes, and employee burden. Especially in cross-professional scenarios, the question is not just whether the model can provide answers, but who is responsible for the final results.
The study period is only four months, so it is not possible to determine whether these new tasks will be retained in the long term. The seasonal nature of the project, the freshness of the product, and organizational promotion may all contribute to short-term reuse. The continuous sample of about 6,200 people is also smaller than the overall population of 1.5 million messages. Further studies over a longer period of time, involving more industries and comparisons with non-user groups, will be needed to determine whether there has indeed been a change in the job structure.
Privacy measures are also worth noting. OpenAI states that information is aggregated and anonymized, and researchers do not read individual messages, which reduces the risk of direct exposure. However, job classification and task classification still rely on user and system data, and external readers cannot replicate the entire process from the announcements. The research conclusions should be regarded as early evidence based on internal platform samples, rather than a census of the entire workforce in society.
For employees, the most realistic impact of AI may not be "being replaced tomorrow," but rather "doing one more type of task today." This can increase autonomy and learning opportunities, but it may also blur professional boundaries. Individuals need to know when they can complete tasks independently with the help of AI, and when they should pass them on to professional colleagues; companies, on the other hand, need to ensure that cross-functional work does not bypass approval and qualification requirements.
This study does not prove that jobs will disappear universally, nor does it prove that work will necessarily become better. The more cautious conclusion it draws is that some employees are being forced to reuse their skills (reflected by AI) to handle tasks beyond their regular responsibilities, with marketing and customer communication activities being particularly likely to remain. Changes in the division of labor (reflected by AI) may first occur in the task lists beneath job titles. Whether these changes can be transformed into increased productivity and career development, rather than the spread of mistakes and additional unpaid work, depends on how organizations redesign their work processes.
