A new study ranking AI platform privacy risks across 13 of the most widely used large language models has found that Inflection AI’s Pi, not Mistral AI’s Vibe, achieved the lowest cumulative privacy-risk score overall in Incogni’s Gen AI and LLM Data Privacy Ranking 2026.
The report, published by data broker removal service Incogni, assessed ChatGPT, Claude, Gemini, Grok, Vibe (formerly Le Chat), Perplexity, Qwen, DeepSeek, Z.ai, Kimi, Meta AI, Pi, and Copilot. Each platform was assigned a score based on the privacy risks it poses; the lower the score, the better.
The conflict between the report’s implied ranking and Incogni’s own cumulative scores matters: Vibe placed first in the ‘what happens to user data’ category, scoring just below 4 points out of a maximum 9, while Pi led on the overall ranking. Z.ai placed third among top performers.
What Drives AI Platform Privacy Risks Up, and Down
Incogni’s methodology grouped its questions into three areas: what happens to user data (whether conversations train models, and whether prompts are shared with third parties); how transparent platforms are (whether privacy policies are easy to find and understand); and what personal data is gathered and where it goes.
For the top-performing platforms, restricted data sharing with third parties and highly limited app data collection practices played a major role in keeping overall privacy-risk scores low, according to the Incogni report.
At the other end of the table, Microsoft’s Copilot, Meta AI, and Moonshot AI’s Kimi drew the harshest assessments. Incogni described Kimi as representing ‘the greatest privacy risk among the platforms assessed’ with regard to user data handling, with its app using data to track users and limited ability to opt out of model training. Meta AI was considered the most ‘data-hungry’ model in the sample, with claims that training occurs only in private conversations rather than group chats described as difficult to verify. Copilot, meanwhile, was found to share data with third-party advertisers and to purchase data from data brokers, compounding transparency concerns over its broad privacy policy.
Google’s Gemini relies on what Incogni calls a ‘partial’ privacy policy that defaults to Google’s general policy, making platform-specific answers difficult to obtain. xAI’s Grok trains not only on Grok user data but also on public conversations from X.
Anthropic Shifts Claude’s Default on Training Data
One of the more consequential recent changes to any AI platform’s data practices came from Anthropic. Its Anthropic Privacy Centre confirms that a policy update, effective 8 July 2026, now means Claude users on consumer plans must actively opt out if they do not want their inputs and outputs used for model training.
The updated policy states: ‘We may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings.’ Even users who opt out may still have conversations used for model improvement if those conversations are flagged for safety review or explicitly reported via feedback.
On data retention, Anthropic’s consumer terms update draws a direct line between training consent and how long data is kept. Users who allow their data to be used for training face a retention period of five years on new or resumed chats and coding sessions. Users who opt out continue under a 30-day retention period. Conversations that are deleted will not be used for future model training.
The scope of these changes is also worth noting. As Cyberdelegate reports, Anthropic’s July 2026 policy update applies only to consumer accounts on Free, Pro, and Max plans. Enterprise, Team, API, Amazon Bedrock, and Google Cloud Vertex AI customers operate under separate commercial terms with stronger privacy protections and are not subject to these consumer policy changes.
Anthropic states it does not sell user data and that Claude remains ad-free.
No Platform Offers Full Privacy, But Some Are More Transparent
Incogni is clear that its rankings are not an endorsement. No platform currently allows users to retrieve data once it has been used to train a model, and no retroactive withdrawal of consent is possible across any of the 13 assessed.
The rankings are better read as a map of transparency and data-sharing practices rather than a binary safe-or-unsafe verdict. ChatGPT, for example, performed well despite belonging to one of the industry’s largest players, demonstrating that scale does not inevitably mean more invasive data practices.
The advice from Incogni’s information security manager, Miguel Fornés, is practical: ‘Treat anything you enter into an LLM as information you are giving to a third-party service. Avoid sharing passwords, financial details, confidential work information, medical records, or other identifying information, and remove names or other identifiers where possible. For more sensitive tasks, consider using the web version rather than a mobile app, pay attention to the provider’s privacy practices, and, where appropriate, use a locally run model that can keep data on-device.’
Disclosure: Ziff Davis, ZDNET’s parent company, filed a copyright infringement lawsuit against OpenAI in April 2025, as Reuters reported. That litigation is ongoing.
For anyone deciding which platform to use, Claude’s opt-out deadline and Copilot’s data-broker purchases represent the two most concrete near-term triggers to act on before signing in again.