ทันเอไอ

Study reveals major tech companies' AI usage reports hide nearly half the truth

AI usage reports published by developer companies may not tell the whole story, after independent researchers found that nearly half of personal data was filtered out.

In a world where AI plays a major role, data on how people use Generative AI largely comes from reports by the developers themselves, such as OpenAI and Anthropic, which typically focus on work and productivity.

However, Anka Reuel, a computer science PhD student at the Stanford Trustworthy AI Research (STAIR Lab) and the Stanford Internet Observatory, noted that this data lacks independent verification.

Recently, researchers from MIT Media Lab, Stanford University, and the independent research network Data Provenance Initiative (DPI) launched the "AI Observatory" project to transparently collect and analyze real-world AI usage data. They analyzed 24,521 consented conversations (totaling 85,633 interaction turns) across 52 language models.

The analysis found that when applying data filtering methodologies used by AI companies (such as the Anthropic Economic Index), approximately 48% of actual conversations were discarded. The filtered content tended to be personal matters—such as health, relationships, and sensitive issues—indicating that official reports may not reflect the full picture of how people actually use AI.