Data protection by vendor
This page compares what the main AI vendors do with your data, plan by plan, with every answer linked to the vendor's own page as checked on 2026-10-28. For what the questions mean and what to do about them, start with Data protection when you use AI.
This is not legal advice. It summarizes public vendor pages to help you ask the right questions. Policies change often, your contract may say something different, and laws such as the GDPR put obligations on you as the app maker. For a real decision, read the current terms and ask a lawyer or your data protection officer.
Comparison table
Short answers; the linked page has the exact wording. "Not stated" means we did not find it on the vendor's pages on 2026-10-28, not that it does not exist. Pages marked † refused automated access on that day; their cells rest on the page text as shown in search results, and need a manual check.
| Vendor and plan | Trains on your data by default? | Opt-out | Default retention | Zero data retention | Location / EU option |
|---|---|---|---|---|---|
| Anthropic: Claude Free, Pro, Max | Only if you allow it; you choose at signup or in settings (news) | Privacy Settings, any time (news); Incognito chats never used (help) | 30 days if training is off; 5 years if on (news); deleted chats removed from back-end within 30 days (help) | No (help) | Not stated |
| Anthropic: API, Claude for Work | No (help) | Not needed; feedback you submit may be used (same page) | API inputs and outputs deleted within 30 days (help) | Some API and Claude Code for Enterprise customers, on approval (help) | Inference "us" or "global" (default); storage US only; no EU option (docs) |
| OpenAI: ChatGPT Free, Plus, Pro | Yes (help †) | "Improve the model for everyone" in Data controls (same page †) | Deleted chats scheduled for deletion within 30 days (help †); Temporary Chat copies up to 30 days (help †) | Not stated for consumer plans | Not stated |
| OpenAI: ChatGPT Business, Enterprise, Edu | No, unless the organization agrees otherwise (help †) | Not needed | Not stated (page †) | Not stated | Not stated |
| OpenAI: API | No, unless you opt in (docs) | Not needed | Abuse-monitoring logs up to 30 days; Responses API state 30 days by default (same page) | On approval (same page) | Regions incl. Europe (EEA + Switzerland); non-US regions need approval and a contract amendment (same page) |
| Google: Gemini app | Yes (Privacy Hub) | Turn off Keep Activity (same page) | Auto-delete after 18 months by default (3, 36 months or never); 72 hours with Keep Activity off; reviewed chats up to 3 years (same page) | Not applicable | Provided by Google Ireland in the EEA and Switzerland; storage location not stated (same page) |
| Google: Gemini API, unpaid | Yes; human reviewers may read inputs and outputs (terms) | Use the paid tier (same page) | Abuse monitoring: 55 days (policy) | No | In the EEA, UK and Switzerland, paid terms apply to all use (terms) |
| Google: Gemini API, paid | No (terms) | Not needed | Abuse monitoring: 55 days (policy); logs on by default for projects created since 2026-10-26, kept 55 days by default (logs) | You configure it by avoiding storing features; some features always store (ZDR) | Not stated |
| Google Cloud: Vertex AI | No, without your permission (docs) | Not needed | Cache deleted within 24 hours; prompts may be logged for abuse detection (same page) | You configure it (disable logging and caching) (same page) | Choose a region; eu endpoint keeps data in EU member
states (residency) |
| Microsoft: Azure OpenAI in Foundry | No, without your permission; not available to OpenAI (docs) | Not needed | Flagged content may be stored for abuse review; period not stated (same page) | Modified abuse monitoring, by application (docs) | Your chosen geography, unless Global or DataZone deployment; EU DataZone available (docs) |
| Mistral: Vibe (consumer chat app) | Yes (help) | Toggle in settings or admin panel (help) | Until you delete the conversation or account (privacy policy) | No (help) | EU by default (help) |
| Mistral: Enterprise | No, opted out by default (help) | Not needed | Not stated | No; ZDR is for pay-as-you-go API calls only (help) | EU by default (help) |
| Mistral: API | Free mode: may be used; pay-as-you-go: you can opt out, default not stated (help) | Admin panel toggle (help) | 30 rolling days for abuse monitoring, except specific APIs (privacy policy) | Pay-as-you-go, stateless calls only, at Mistral's discretion (help) | EU by default; US endpoint optional (help) |
| Meta: Meta AI app | Yes (help) | In the EU, an objection form (Meta newsroom); elsewhere not stated | Not stated | Not applicable | Not stated |
| DeepSeek: app | Yes, after de-identification (terms) | "Improve the model for everyone" setting (same page) | As long as you have an account (privacy policy) | Not stated | Stored in the People's Republic of China (same page) |
| DeepSeek: API | Not stated (platform terms) | Not stated | Not stated | Not stated | Not stated |
| Any open-weight model you run yourself | Only what you do | Not applicable | What you configure | Yes, if you store nothing | Wherever you run it |
All pages checked on 2026-10-28. † = page refused automated access; manual check needed.
Details and exact wording, by vendor
Anthropic
- Consumer training is a choice. The update gives "users the choice to allow their data to be used to improve Claude"; "You're always in control of this setting"; new users "can select your preference in the signup process" (news). The page does not say which way the switch starts.
