No AI Reveals the Data Used for Training

None of the 13 audited platforms allow for the retrieval of data once it has been used to train a model, and no retroactive withdrawal of consent is possible. In this context, Incogni publishes a ranking of the leading AI services, from least to most risky for privacy, with the idea that larger products expose users more, except for some exceptions.

The evaluated platforms do not return data after training

Incogni states that none of the reviewed platforms offer the option to retrieve personal data once it has been integrated into the training of a language model. It is not possible to perform a retroactive deletion or to withdraw consent afterward.

Vibe is ranked safest; Kimi, Meta, and Copilot concentrate the risks

Vibe (formerly Le Chat) stands out with easy-to-understand policies, mobile applications deemed privacy-friendly, data collection presented as sourced from public information, and a sharing of conversations limited to a "minimal" number of third parties. In contrast, Kimi is credited with the highest risk and makes disengagement from training difficult according to Incogni. Meta is labeled as the most "data-hungry" model, and its claims of training confined to "private conversations" are hard to verify. Copilot, on the other hand, shares data with third-party advertisers and relies on a policy considered too broad, leading to transparency issues.

Transparency and disengagement vary by service

ChatGPT displays a policy deemed readable and accessible, accompanied by a FAQ and resources, and indicates that withdrawal from training is presented as easy, while noting that information is collected and may be shared for security or marketing purposes. Claude provides a detailed policy along with resources but now requires users to take the initiative to withdraw so that their data is not used for training purposes. Perplexity offers an easy-to-follow but somewhat vague policy; opting out of training is straightforward, while sharing within the group and with affiliates remains possible. Qwen details the data used and its purposes but leaves unclear what is shared and with whom. Pi combines a complex policy, specific treatment of EU users for GDPR compliance, potential sharing with business and research partners, and an application considered privacy-friendly. Regarding Gemini, Incogni notes that Google can already aggregate vast amounts of data through other services and describes its privacy policy as "partial." Finally, Grok may transmit data to group entities and also trains on public conversations.

Local or hosted: the open-weight option changes the game

DeepSeek allows for data to be kept on local networks via open-weight models, while hosted versions raise more serious privacy considerations. Z.ai follows a similar logic: open weights for local use, but specific privacy implications for hosted deployments.

Incogni evaluates 13 AIs and values the least intrusive practices

Incogni's ranking covers 13 platforms: ChatGPT, Claude, Gemini, Grok, Vibe, Perplexity, Qwen, DeepSeek, Z.ai, Kimi, Meta AI, Pi, and Copilot. Each service receives a score where a lower level means a lower risk, with the scale taking into account the discovery of intrusive practices, the readability of sharing practices, and the fate of user data. The axes of analysis cover what happens to the data, the degree of transparency, and the management of information on the platforms. According to Incogni, larger platforms tend to be more invasive, while noting a notable exception. The organization clarifies, however, that its results do not mean that a given service is automatically the best or worst option for all uses. The report is published under the title "Gen AI and LLM Data Privacy Ranking 2026."