Independent research finds sensitive behaviors and clear model preferences not captured in company reports.

The AI Observatory, a new independent research project, has released data showing how people actually use AI models. The analysis reveals more sensitive behaviors than major vendors like OpenAI and Anthropic report, and it exposes significant differences among ChatGPT, Gemini, and Claude.

Researchers behind the project say companies publish only the data they want to share, and no independent source exists to verify the claims. The AI Observatory aims to fill that gap by tracking real-world interactions at scale. Early findings from the project show that people use Anthropic’s model more often for coding, Gemini for social and roleplay uses, and ChatGPT for homework assistance.

For builders and operators, these results matter because product design and benchmark expectations often rely on vendor-reported usage patterns. Independent data can reveal underserved use cases, privacy-sensitive behaviors, and the need for stronger safeguards. Teams should consider how model choice affects real user behavior, not just stated preferences.

The observatory is still in its early stages, and its goal is to provide ongoing independent analysis. Watch for broader studies, deeper model comparisons, and increased scrutiny of how AI products handle sensitive interactions.

What matters

• A new independent research project analyzes real AI usage patterns beyond company-reported data.
• Builders should note usage varies by model: coding on Anthropic, roleplay on Gemini, homework on ChatGPT.
• Watch for more independent AI usage studies as researchers continue to fill data gaps.

Why it matters

Watch for more independent AI usage studies as researchers continue to fill data gaps.

This GenAI News article was prepared in original wording using reporting and materials published by MIT Technology Review AI. Source reference: https://www.technologyreview.com/2026/08/18/1142229/the-download-how-people-use-ai-flock-cameras-design/.

Drafted by the GenAI News review pipeline.