ZDNET、13 AI プラットフォームのプライバシーリスクを調査し順位付け

*Follow ZDNET: *Add us as a preferred source* on Google.*

ZDNET's key takeaways

• Incogni's new study of 13 AI platforms ranks them by privacy risk.
• The largest AI platforms turn out to be the most privacy-invasive - with one exception.
• Here's what to know about each option before signing up for an account.

The use of generative AI platforms such as ChatGPT, Claude, and Gemini continues to change how we live and work. But do the privacy risks outweigh the convenience?

AI engines have had a cataclysmic impact on many of our lives, comparable to that of smartphones. Businesses are pouring billions of dollars into AI-related developments; it seems like almost every organization wants to project an "AI-first" image, and even in our daily queries and quests for information, AI citations are replacing the older methods we used to search.

Also: How to keep your conversations with ChatGPT, Gemini, Copilot, or Claude private

We're asking AI for advice before making real-world decisions, and some of us are even seeking out recommendations on our personal lives and finances.

*(Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)*

But how safe is the data we feed into our AI companions, and do the organizations developing today's AI platforms consider user privacy to be their responsibility? If you've ever asked these questions, new research from Incogni sheds some light on them.

Incogni's Gen AI and LLM Data Privacy Ranking 2026 report

Published on Thursday, data broker removal service Incogni's new research analyzes 13 AI platforms and the potential risks they pose to consumer privacy.

Also: AI cybersecurity attacks are the future: 43% of companies have already experienced it

The "Gen AI and LLM Data Privacy Ranking 2026" report looks at these AI platforms: ChatGPT, Claude, Gemini, Grok, Vibe (formerly Le Chat), Perplexity, Qwen, DeepSeek, Z.ai, Kimi, Meta AI, Pi, and Copilot. In the study, each AI has been assigned a score "based on the privacy risks they pose," Incogni says. The lower the score, the better, as each point added reflects the discovery of privacy-invasive practices, how easy data-sharing practices are to find and understand, and what happens to user data.

Three groups of questions are addressed by the study:

• What happens to user data? Are user conversations fed into training platforms, can users opt out, and are prompts shared with third parties or other systems?
• Are AI platforms transparent? Are data and privacy policies easy to find and understand?
• What happens to data on AI platforms? What personal data is gathered, where does it come from, and if it is shared, where does it go?

The larger the platform, the bigger the risk - with one exception

Some of the largest AI platforms appear to pose the greatest risk to user privacy, except for OpenAI's ChatGPT. Gemini and Meta AI received some of the highest risk scores, whereas Vibe, ChatGPT, and Pi achieved some of the best (and lowest) risk scores in Incogni's assessment.

*Incogni*

In summary, this is how each AI model performed overall, ranked from lowest to highest risk:

• Mistral AI's Vibe (fka, Le Chat): Easy-to-understand privacy policies, privacy-friendly mobile apps, information collected from public sources, and user conversations are only shared with a "minimal" number of third parties.
• OpenAI's ChatGPT: A clear, simple privacy policy, FAQs, and resources; opting out of data collection for model training is straightforward, and it is clearly stated that data is collected and may be shared for security or marketing purposes.
• Inflection AI's Pi: A complex privacy policy that handles EU users separately (for GDPR purposes), potentially shares user data with corporate, commercial, and research partners, but has a privacy-friendly app.
• Perplexity AI: An easy-to-follow but not very detailed privacy policy; opting out of model training is easy to find and enable, but Perplexity user data may be shared with corporate group members and affiliates.
• Alibaba's Qwen: Privacy policies describe what data is used and for what purposes, such as sharing personal data with analytics providers, search engine providers, and other third-party services, but it is less clear exactly what user information is shared and with whom. Incogni says it is also difficult to determine whether user conversations are used for training, as this isn't covered beyond a statement that user inputs are stored for "model improvement and other purposes."
• DeepSeek: DeepSeek allows user data to stay on local networks via open-weight models, but the researchers say privacy considerations become "more serious" on hosted versions of its models. DeepSeek allows users to reach out if they find inaccurate personal information and want it corrected or removed. DeepSeek provides detailed information on how its AI models are trained.
• Z.ai: Z.ai also offers open-weight models, but again, hosted versions have privacy implications. Z'ai's privacy policy is reasonable but not very detailed, and there is ambiguity surrounding AI model training.
• Google's Gemini: Google can already collect vast amounts of data from its other services, and Gemini has what the team calls a "partial" privacy policy that defaults to its general policy. As a result, "getting answers to specific questions can be a real challenge."
• Anthropic's Claude: Anthropic has an extensive privacy policy and data-related resources, and claims not to use user data for AI training by default, although its latest policy (July 2026) suggests users now have to opt out. The organization may publish research using aggregated, de-identified user data.
• xAI's Grok: Corporate group members and affiliated parties can receive user data, and its model is trained not only on Grok data but also on public X conversations.
• Moonshot AI's Kimi: Incogni said Chinese-owned Kimi "represented the greatest privacy risk among the platforms assessed" with regard to user data handling, and its app uses data to track users, according to the report. It's also difficult to opt out of AI model training.
• Meta: Considered the most "data-hungry" model in the sample, Meta's claims that training only takes place in "private conversations but not group chats" are difficult to verify due to a broad privacy policy.
• Microsoft's Copilot: The app shares data with third-party advertisers and has an overly broad privacy policy, creating transparency issues. Incogni also says that data is purchased from data brokers.

Which AI is the best for data privacy?

Incogni's results and rankings do not mean that Vibe, ChatGPT, Meta, or Copilot are automatically the best or worst options for you. For example, no platform allows users to retrieve their data once it has been used to train a large language model, and no retroactive removal or withdrawal of consent is possible.

Instead, you should look at these privacy-weighted scores as indicators of what some AI models are doing well, and what others need to improve. New security and privacy concerns are emerging week by week -- the accidental indexing of Claude chats being only one of many recent incidents -- and your privacy is not guaranteed with any AI model on the market today.

Also: Google's AI can see your business data by default in Workspace - unless you disable it

When you use any large language model, it's best to choose one with transparent data collection and handling policies, and be careful about what information you share with it. If the idea of your financial statements being used to train an AI isn't palatable, for example, then don't submit them -- and apply this concept to all of your future conversations.

"Treat anything you enter into an LLM as information you are giving to a third-party service," Incogni's information security manager Miguel Fornés told ZDNET. "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."