Every prompt sent to a large language model is stored somewhere for some period, and whether it can be used to train future models depends on the provider, the product and the plan, not on the model. This page sets out, side by side, what six providers publish about training use and retention for their consumer apps and their business and API services: the training default, the opt-out control, how long content is kept, and whether zero data retention (ZDR) can be arranged. It is written for privacy officers, security teams, procurement and anyone deciding which AI tool staff may use with which data. Providers are listed alphabetically; order is never a ranking. Every cell is taken from the provider's own help centre, policy or documentation page. Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Key takeaways
- Consumer and business terms differ at every provider checked. Anthropic, Google, OpenAI and xAI all state that business, enterprise or paid API content is not used for training by default, while their consumer apps use separate settings.
- The opt-out is a named setting, and it is per account. "Model Improvement" (Anthropic), "Keep Activity" (Google Gemini Apps), "Improve the model for everyone" (OpenAI) and "Improve the model" (xAI) each sit in the user's own settings, so an organisation cannot rely on staff having changed them.
- 30 days is the most common published API retention period. Anthropic, OpenAI and xAI each publish a deletion or abuse monitoring window of 30 days for API content, with exceptions for flagged content and legal holds.
- Zero data retention is by approval, not by default. Anthropic, Mistral AI and OpenAI describe ZDR as an arrangement each customer must request and have approved; xAI lets a team admin enable it in its Console and states it does not recommend ZDR for most customers; Google describes the configuration steps for ZDR on its cloud platform and lists features that retain data regardless.
- Feedback buttons override the opt-out. Anthropic, Mistral AI and xAI state that thumbs up or down ratings and bug reports may be used for training even where training is otherwise off.
Why the product and plan decide data handling, not the model
The same model can be reached through a consumer chat app, a business workspace plan, a pay-as-you-go API and one or more cloud marketplaces. Each route has its own terms. A prompt typed into a free consumer app may enter a training pipeline; the same prompt sent through the paid API of the same provider usually does not, and may be deleted within 30 days. When an assessment asks "is this model safe for confidential data", the useful question is which product, which plan and which settings.
- Consumer apps: free and individual paid plans governed by consumer terms and a privacy policy, with user-level settings.
- Business and enterprise plans: workspace products governed by commercial terms and usually a data processing agreement, with admin-level settings.
- API services: developer access billed per token, governed by commercial or API terms. Some providers split API terms into unpaid and paid tiers (Google) or pay-as-you-go and free modes (Mistral AI).
- Cloud platforms: models hosted inside a hyperscaler, where the hyperscaler's own data terms apply. This page covers Google Cloud's documentation as one example; hosting through other clouds follows those clouds' terms.
For what ZDR means in principle, see the Zero Data Retention for AI. For retention and deletion controls inside your own AI systems, see AI Data Retention and Deletion.
Consumer apps: training default and opt-out control
As of 2026-11-15. "Not stated" means the page checked did not state the default; it does not mean the default is on or off.
| Provider | Consumer training use | Opt-out control | Source |
|---|---|---|---|
| Anthropic (Claude Free, Pro, Max, including Claude Code on those plans) | Chats and coding sessions are used to improve models if the user allows it, if a conversation is flagged for safety review, or if the user opts in (for example the Trusted Tester Program). Default state not stated on the page checked. | "Model Improvement" in Privacy Settings. Incognito chats are not used, even with Model Improvement on. | Anthropic Privacy Center |
| DeepSeek (DeepSeek app and web) | The privacy policy (updated 10 February 2026) states personal data, including inputs, may be used to train and improve its models. | The policy lists a right to opt out of use for training, exercised by contacting [email protected]. An in-app setting is not stated in the policy. | DeepSeek Privacy Policy |
| Google (Gemini Apps) | With Keep Activity on, activity is used to improve services, including training generative AI models. Default state not stated on the page checked. | "Keep Activity" setting. With it off, future chats are not used for training unless the user sends feedback. | Gemini Apps Privacy Hub |
| Mistral AI (Vibe individual plan) | Input and output are used by default for training. | Opt out at any time. Thumbs up or down ratings authorise use of the rated input and output on every plan. | Mistral AI Help Center |
| OpenAI (ChatGPT individual plans) | ChatGPT trains on conversations unless the user opts out. | "Improve the model for everyone" in Settings > Data Controls, or "do not train on my content" in the privacy portal. Temporary Chats are not used for training. | OpenAI Help Center; Data Controls FAQ |
| xAI (Grok apps and Grok.com) | Default state not stated on the page checked. Private Chat content is not used for training. | "Improve the model" under Settings > Data Controls (mobile) or Settings > Data (web). Voluntary feedback may still be used after opting out. | xAI Legal FAQ |
Consumer apps: how long content is kept
As of 2026-11-15.
