Researchers also identified a decline in what they call “sensitive use”—conversations involving potentially harmful or restricted content, including sexual harassment and hate speech. This trend may indicate that AI platforms are improving their safety features and content moderation systems.
The AI Observatory also found that AI usage varies significantly by model. Users differed in the topics they explored, their interaction styles, and the structure of their conversations. The likelihood and type of sensitive AI use cases also varied depending on the chatbot.
For example, users were more likely to use Grok and Gemini for information searches. Grok was especially popular for news and political information, but researchers also identified it as a potential hotspot for misinformation. This finding aligns with other research showing how easily misinformation can spread on Grok. xAI did not respond to requests for comment.
By comparison, users were more likely to use Anthropic’s Claude for coding, Gemini for social interaction and role-playing, and ChatGPT for homework assistance.
Researchers also observed differences between versions of the same AI model. Conversations were generally shorter when ChatGPT was powered by GPT-3.5, while GPT-4o generated longer and more repetitive interactions. This finding is notable because GPT-4o has faced concerns related to emotional dependence and user attachment.
However, corporate AI usage reports have not typically captured these differences between—and even within—models. “A single company report never tells the whole story,” said Shane Longpre, a recent MIT Media Lab doctoral student who led the study with Royer.
To create the AI Observatory, Reuel and researchers from MIT, Stanford University, the Data Provenance Initiative, and other institutions combined 24,521 conversations—consisting of user prompts and corresponding AI responses—across 85,633 conversational turns. The data came from seven real-world datasets collected in previous studies and represented 5,000 users who interacted with 52 AI models between 2023 and 2025, including ChatGPT, Gemini, Claude, and Grok.
Still, these conversations represent only a small portion of the data available to major AI companies. For example, Anthropic’s latest Economic Index report analyzed 1 million Claude conversations, while OpenAI’s How People Use ChatGPT report examined 1.5 million conversations.
Source: www.technologyreview.com