The News, Explained
OpenAI analyzed more than 1.5 million work-related ChatGPT messages from April through July 2026 to study how workers used AI for tasks outside their usual job scope. The analysis found cases where people tried an out-of-role task and then used AI for the same kind of task again in a later month. Source
Repeated out-of-role tasks
Among roughly 6,200 workers observed throughout April to July, the share of out-of-role AI activity involving tasks they had used before rose from 13.1% in April to 25.9% in July. The researchers said the pattern is consistent with some tasks moving from one-time experimentation into repeated workflows. Source
In a separate one-month follow-up sample, workers repeated an out-of-role task they had done in the previous month at a rate of 23.6%. Comparable workers with no observed use of that task in the previous month used it at a rate of 8.4%. Source
Reuse rates varied by task
Next-month reuse reached 54% for discussing products and services with customers, 44% for writing advertising or promotional copy, and 37% for producing marketing materials. Explaining financial information was around 15%, while the overall average for out-of-role tasks was 18.5%. Source
Prompting patterns also differed
For out-of-role tasks, workers used shorter prompts on average than they did for in-role tasks. They asked less often for explanations, procedures, specific formats, or advice, and were more likely to provide examples or background information and ask the model to review or verify an output. The researchers interpreted this as a possible pattern of drawing on knowledge from another domain and applying it to the current problem. Source
Data included in the study
The study used role and department information supplied during ChatGPT Business onboarding to classify occupations and analyzed a subset of conversations that could be classified. Onboarding is the initial setup process in which users or organizations configure account and role information. The analysis was based on aggregated and anonymized data. Aggregated data combines many records into summary statistics, while anonymization removes or obscures information that could identify individuals. Data with training use disabled and messages without the classification information were excluded. Source
OYOPICK’s Take
Repeated use is worth treating as a signal that a new task is entering a workflow. Output quality and responsibility remain questions to evaluate alongside it. For a hypothetical team repeatedly drafting promotional copy outside its usual roles, reviewer corrections and whether drafts are used could be recorded separately from usage counts. This is a proposed local evaluation.
The sample covers only classifiable ChatGPT conversations. Within that limit, planning knowledge-sharing and responsibility for recurring new tasks could help AI support collaboration across roles.