| Provider | Deleted or temporary chats | Longer retention cases | Source |
|---|---|---|---|
| Anthropic | Deleted chats leave history immediately and are deleted from back-end storage within 30 days. | With Model Improvement on, de-identified chats may be kept up to 5 years in training pipelines. Content flagged for Usage Policy violations up to 2 years; classification scores up to 7 years; feedback 5 years. | Anthropic Privacy Center |
| DeepSeek | Data kept as long as necessary to provide the services; account data and inputs kept for as long as the user has an account. Chat history can be deleted in settings. | Data may be kept longer where terms are violated. The policy states data is collected, processed and stored in the People's Republic of China. | DeepSeek Privacy Policy |
| With Keep Activity off, future chats are still saved for 72 hours. | With Keep Activity on, activity auto-deletes after 18 months by default (changeable to 3 or 36 months, or off). Chats reviewed by human reviewers are retained up to 3 years. | Gemini Apps Privacy Hub | |
| Mistral AI | Not stated by Mistral AI on the pages checked (2026-11-15). | Not stated by Mistral AI on the pages checked (2026-11-15). | Mistral AI Help Center |
| OpenAI | Temporary Chats are removed from systems after 30 days. | A removal period for deleted standard chats is not stated on the Data Controls FAQ page checked. | OpenAI Data Controls FAQ |
| xAI | Deleted conversations and Private Chat content are removed within 30 days. | Unless de-identified and disassociated from the account, or kept for safety, security or legal reasons. | xAI Legal FAQ |
Business plans and APIs: training default and retention
As of 2026-11-15.
| Provider | Business and API training default | API retention | Source |
|---|---|---|---|
| Anthropic | Inputs and outputs from commercial products (Claude for Work, the Anthropic API, Claude Gov) are not used for training by default. Exceptions: feedback and bug reports, and explicit opt-in. | API inputs and outputs deleted within 30 days of receipt or generation, unless an exception applies (for example a ZDR agreement, legal requirement or flagged content). | Commercial training; Commercial retention |
| DeepSeek | Not stated by DeepSeek on the pages checked (2026-11-15). | Not stated by DeepSeek on the pages checked (2026-11-15). | DeepSeek Privacy Policy |
| Gemini API Paid Services: prompts and responses not used to improve products. Unpaid Services: used to improve products and machine learning, with possible human review. Google Cloud: no training or fine-tuning on customer data without prior permission or instruction. | Gemini API Paid Services: logged for a limited period solely for abuse detection (period not stated). Google Cloud: implicit cache data in memory and durable storage deleted within 24 hours of writing. | Gemini API Additional Terms; Google Cloud ZDR page | |
| Mistral AI | Vibe Enterprise customers are opted out of training by default, with an admin opt-in toggle. Mistral Studio pay-as-you-go customers can opt out at any time. | A default API retention period is not stated on the pages checked. | Training article; ZDR article |
| OpenAI | Since 1 March 2023, data sent to the API is not used to train models unless the customer opts in. ChatGPT Team and Enterprise content is not used for training. | Abuse monitoring logs kept up to 30 days, unless law requires longer or retention is needed to protect services or third parties. | API data controls |
| xAI | Business data, including prompts and outputs, is not used for training. Free credits may be offered in exchange for permission to train. | Inputs and outputs automatically deleted within 30 days, unless agreed otherwise in writing or required by law (for example flagged content). | xAI Enterprise FAQ |
Is zero data retention available?
As of 2026-11-15. ZDR terms are customer specific: the published pages describe eligibility and scope, not contract wording.
| Provider | ZDR availability and route | Scope and exclusions stated | Source |
|---|---|---|---|
| Anthropic | Available to some Claude API and Claude Code for Enterprise customers, subject to approval, applied per organisation. Status visible under Settings > Privacy Controls > Data retention period. | Covers eligible APIs, products using a commercial organisation API key (including Claude Code via the API) and Claude Code for Enterprise. Excludes consumer plans. User Safety classifier results are still retained; Covered Models have separate limited retention and review for safety work. | Anthropic Privacy Center |
| DeepSeek | Not stated by DeepSeek on the pages checked (2026-11-15). | Not stated by DeepSeek on the pages checked (2026-11-15). | DeepSeek Privacy Policy |
| Google (Google Cloud) | Configured by the customer: keep request-response logging disabled (the default), set store to false in Interactions API requests, turn off caching through cacheConfig, leave Live API session resumption disabled, avoid the Sandbox Snapshot API. Under Section 4.3 of the Google Cloud Platform Terms of Service, Google may log prompts to detect abuse. | Retained regardless: Grounding with Google Search query logs up to 3 days, Deep Research sessions 7 days, Grounding with Google Maps 30 days. Some advanced models or features may not support ZDR. | Google Cloud ZDR page |
| Mistral AI | Pay-as-you-go customers, stateless API calls only, by request through the Help Center or support; approved or denied at Mistral AI's discretion. Approval shows in Admin panel Privacy settings. | Applies to chat completions, FIM completions, embeddings, moderation and classification, OCR and audio endpoints. Excludes Vibe Work, Chat, agents, conversations, libraries, batch files and /v1/files. | Mistral AI Help Center |
| OpenAI | ZDR or Modified Abuse Monitoring by prior approval; contact sales for eligibility. Approved customers choose one per organisation or project. | Under ZDR the store parameter is treated as false for /v1/responses and /v1/chat/completions; some endpoints can still retain application state. | API data controls |
| xAI | API: team admins can enable ZDR team-wide in the Console after deleting existing Files and Collections; xAI states it does not recommend ZDR for most customers. Enterprise FAQ: the default 30-day deletion can be changed by written agreement. | ZDR disables features that depend on stored data, including the stateful Responses API, Files and Collections, the Batch API and deferred completions. | xAI API Security FAQ; xAI Enterprise FAQ |
How to read these policies: factual criteria
These are criteria for matching a product to a data class. They are not a ranking of providers.
