Study reveals major AI usage data blind spots in company reports
MIT Tech Review AI · rss · 2026-08-18
The AI Observatory project by MIT and Stanford analyzed 24k real conversations, finding that vendor reports like Anthropic's filter out 48% of non-work interactions. Filtered data showed much higher rates of sensitive topics like health, harassment, and adult content. Usage patterns also vary significantly by model: Grok for news, Claude for coding, and ChatGPT for homework.
More from Safety
- Uncensored Models: Tools for Red-Teaming, Not for Deployment — Moin_Chaudhary26 · 2026-08-18
- Apple's Tahoe 26.6.2 credits a surprising number of bug fixes to OpenAI's Codex team — mircomusolesi · 2026-08-18
- Researcher: dodging facts about China's AI leads DC into a parallel reality — pstAsiatech · 2026-08-18
- Offensive hacking agents see zero incidents, sparking safety debate — OwariDa · 2026-08-18
- MIT study: AI usage reports are biased, work use cases overestimated — nordicinst · 2026-08-18
- a16z Podcast: Defending Against AI That Can Hack — a16z Podcast · 2026-08-18