What happened
Pebblous says OpenAI’s Signals v3.0, published on 26 October 2026, contains 24 CSV files and a data dictionary under a CC BY 4.0 license. The dataset includes country-by-quarter work-activity tables covering seven quarters, from the first quarter of 2025 through the third quarter of 2026. OpenAI describes the release as covering work activities drawn from O*NET’s Intermediate Work Activities (IWA) taxonomy.
To test the published activity count, Pebblous compared the classification prompt in section 9 of OpenAI’s v3.0 data dictionary with the O*NET 30.3 activity list. It reports that the prompt requires the classifier to select one of 332 IWAs, except where a message is unclear or ambiguous. Pebblous says all 332 prompt labels matched the O*NET list. It then counted the distinct activity names in the published tables: 165 values, including the catch-all “Other IWA.” That leaves 164 activities with individual names and 168 grouped into the catch-all category.
The data dictionary calls this “rare-code grouping” and says rare IWAs are grouped into “Other IWA” for privacy and robustness. Pebblous reports that the dictionary does not identify which activities were grouped or the count in that category. In its recount, five O*NET parent groups had no individually named activities remaining: machine and process control; coaching, directing, and motivating subordinates; repairing and maintaining mechanical equipment; repairing and maintaining electronic equipment; and operating vehicles, mechanized devices, or equipment. Other groups retained most of their names. The report characterizes the broad pattern as desk-work categories remaining visible while hands-on and supervisory work are more heavily grouped.
Pebblous makes a separate comparison between the grouped activities’ share of ChatGPT work messages and their share of US occupational task composition. It reports that “Other IWA” accounted for 2.40% of classified US work messages in the third quarter of 2026. The approximately 43% figure in OpenAI’s announcement refers instead to the employment-weighted task composition of mapped US occupations. Pebblous says it independently approximated that figure at 42.10%–42.95% using O*NET 30.3 and US Bureau of Labor Statistics employment data, across six calculation variants. It notes that this is not the same method as OpenAI’s analysis.
The report also emphasizes the scope of the Signals dataset: its activity tables cover personal ChatGPT accounts, exclude Codex and enterprise accounts, and use sampled messages. Pebblous says the sample includes adult personal accounts and that aggregate values receive differential-privacy noise before normalization. These scope boundaries matter when interpreting the dataset as a measure of work involving AI tools more broadly.
What the archive adds
New Runtime’s earlier coverage noted that OpenAI has published provenance signals across content produced with its tools (24 October 2026). That is not the same event or data question. Pebblous’s recount opens a new thread: the visibility limits built into OpenAI’s work-activity taxonomy.
Trend signal
This supports the pattern that agent economics moves to completed work: measures of AI’s economic footprint depend on defining the work unit being counted. The report does not measure agent task completion or cost, but it shows how aggregation choices can obscure which kinds of work appear in a dataset. The link is a methodological connection, not evidence that Signals measures completed-task economics.
Why it matters
A dataset can be open and still present a narrower picture than its classification system suggests. If a researcher, company, or policymaker reads the named categories as a complete map of AI-assisted work, the “Other IWA” bucket can hide precisely the activities whose adoption is less common or harder to classify. Pebblous’s examples point to hands-on and supervisory work, not just a generic loss of detail.
For teams building internal AI-adoption dashboards, the practical lesson is to track the taxonomy as well as the headline metric. Record the classifier’s full label space, the reported labels, the size and trend of catch-all buckets, and the share of messages that cannot be classified. Where possible, test conclusions against an external task taxonomy and distinguish personal-product usage from enterprise and coding-agent activity. A rise in a named category, or a fall in one, is not a complete account of work moving into or out of AI.
Limits and uncertainty
Pebblous’s analysis is a recount of OpenAI’s published tables and a separate reconstruction of the employment-weighted estimate; it is not an audit of OpenAI’s underlying message sample or classification pipeline. Its 43% reconstruction uses O*NET task ratings and employment data, which Pebblous says differ from OpenAI’s method. The source also reports low agreement among human annotators for fine-grained IWA labels in prior validation work, making individual category estimates harder to interpret than broad trends. The precise contents and individual prevalence of the 168 grouped activities remain unknown from Pebblous’s account of the public tables.
Reader takeaway
When an AI-use dataset reports work by category, ask what the classifier could choose and what the published table actually names. In Signals v3.0, those are not the same: 332 possible activities became 164 individually visible labels plus “Other IWA.” That gap is part of the measurement, not a footnote.
Open question
Will OpenAI publish a more detailed breakdown or sensitivity analysis showing how rare-code grouping affects comparisons across countries and quarters?