- Which terms apply: confirm whether the account is on consumer terms or commercial terms. Paid individual plans can still be consumer plans: Anthropic lists Claude Pro and Max with Free under its consumer terms.
- Who controls the setting: consumer opt-outs are user-level and can be changed back; enterprise defaults are admin-level. If a control is user-level, your policy has to cover it.
- What overrides the setting: feedback ratings, bug reports, safety flags and legal holds are stated exceptions. Decide whether staff may use feedback buttons with sensitive content.
- Where retention restarts: features such as web grounding, research agents, file stores and stateful conversation APIs carry their own retention, as the Google and Mistral AI exclusion lists show. Map every feature you enable, not only the model endpoint.
- Where data is stored: DeepSeek's policy states storage in the People's Republic of China. Residency options for other providers are documented separately (OpenAI points to a data residency guide); check them before transfer assessments.
- Whether the page is the contract: help centre pages summarise; the commercial terms and data processing agreement govern. Record the page and date you relied on.
Turning the comparison into controls
- Inventory routes, not models. List every AI product in use by plan and access route (consumer app, business plan, API key, cloud project). The AI Agent Products Release Tracker and LLM Pricing Tracker help name them.
- Classify each route against your data classes. Consumer routes with training on by default should not receive confidential data. Record the source page and date beside each decision.
- Move work use onto commercial terms. Steer staff to business plans or API access where training is off by default, and block or coach on personal accounts. See Blocking Personal ChatGPT, Claude and Gemini Accounts at Work.
- Set admin controls. Turn off feedback submission where offered (for example Anthropic's "Rate chats" organisation setting) if feedback content could be sensitive.
- Request ZDR only where the use case needs it, and test it: confirm the status page or admin panel shows it, and list the features excluded from it.
- Add DLP on prompts. Retention terms limit what the provider keeps; they do not stop sensitive data leaving. See the Generative AI Data Loss Prevention Guide.
- Review on change. Re-check the source pages when a provider changes terms or launches a product, and log the date.
Measuring whether the controls hold
- Share of AI traffic on commercial terms: proportion of AI requests or users going through sanctioned business plans or API keys rather than consumer accounts.
- Routes with a recorded policy check: percentage of inventoried AI routes with a dated source page for training and retention.
- ZDR verification: number of ZDR arrangements with evidence from the provider's status page or admin panel, against the number claimed.
- Excluded feature use: count of projects using features listed as retaining data regardless of ZDR (grounding, research agents, file stores).
- DLP events on AI destinations: blocked or coached prompts containing classified data, trended monthly.
Common failure patterns
- Assuming a paid personal plan equals business terms. Anthropic, for example, lists Claude Pro and Max alongside Free as consumer plans, so a personal subscription does not bring commercial training defaults.
- Treating ZDR as account-wide. Published exclusions (batch files, file stores, agents, grounding) mean data can still be retained for parts of a workload.
- Forgetting feedback buttons. A thumbs up on a sensitive conversation can send it into a training set even with training off.
- Relying on a user's opt-out. Settings are per user and per device or account, and no organisation can see them for personal accounts.
- Citing a help page as the contract. Help pages change; keep the dated copy and the governing agreement together.
Practitioner checklist
- List every AI product in use by provider, plan and access route.
- For each route, record training default, opt-out owner, retention period and ZDR status, with the source URL and date checked.
- Prohibit confidential data on consumer routes where training is on by default or the default is not stated.
- Turn off feedback submission at organisation level where the product offers it and the data warrants it.
- Confirm any ZDR arrangement on the provider's status page or admin panel and list excluded features.
- Check data storage location statements before cross-border transfer assessments.
- Pair retention controls with prompt DLP and logging of AI destinations.
- Re-check the source pages when providers announce new products or terms.
Standards and references
- Anthropic: Zero data retention products
- DeepSeek: Privacy Policy
- Google Cloud: Zero data retention
- Google: Gemini API Additional Terms of Service
- Mistral AI: Can I activate Zero Data Retention (ZDR)?
- OpenAI: Data controls in the OpenAI platform
- xAI: Enterprise FAQ
- xAI: API Security FAQ (accessed 2026-11-15)
Related NHI Mgmt Group resources: Zero Data Retention for AI · AI Data Retention and Deletion · Anthropic Models · DeepSeek Models · Google Gemini Models · Mistral AI Models · OpenAI Models · xAI Models