# Independent Research Reveals Truth About AI Usage Company reports may not reflect real AI usage patterns - **Source:** [MIT Technology Review](https://www.technologyreview.com) - **Published:** 2026-08-18 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/de978e47-8feb-464a-8d94-8d59741d4cbd ## Gaps in AI Usage Data Researchers from Stanford and MIT found that usage data published by major AI companies is limited and biased. The AI Observatory project analyzed 24,521 real conversations collected with user consent to create an independent data source. They emphasize that critical decisions about AI's benefits and risks are being made based on this incomplete data. The project includes 85,633 interactions with 52 different models between 2023-2025. ## Blind Spots in Company Reports Popular reports like Anthropic's Economic Index only show work-related AI usage. When Anthropic's filters were applied to AI Observatory data, 48% of conversations were excluded. Non-work uses included health (44.2%), adult content (7.9%), hate speech (27.5%) and sexual content (16.7%). OpenAI data also shows only 30% of consumer use is work-related. ## Changing Usage Patterns Over Time The research revealed that AI conversations became longer and more complex over time. In the WildChat dataset, both prompt and response token counts increased while small talk became more frequent - indicating growing demand for AI companionship. During the same period, chatbots' 'I am an AI' disclosures decreased while sensitive content conversations became less frequent, suggesting improved platform safeguards. ## Significant Differences Between Models Different AI models show vastly different usage patterns. Grok and Gemini are primarily used for information retrieval, with Grok being particularly popular for political news - but also a hub for misinformation. Claude is preferred for coding, Gemini for social roleplay, and ChatGPT for homework help. Even different versions of the same model (GPT-3.5 vs GPT-4o) showed noticeable differences in conversation length and iteration. ## Limitations of Data Collection Methods AI Observatory combined 7 existing datasets to analyze interactions from 5,000 users. However, this data pales in comparison to the millions of conversations held by major companies - Anthropic analyzed 1 million conversations, OpenAI 1.5 million for their reports. Voluntary data sharing methods may underrepresent sensitive uses. ## Shortcomings in Industry Reports Company reports fail to fully capture even the usage differences between their own models. MIT Media Lab's Shayne Longpre states 'No single company report tells the whole story.' While companies publish separate reports on specific use cases, having consolidated independent analysis helps understand AI usage more holistically. ## Model-Dependent User Behavior The research shows users employ different AI models for distinct purposes. Grok is favored for political news and information retrieval while becoming a misinformation hub. Claude sees more professional use, Gemini social interactions, and ChatGPT educational support. Even different versions of the same model (GPT-3.5 vs GPT-4o) show clear behavioral differences. ## The Future of AI Usage Data The AI Observatory project aims to provide more transparent and comprehensive data about AI usage. Researchers stress the need for independent data sources to support policymaker decisions. However, the voluntary data sharing method risks underrepresenting sensitive use cases. They caution that findings may not reflect all AI usage patterns. ## Sources covering this story - https://www.technologyreview.com/2026/08/18/1142226/how-people-use-ai/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._