Incogni ranks 13 AI assistants by privacy risk: bigger is worse, except ChatGPT

Incogni, a data broker removal service, published its "Gen AI and LLM Data Privacy Ranking 2026" on Thursday, scoring 13 generative AI platforms on the privacy risks they pose to consumers ^[1][2]. The headline finding is a size correlation with one break in it: the largest platforms score worst, with the exception of OpenAI's ChatGPT ^[3].

The scoring mechanics matter more than the league table. Incogni assigns each platform a score in which lower is better, and each added point reflects the discovery of privacy-invasive practices, how easy data-sharing practices are to find and understand, and what happens to user data ^[4]. The study is organised around three question groups: whether conversations feed model training and whether users can opt out or have prompts passed to third parties; whether policies are easy to find and understand; and what personal data is collected, where it comes from, and where it goes ^[5]. That construction means a vendor can improve its position partly by writing clearer documentation, not only by collecting less. For a team that has to justify which assistant gets access to internal data, the transparency component is the cheapest thing to verify yourself and the least predictive of what actually happens to a prompt.

The ordering, from lowest to highest risk, starts with Mistral AI's Vibe, formerly Le Chat, credited with easy-to-understand privacy policies, privacy-friendly mobile apps, collection from public sources, and conversations shared with a "minimal" number of third parties ^[6][7]. ChatGPT is next, with a clear privacy policy plus FAQs and resources, a straightforward opt-out from data collection for model training, and explicit statements that data is collected and may be shared for security or marketing purposes ^[8]. Inflection AI's Pi follows, with a complex policy that handles EU users separately for GDPR, potential sharing with corporate, commercial and research partners, and a privacy-friendly app ^[9]. Perplexity is fourth: an easy-to-follow but not very detailed policy, an easy-to-find training opt-out, and data that may be shared with corporate group members and affiliates ^[10]. Alibaba's Qwen is fifth, described as naming purposes such as sharing personal data with analytics providers, search engine providers and other third parties while leaving it unclear exactly what is shared and with whom, and difficult to determine whether conversations are used for training ^[11].

Gemini and Meta AI received some of the highest risk scores; Vibe, ChatGPT and Pi some of the lowest ^[12]. Beyond that, the itemised detail available covers five of the 13 platforms, leaving eight positions unspecified, including Claude, Copilot, Grok, DeepSeek, Z.ai, Kimi, Gemini and Meta AI ^[13].

Two caveats before this goes into a vendor matrix. Incogni sells data removal services, so privacy salience is its business ^[1]. And the assessment is framed around consumer privacy ^[2], which is not the same document set as an enterprise agreement with data processing terms.

Worth watching: where Claude and Copilot land, since both sit inside enterprise stacks; whether the opt-out paths credited to ChatGPT and Perplexity ^[8][10] hold at the API and workspace tier rather than only in consumer apps; and the disclosure attached to the report's coverage, that ZDNET parent Ziff Davis filed an April 2025 lawsuit against OpenAI alleging infringement of its copyrights in training and operating its AI systems ^[14].