What AI publishers' usage reports don't show: 48% of conversations discarded by the filter

An independent observatory has applied Anthropic's methodology to its own conversation data. The result: the filtered share is massive, and personal uses are much more present than advertised.

Anthropic and OpenAI regularly publish analyses of how the public uses their assistants. These reports are, to date, the main window available on the actual uses of general AI — and no one can cross them. This is the starting point of Eileen Guo's article for the MIT Technology Review, published on August 18, 2026.

The new AI Observatory was an independent body: seven aggregated data sets, 24,521 conversations and 85,633 dialogue rounds, from approximately 5,000 users and 52 models. By applying to this corpus the filtering that the Anthropic Economic Index uses to rule out non-professional uses, 48% of conversations disappear — a proportion that does not appear anywhere in official reports.

The differences are concentrated where they are least expected in a productivity-oriented narrative: health and personal relations account for 44.2% of the exchange in the observatory's data, compared with 31.2% on the official report side, and sexual content 16.7% against 2.4%. On 2023-2025, conversations lengthen, ordinary conversation progressed, sensitive subjects receded. As researcher Anka Reuel (Stanford) points out, there is no way of independently corroborating the figures published by publishers — and the corpus collected on the basis of volunteerism probably under-represents the most intimate uses as well.