ChatGPT at work: What OpenAI Signals shows about 165 work activities
ChatGPT at work: OpenAI releases data on 165 activities across nearly 100 countries. What the data shows and why it cannot prove job replacement.
On October 26, 2026, OpenAI released Signals data mapping work-related ChatGPT messages to 165 standardized activity categories across nearly 100 countries. The dataset covers individual ChatGPT usage, not all workplace AI: enterprise messages and Codex are excluded, and usage shares do not demonstrate job displacement.
CLAIM-LEVEL EVIDENCE
Evidence and remaining unknowns
- Primary source:
OpenAI – announcement
OpenAI provides aggregate data for 165 activities.
- Open question:
Evidence limits
OpenAI covers messages on individual ChatGPT plans, excluding enterprise and Codex. Prompt patterns do not measure productivity, job replacement or agent reliability.
OpenAI covers messages on individual ChatGPT plans, excluding enterprise and Codex. Prompt patterns do not measure productivity, job replacement or agent reliability.
A more detailed map of work-related prompts
ChatGPT at work now has a more detailed observational dataset, although it is not a census of employment. OpenAI uses 165 Intermediate Work Activities from the O*NET framework to classify work-related ChatGPT messages. The categories span tasks such as preparing teaching material, developing marketing content, programming and examining business data. The company says the new release includes almost 100 countries and quarterly observations from the first quarter of 2025 through the third quarter of 2026.
The appeal of this framework is comparability. Researchers can look at the activities people bring to ChatGPT rather than only counting total usage. Yet a classified prompt establishes neither task completion nor the quality of a finished result. One message might ask for a suggestion, another might request a draft, and an actual employee may still need to verify every step. It is a measurement of interaction patterns, not an autonomous-agent performance test.
How to read the 51.1% finding
OpenAI reports that, across 74 countries included in every observed quarter, the ten most common work activities accounted for an average of 51.1% of classified work messages in the third quarter of 2026, down from 56.0% in the first quarter of 2025. In this dataset, usage appears less concentrated in a handful of categories. That does not establish whether jobs are disappearing, whether output quality has improved, or whether one country has gained an economic advantage.
Country differences can reflect different kinds of employment as well as differences in access to AI tools. OpenAI acknowledges both possibilities. Comparing shares of ChatGPT messages directly with shares of all workplace activities would also be misleading. Some physical, in-person and record-keeping tasks are not individually represented in the published classifications, and the original sampling frame does not cover every workplace technology.
Privacy protection and what is missing
The release describes aggregate statistics derived from individual ChatGPT plans. It does not include enterprise conversations or Codex usage. That omission matters for studying software engineering teams and agents that run longer workflows with company tools. OpenAI says automated systems classify messages after identifying information has been removed, with small groups suppressed and statistical noise added. Researchers do not read individual user conversations in order to build the published tables.
AgentenCode assessment: The dataset offers evidence about demand for assistance across kinds of work, and it is a useful starting point for research questions. However, it cannot on its own tell employers which AI agents are dependable, which permissions are safe, how much human review is necessary, or how hiring responds. Those questions require independent workplace observations, outcome measurements and testing of specific agent systems.
Primary sources and further reading
AgentenFit: explore suitable tasks →
A primary announcement is not an independent field test.
Frequently asked questions
Does it cover every job?
No. OpenAI categorizes messages into 165 task groups.
Are Enterprise and Codex included?
No. OpenAI expressly excludes both.
Does this establish layoffs?
No. Message classifications do not measure job losses.