Privacy and training data

How long each provider says it keeps what you type, and whether it may use your chats to train future models.

Please note: This site is only a guide to help people understand. Consult your IT, security and/or privacy department for clarification.

No provider is ranked or recommended here. Providers appear in the same order every time. Why these tools? Facts were last reviewed on 2026-10-07. If we could not find a fact on a provider's own pages, the table says “Not found in reviewed sources.”

Privacy: Frontier model providers
Claude	ChatGPT	Gemini	Grok	Llama
Free, Pro and Max: deleted chats are removed from Anthropic's systems within 30 days. If you allow your chats to be used for training, they may be kept (with your identity removed) for up to 5 years. Source for Claude, PrivacyApplies to: Claude (Free, Pro and Max)Checked 2026-10-07 Team, Enterprise and API: deleted chats are removed from Anthropic's systems within 30 days, and API inputs and outputs are deleted within 30 days unless a different agreement applies. Source for Claude, PrivacyApplies to: Claude (Team, Enterprise and API)Checked 2026-10-07	Personal plans: saved chats stay until you delete them. Deleted chats are scheduled for permanent deletion within 30 days. Temporary chats may be kept up to 30 days for safety. Source for ChatGPT, PrivacyChecked 2026-10-07 Business and Enterprise: admins control how long chats are kept, and deleted chats are removed within 30 days. API inputs and outputs may be kept up to 30 days to check for abuse. Source for ChatGPT, PrivacyApplies to: Business, Enterprise and APIChecked 2026-10-07	Personal Google accounts: activity is auto-deleted after 18 months by default (you can pick 3 or 36 months, or off). Some chats are read by human reviewers and kept up to 3 years, separated from your account. Source for Gemini, PrivacyChecked 2026-10-07 Workspace accounts: chats stay within your organization and are not reviewed by people outside it without permission. Admins choose how long Gemini app chats are kept (up to 36 months). Source for Gemini, PrivacyApplies to: Google WorkspaceChecked 2026-10-07	Personal accounts: deleted chats, and chats in Private Chat, are removed within 30 days. A limited number of xAI staff may review chats, including to improve the model. Source for Grok, PrivacyChecked 2026-10-07 Business and API customers: inputs and outputs are deleted automatically within 30 days unless agreed otherwise in writing or required by law. Source for Grok, PrivacyApplies to: Business and API customersChecked 2026-10-07 x.ai is run by SpaceXAI. Some older claims below say "xAI", the name the same pages used when we first read them.	From the model developer (Meta) Meta's own assistant, Meta AI, now runs on Meta's "Muse Spark" model, not Llama (announced April 2026), so Meta AI's terms do not describe Llama. Source for Llama, PrivacyChecked 2026-10-07 From the host: Amazon Web Services (Amazon Bedrock) Example host, Amazon Bedrock: under the default setting AWS may keep prompts to check for abuse, but the model maker does not receive them. Zero data retention (nothing stored) can be set through the AWS API, not the console. Some models require retention and are blocked when it is off. Source for Llama, PrivacyApplies to: Amazon BedrockChecked 2026-10-07 Open-weight: the developer's license and the host's terms are separate. Whoever hosts the model is responsible for its security.

