The AI Observatory, a public platform aggregating real AI conversations donated with user consent across seven datasets, applied Anthropic's work-focused classification approach to its own corpus. MIT Technology Review reported the results on 18 August.
What the filter removes
Nearly half — 48% — of conversations would be filtered out as non-work-related. The discarded portion is not a random remainder. Compared with the published analysis, excluded conversations showed markedly higher rates of health and relationships (44.2% against 31.2%), adult and illicit topics (7.9% against 2.1%), harassment and hate (27.5% against 5.66%) and sexual content (16.7% against 2.4%).
What the common telling gets wrong
This is not a finding that a lab hid anything. A methodology built to study economic use of AI is supposed to exclude non-work conversation; that is its stated purpose. The finding is about what happens when such figures escape their methodology and get quoted as descriptions of how people use chatbots generally. The half that gets removed is disproportionately the sensitive half — which means any public debate resting on work-classified usage statistics is arguing from a sample that systematically excludes the interactions most likely to concern it.
The comparison also has real limits, and they cut against the finding. The donors are self-selected — people who volunteer chat logs to a research project are not a random sample of users, and there is no obvious reason to assume they skew the same way as the general population. And the corpus covers 2023 to 2025, so it describes a previous generation of chatbots, not the models shipping now.
Why independent measurement is scarce
Almost everything the public knows about how people use AI comes from the companies selling it, measured on their own terms, on data nobody else can see. A donated-log corpus is a weak instrument by comparison — but it is one of the few that allows anyone outside a lab to re-run a published method and see what the method leaves out.