ANALYSIS | What new ChatGPT data reveals about South African work use

Writing remains a major part of South African work-related use of ChatGPT, but newly updated OpenAI data also points to a growing share of messages about visual design. The figures offer a view of what people ask the service to help with — not a count of jobs automated or hours saved.

EBNewsDaily’s analysis of the South African rows in OpenAI’s Signals v3.0 dataset, updated on 26 October, found that 39.3% of messages in September 2026 were classified as work-related. The remaining 60.7% were classified as non-work use.

Those percentages describe messages in the dataset, not the proportion of South Africans using AI at work. They also do not describe the share of workers who use ChatGPT.

Writing leads, with design gaining ground

Among South African work-related messages in September, writing accounted for 39.1%. Practical guidance followed at 24.1%, while seeking information represented 10.5% and technical help 8.1%.

A separate, more detailed quarterly breakdown groups work-related messages by activity. In that table, editing written material or documents was the largest South African category in both the first quarter of 2025 and the third quarter of 2026.

Its share fell from about 22.0% to 19.1% over that period. Messages classified as creating visual designs or displays rose from about 2.3% to 8.1%, making that the second-largest category in the latest quarter.

The comparison suggests a change in the mix of tasks represented in South African work-related messages. It does not establish that fewer editing requests were sent in absolute terms: a category’s share can fall even while its total volume grows. The published tables used here do not supply those absolute volumes.

The broad monthly writing category and the more specific quarterly editing category also measure different things. Their percentages should not be treated as interchangeable.

What the figures cannot tell us

OpenAI says its global dataset draws on one million consumer ChatGPT messages per month, from July 2024 to September 2026. That is the global sample, not a million South African messages.

It covers individual accounts with a declared age of at least 18. Enterprise messages and Codex use are excluded, as are users who deleted their accounts or opted out of allowing their data to be used for training. Those exclusions limit how well it describes the wider workplace.

Work-relatedness and activity labels are inferred by automated classifiers rather than supplied as survey answers by users. OpenAI says researchers did not read raw message text and that the published statistics incorporate privacy-protecting noise. Smaller subsets have greater variability, so small differences should not be overstated.

For South African employers and workers, the useful distinction is between evidence of demand for assistance and evidence of a completed result. These data identify categories of requests. They do not show whether the output was accurate, whether a task was finished, whether productivity improved or whether anyone’s employment changed.

The strongest conclusion is narrower: writing remains prominent in this slice of South African ChatGPT use, while visual-design requests occupy a larger share than they did at the start of 2025.

Data: OpenAI Signals v3.0, by Aaron Chatterji and colleagues, licensed under CC BY 4.0. EBNewsDaily calculated percentages and comparisons from the published South African data.