- Consumer retention. If you allow training, retention is extended "to five years"; otherwise the "30-day data retention period" applies (same page). For conversations flagged by safety systems: "We retain inputs and outputs for up to 2 years and trust and safety classification scores for up to 7 years" (help).
- Commercial. "By default, we will not use your inputs or outputs from our commercial products" to train (help). For the API, "we automatically delete inputs and outputs on our backend within 30 days of receipt or generation", with exceptions such as ZDR agreements and the Files API (help).
- Location. "Inference geo: Only
"us"and"global"are available", and "Workspace geo: Only"us"is currently available" (data residency). Through Amazon Bedrock or Google Cloud, those platforms' own regional options apply instead.
OpenAI
- API. "data sent to the OpenAI API is not used to train or improve OpenAI models", unless "you explicitly opt in"; "abuse monitoring logs are generated for all API feature usage and retained for up to 30 days"; ZDR and modified abuse monitoring are "subject to prior approval by OpenAI" (Your data). Data residency regions include "Europe (EEA + Switzerland)"; some regions store data at rest only, without regional processing (same page).
- ChatGPT consumer and business. OpenAI's help center refused automated access on 2026-10-28. As indexed by search, it says: "When you use our services for individuals (Free, Plus, Pro), we may use your content to train our models", with an opt-out in Data controls (help †). Check these pages yourself in a browser.
- Court-ordered retention. OpenAI has published statements about a US court case that required it to keep some ChatGPT and API data longer than usual. We could not open those pages on 2026-10-28, so we do not summarize them here. If this matters for you, read OpenAI's statement † directly.
- Gemini app. "Google uses your activity to provide, develop, and improve its services (including training generative AI models)"; "To stop your future chats from being reviewed to improve Google services, turn off your Keep Activity setting"; "Temporary chats and chats you have when Keep Activity is off are retained with your account for 72 hours"; "Reviewed chats are retained for up to three years" (Gemini Apps Privacy Hub). Work and school accounts "may be subject to different data handling terms" (same page).
- Gemini API. On unpaid services, Google "uses the content you submit … to provide, improve, and develop" its products and "human reviewers may read, annotate, and process your API input and output"; on paid services, "Google doesn't use your prompts … or responses to improve our products" (Gemini API terms). For abuse monitoring, "Google retains the following data for fifty-five (55) days" (usage policies). "To achieve zero data retention, customers must take specific actions or avoid specific features", and grounding with Google Search stores data for 30 days (ZDR).
- Vertex AI (Google Cloud's AI platform, now
presented as Gemini Enterprise Agent Platform). "Google won't use your
data to train or fine-tune any AI/ML models without your prior
permission or instruction" (data
governance). The
euendpoint "strictly covers data residency within EU member states"; the global endpoint "may be processed in any Google Cloud location around the world" (data residency).
Microsoft (Azure OpenAI in Microsoft Foundry)
- Prompts and completions "are NOT available to OpenAI" and "are NOT used to train any generative AI foundation models without your permission or instruction"; data is "processed within the customer-specified geography (unless you are using a Global or DataZone deployment type)" (data privacy).
- Customers "who meet additional Limited Access eligibility criteria" can "apply to modify abuse monitoring" (abuse monitoring).
Mistral
- Mistral's help pages now call its consumer chat app Vibe. "Input and output data are used by default to train our artificial intelligence models" on Vibe, while "Enterprise customers are opted out of training by default" (help). Opt-out toggles for Vibe and the API are separate (help).
- API retention: "we keep your Input and Output for the period necessary to generate the Output and then for thirty (30) rolling days to monitor abuse (unless zero data retention is activated)" (privacy policy). ZDR is "available only with pay-as-you-go and only for stateless API calls", and Mistral may "at our discretion, approve or deny it" (help).
- "By default, your data is hosted in the European Union" (help).
Meta
- "Meta uses your interactions with AIs to improve AI at Meta" (Meta help). In the EU, interactions with Meta AI "will also be used to train and improve our models", with "a link to a form where people can object" (Meta newsroom).
- Since December 2025, Meta says it will "use your interactions with AI at Meta to personalize the content and ads you see" (Meta newsroom).
- Meta's generative AI privacy pages on facebook.com could not be loaded on 2026-10-28, so retention and location are not covered here.
- Llama models are open-weight: if you run them yourself, Meta receives nothing.
DeepSeek
- The app may "use Inputs and Outputs to provide, maintain, operate, develop or improve the Services", subject to de-identification, and you "can opt out by turning off "Improve the model for everyone"" (terms of use).
- "we directly collect, process and store your Personal Data in People's Republic of China"; data is kept "for as long as you have an account" (privacy policy).
- The API terms we found do not address training, retention or location (open platform terms). DeepSeek's models are also open-weight, and other providers serve them under their own terms.
Related pages
See also
Glossary terms used on this page: Agent, API, Cloud, Generative AI, Inference, Logs, Model, Open-weight model, Prompt, Provider, Zero data retention (ZDR).
Sources
All vendor pages are linked in the table and in the details above, and were checked on 2026-10-28. Pages marked † refused automated access on that date. Corrections with a link to the current vendor page are especially welcome.
First version written with AI assistance and checked against the vendor pages listed under sources on 2026-10-28. Not legal advice. Found a mistake or a policy change? Fix it or open an issue. Text licensed under CC BY 4.0.