Who Uses AI? Platform Selection and the Measurement of Occupational AI Exposure
Conversation logs from AI platforms are increasingly used to measure who is exposed to artificial intelligence at work. But the users observed in those logs are not the workforce. We show that platform-derived exposure scores conflate task-level applicability with the occupational composition of each platform's user base — and that this selection alone moves the post-ChatGPT employment coefficient by a factor of 1.9.
Figure 2. Theoretical AI capability vs. each platform's published exposure measure
Three-panel polar visualization across 22 SOC major occupation groups. The angular position fixes an occupation; the radial position is the published exposure score. Panel A: Massolinik & McCrory (2026) original composite measure. Panel B: Anthropic Economic Index (2026), IP API Wave 5 share. Panel C: Microsoft Copilot (Tomlinson 2025) AI applicability score. The same theoretical capability benchmark in blue is fixed across panels; the red shapes are what each platform's user base produces. Same underlying technology, three different pictures of the workforce.
Yin, M., & Ogut, B. (2026). Figure 2: Theoretical AI capability versus each platform's published main exposure measure. In Who uses AI? Platform selection and the measurement of occupational AI exposure (arXiv:2605.21743).
https://michelleyin.org/pub-who-uses-ai.html#figure2
Key Findings
The User Base Is Not the Workforce
Across three widely-used platform-derived AI exposure measures, the occupational composition of the users producing the score departs sharply from the BLS workforce. The published scores therefore mix two unrelated things: how applicable AI is to a task, and who happens to be on that platform.
The Platform Choice Drives the Result
Holding the empirical design fixed and changing only the platform input, the post-ChatGPT employment coefficient moves by a factor of 1.9. Consumer and enterprise channels within the same vendor disagree in sign. The conclusion an applied paper reaches depends on which conversation log it draws from.
This Is Non-Classical Measurement Error
We formalize the mechanism and decompose the bias into between-occupation selection (which occupations show up in the logs) and within-occupation selection (which workers in that occupation show up). Both move in the same direction and reinforce each other; neither is mean-zero.
Workforce Reweighting Closes Most of the Gap
Reweighting platform-derived scores to BLS employment shares attenuates downstream employment estimates by 42 to 93 percent across specifications. What is left over after reweighting is closer to the parameter applied work intends to estimate.
Augmentation, Not Substitution
What platform logs actually capture is how observed users augment their own work with AI. That is informative, but it is not the same parameter as substitution in the workforce. Conflating the two has produced employment forecasts that lean far harder on platform sociology than on AI capability.
Why It Matters
Forecasts from the BLS, OECD, IMF, and major consultancies increasingly cite platform-derived exposure scores. If the platform input drives the answer, those forecasts inherit a sociological selection effect they do not name. Workforce reweighting is a small empirical fix with a large policy implication.
This paper sits inside a broader research program at the RISEI Lab on the stability and identification of AI exposure measurement. The companion NBER Working Paper 35110 documents instability within a fixed measurement instrument across LLMs; this paper documents instability between instruments driven by platform user composition. Together they argue for partial-identification bounds and continuous recalibration, not point-estimate exposure scores.
Abstract
Conversation logs from AI platforms are increasingly used to measure occupational exposure to artificial intelligence, but the users observed in these logs are not the workforce. We show that platform-derived exposure scores combine task-level AI applicability with the occupational composition of the platform's user base. Holding the empirical design fixed, changing only the platform input changes the post-ChatGPT employment coefficient by a factor of 1.9, and consumer and enterprise channels within the same vendor disagree in sign. We formalize the resulting non-classical measurement error, decompose it into between- and within-occupation selection, and construct workforce-reweighted partial-identification bounds. Reweighting to Bureau of Labor Statistics employment shares attenuates estimates by 42 to 93 percent. The bias captures augmentation among observed users more directly than substitution in the workforce.