# OpenAI Signals Work Data: What 165 Activities Can Tell Us

OpenAI's new work-activity dataset makes ChatGPT usage easier to study, but interpreting it starts with understanding what the numbers leave out.

By Paras Madan | 2026-10-27


OpenAI released new Signals data on October 26 covering 165 work activities across nearly 100 countries. It tracks work-related use of consumer ChatGPT. For founders and AI teams, its value is in understanding workflows, provided message counts are not mistaken for economic outcomes.

## What OpenAI Signals added

The [new release](https://openai.com/index/chatgpt-expanding-world-of-work/) maps usage to O*NET Intermediate Work Activities, categories shared across occupations. Its quarterly work-activity series runs from Q1 2025 through Q3 2026.

Across 74 countries present in every quarter, the average share of classified work messages accounted for by each country's ten most common activities fell from 56.0% in Q1 2025 to 51.1% in Q3 2026. That describes a less concentrated mix of work-related usage. It does not establish that more jobs were automated.

The distinction matters when deciding what to build. A task becoming more visible in conversations could be a useful lead for customer research. It cannot tell a product team whether people will pay for a dedicated tool, whether that tool saves time, or whether its output survives review.

## Read the denominator before the percentage

The [data dictionary](https://cdn.openai.com/signals/data-dictionary.pdf) defines the work-activity measure as a share of work-related messages within a country and quarter. The wider dataset samples one million consumer messages monthly from accounts declaring an age of at least 18. It excludes deleted accounts, training opt-outs, enterprise messages and Codex.

Work categories are model-inferred labels, not users' self-reported occupations. Researchers use automated classification of de-identified content; published aggregates include differential-privacy noise. Smaller groups have less precise estimates.

**A message share is not a worker share.** Consider a hypothetical developer who needs ten exchanges to resolve a bug and a marketer who needs one exchange to revise a paragraph. Counting their messages would not establish that the developer received ten times the value, spent ten times as long working, or represents ten times the market opportunity.

A category's share can also decline while its absolute usage grows, if other categories grow faster. Before describing a workflow as losing relevance, an analyst would need evidence beyond its percentage of the mix.

## Why the 43% “Other” figure matters

OpenAI says roughly 43% of the employment-weighted task mix for mapped US occupations sits in “Other IWAs”, outside the individually reported categories. Those activities failed its consistent-representation threshold in the original US ChatGPT sample.

That makes workforce comparisons easy to overstate. A chart organized around observable chatbot activity will naturally be more informative about some tasks than others. An activity appearing rarely in this view is not evidence that it has little economic value, nor proof that AI cannot assist it through another interface.

For an enterprise team, the practical implication is to separate discovery from evaluation. Use the dataset to generate questions about a workflow, then investigate the actual process: what inputs workers have, where mistakes matter, who approves the output and how success is measured. That second step needs evidence from the organization concerned.

## How to use the data responsibly

OpenAI provides [downloadable CSVs and methodology](https://openai.com/signals/data-download/) under CC BY 4.0, allowing reuse and adaptation with attribution. That makes the release useful for researchers who want to build their own comparisons rather than rely on an announcement's charts.

A sensible analysis starts with one activity, a clearly defined comparison group and a time period. Keep the same denominator throughout. Explain missing coverage beside the chart, not in a distant footnote, and avoid treating a small movement as a meaningful change without checking uncertainty.

For product planning, I would treat a pattern here as a reason to interview users, not a substitute for doing so. The useful next question is whether a recurring conversation reveals a workflow that deserves a better product. Answering that requires connecting observed interest to a measurable result outside the chat window.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn complex AI research and product updates into clear stories that developers and founders can use and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min)