OpenAI Signals Work Dataset: Coverage, Privacy and Limits

OpenAI Economic Research released version 3.0 of its Signals data on October 26, 2026. The dataset connects classified work-related messages from individual ChatGPT plans to 165 Intermediate Work Activities derived from the O*NET framework. It covers nearly 100 countries in quarterly observations from the first quarter of 2025 through the third quarter of 2026. It is a dataset about a subset of ChatGPT use, not a census of work, employment or productivity.

The dataset measures messages, not jobs completed

OpenAI says researchers can compare activity mixes across countries and over time. Across 74 countries observed in every quarter, the ten most common activities accounted for an average 51.1% of classified work messages in the third quarter of 2026, down from 56.0% in the first quarter of 2025. The company interprets this as work-related use becoming less concentrated.

A message classification cannot reveal whether the answer was correct, adopted, valuable or harmful. One task may require many messages while another needs one. The country mix can also reflect plan access, language support, occupational structure and user selection. Researchers should avoid turning message share into a claim about labour-market share or productivity.

Enterprise and Codex activity are outside the release

The release covers work-related messages on individual ChatGPT plans and explicitly excludes enterprise messages and Codex use. It also notes that roughly 43% of the employment-weighted task mix for mapped US occupations falls into other activities outside individually reported categories, including material handling, vehicle operation and some medical-record work.

Those gaps matter for Canadian comparisons. Knowledge work visible in consumer ChatGPT may be overrepresented relative to physical, regulated or employer-managed work. A country chart should therefore be joined with independent labour-force, occupation and access data rather than read as a complete national AI-adoption index.

Aggregation and noise reduce identification risk but also affect precision

OpenAI says researchers do not read individual conversations for this dataset. Automated systems classify messages after identifying information is removed; the release aggregates results, withholds small groups and adds statistical noise. These are documented privacy protections, but the public page does not let an outside reader independently audit every classifier or privacy parameter.

Analysts should keep uncertainty visible, avoid tiny subgroup comparisons and retain the original release version and access date. A change smaller than the effect of classification or privacy processing may not support a strong conclusion.

A responsible Canadian analysis needs external denominators

Useful follow-up could compare Canadian activity shares with Statistics Canada occupation counts, education, language and industry measures. It should distinguish user behaviour from employer adoption and task frequency from economic value. Regional findings need sufficient sample support rather than assumptions based on national averages.

Signals can generate questions about how people use ChatGPT for work. It cannot by itself answer whether AI replaced jobs, raised wages, improved output or reached workers evenly. Those outcomes require other datasets and causal research.

Sources and editorial method

We use the linked original statements and reports to distinguish documented facts from our analysis. Company announcements describe their authors’ plans and claims; they are not independent verification of future outcomes. This article is dated and will be revised if material evidence changes. Read our editorial standards and corrections process.

OpenAI: ChatGPT’s expanding world of work ↗ — Official October 26 release; scope, time range, coverage, exclusions, privacy method and summary statistics.
OpenAI Signals data download and methodology ↗ — Official versioned data and methodology page linked from the release.