Recommended Free Tools
Clio is Anthropic’s research platform for finding aggregate patterns in how people use Claude, using AI-assisted analysis rather than relying on human reviewers to read raw conversations as the main path. Its published usage findings are a historical snapshot—not a current census—and its privacy protections are layered safeguards, not a formal differential-privacy guarantee.
What Clio is designed to do
Clio stands for “Claude insights and observations.” Anthropic describes it as a system that summarizes conversations, groups related summaries into clusters, and surfaces broad usage patterns. The goal is to help answer questions such as which capabilities people are adopting and how reported use varies across communities, without making manual reading of raw conversations the primary analysis method.
As an Amazon Associate I earn from qualifying purchases.
The system also supports safety analysis. That use should not be conflated with aggregate research: Anthropic says safety-focused runs can link results to individual accounts, and access to those results is restricted to a small number of authorized staff. Its Privacy Center says the research-oriented aggregate use provides no access for Anthropic employees to raw user conversations or customer-specific data. Anthropic’s explanation of Clio’s research and safety uses is dated March 16, 2026.
What the published usage findings cover
The main Claude.ai analysis in the paper used a sample of 1,000,000 Free and Pro conversations collected from October 17–24, 2024. It excluded Team, Enterprise, and API customers from this published individual-use analysis. These boundaries matter: the results describe that sample and period, not every Claude user or current usage.
#1 Best Overall
Within that sample, the paper reports:
- Web and mobile application development accounted for more than 10% of conversations.
- Writing, research, and educational use each accounted for 6–10% of conversations.
- Some topic patterns differed by language; elder-care discussions were elevated in Japanese- and Chinese-language conversations.
A separate multilingual analysis sampled 2,281,911 conversations from October 24–November 13, 2024. Its population and sampling window differ from the one-million-conversation Claude.ai analysis, so its figures should not be merged with that sample as if they were one continuous census. The authors also report that Claude.ai, LMSYS-1M-Chat, and WildChat have different mixes of non-coding use cases. Platform, language, sampling method, and time period all shape what a dataset reveals.
How the privacy safeguards work
The paper describes a sequence of safeguards across Clio’s analysis pipeline. They are intended to reduce exposure of private information in the summaries and aggregate outputs:
- Conversation summaries: The model is instructed to omit private information when summarizing conversations.
- Cluster thresholds: Clusters that do not meet minimum requirements for unique accounts and conversation counts are discarded.
- Cluster summaries: The model is again instructed to leave private information out when summarizing a cluster.
- Audit and removal: Cluster summaries are audited, and clusters containing private information are removed.
This is a defense-in-depth design. The authors say that formal guarantees such as differential privacy and k-anonymity are difficult to apply to rich text summaries, and characterize Clio’s approach as statistical and empirically evaluated. The safeguards therefore should not be read as proof that re-identification is impossible under every threat model.
What the validation does—and does not—show
The paper reports 94% accuracy in reconstructing topic categories on a synthetic dataset with a known topic distribution. That is evidence about performance on that benchmark; it is not a claim that every real-world conversation is classified correctly at that rate.
For a privacy-layer audit, the authors examined 5,000 conversations. The audit and synthetic benchmark provide useful evaluation evidence, but they do not establish a universal privacy guarantee. The authors explicitly discuss the difficulty of formal guarantees for rich textual outputs, alongside potential misuse and ethical concerns.
Clio’s safety-monitoring role
Beyond broad research on usage, the paper describes using Clio to look for coordinated abuse, identify unexpected issues around launches or world events, and analyze safety classifiers. One example involved finding automated accounts engaged in search-optimized spam, which were then banned. This illustrates a safety application; it does not mean that all aggregate research outputs are account-linked or that every analysis run has the same access model.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to interpret Clio’s results
When assessing a claim based on Clio, check what population was analyzed, when conversations were sampled, how the sample was selected, and whether the output is aggregate research or a safety-focused run. Also distinguish empirical privacy testing from a formal guarantee, and synthetic benchmark accuracy from performance on an entire real-world population. Those details determine what a reported pattern can reasonably support.
The underlying paper, “Clio: Privacy-Preserving Insights into Real-World AI Use,” was posted on December 18, 2024. Its usage figures are historical sample findings; they should not be presented as a current distribution of Claude use. The paper also gives an estimated cost of $48.81 for an example run over 100,000 conversations. That is an estimate for the paper’s example, not a current Clio service price.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