Privacy: Other general AI assistants
Microsoft Copilot	Perplexity	Mistral Le Chat
Prompts, retrieved data and responses are processed and stored under the same contract terms as the organization's other Microsoft 365 content, and stored data is encrypted. Users can delete their Copilot activity history. EU users' traffic stays within the EU Data Boundary, except that models from Anthropic are currently excluded. Source for Microsoft Copilot, PrivacyApplies to: Microsoft 365 Copilot (now named Microsoft Copilot), paid work licenseChecked 2026-10-08 Copilot Chat gets "enterprise data protection" at no extra cost: prompts and responses stay in the Microsoft 365 service boundary and are logged. Web search queries sent to Bing are handled separately under the Microsoft Services Agreement and are not EU Data Boundary compliant. Source for Microsoft Copilot, PrivacyApplies to: Microsoft 365 Copilot Chat (now named Microsoft Copilot Chat)Checked 2026-10-08 Older version of the consumer app, before the 2026-08-18 update: conversations are saved by default and kept 18 months, uploaded files are kept no longer than 18 months, and some conversations get automated and human review. There is no opt-out of human review. Source for Microsoft Copilot, PrivacyApplies to: Microsoft Copilot app, older version (before 2026-08-18)Checked 2026-10-08 Privacy: Not found in reviewed sources: How long chats are kept in the new consumer Copilot app Source for Microsoft Copilot, PrivacyApplies to: Microsoft Copilot app, new version (from 2026-08-18)We checked the overview, privacy controls and activity history pages. They describe deleting chats and memory, not a retention period.Checked 2026-10-08 Users without a Microsoft 365 subscription may see ads in the new consumer app. Ad personalization can be turned off in settings. Source for Microsoft Copilot, PrivacyApplies to: Microsoft Copilot app, new version (from 2026-08-18) (No Microsoft 365 subscription)Checked 2026-10-08	Consumer accounts: personal information is kept while the account is active and removed within 30 days of account deletion. Source for Perplexity, PrivacyApplies to: PerplexityChecked 2026-10-08 Enterprise: files attached to sessions are deleted after 7 days, and organizations with 50 or more seats or one Enterprise Max seat can set custom retention. Perplexity describes zero data retention and zero training agreements with AI providers (OpenAI, Anthropic and others). Source for Perplexity, PrivacyApplies to: Perplexity Enterprise (Enterprise Pro, Enterprise Max)Checked 2026-10-08 API Platform: "No data logging or storage". Enterprise Pro and Max: "never logged or used for training". Source for Perplexity, PrivacyApplies to: Perplexity API Platform; Perplexity Enterprise (API Platform; Enterprise Pro and Max)Checked 2026-10-08	Conversations are kept by default so you can see your history. You can delete a conversation at any time, but Mistral may keep data in back-end systems for different periods, for example to enforce its Usage Policy or as required by law. Team and Enterprise organizations may have their own retention controls. Source for Mistral Le Chat, PrivacyApplies to: Vibe (Team and Enterprise mentioned)Checked 2026-10-08 Data is stored for as long as personal data is processed. Account identity data is kept 5 years after the contract ends, other sign-up data 1 year after account deletion, technical data for 1 rolling year, and invoices 10 years. Source for Mistral Le Chat, PrivacyApplies to: Mistral (company-wide)Checked 2026-10-08 Mistral's privacy policy (effective 2026-09-03): Vibe inputs and outputs are kept until you delete your account or delete the conversation. For the APIs (except some specific APIs), inputs and outputs are kept as long as needed to generate the output and then for 30 rolling days to monitor abuse, unless zero data retention is turned on. Source for Mistral Le Chat, PrivacyApplies to: Vibe; Mistral APIsChecked 2026-10-08

Privacy: Tools built for nonprofits
Givebutter	Grantable
Some Magic features (campaign content generation, contact summaries, smart filters, custom reports, receipt matching, check acceptance) send the data the feature needs to third-party AI providers: Anthropic, Google and OpenAI. That data can include names, email addresses, giving history and contact details. Source for Givebutter, PrivacyApplies to: Givebutter MagicChecked 2026-10-08 Givebutter logs requests to and responses from the AI providers in an AI observability tool (named as Braintrust). These logs may contain personal information. Source for Givebutter, PrivacyApplies to: Givebutter MagicChecked 2026-10-08 Inputs are generally processed to produce the requested output and are not kept by the AI provider, subject to each provider's own retention policies. Source for Givebutter, PrivacyApplies to: Givebutter MagicChecked 2026-10-08 Givebutter keeps personal information for as long as reasonably necessary for the purposes it was collected, with no fixed period stated. Source for Givebutter, PrivacyApplies to: GivebutterChecked 2026-10-08 AI features are on by default. Fundraisers can turn off some AI features one by one with in-product controls, and some features that are core to the platform cannot be turned off. AI Connectors need an explicit opt-in. Source for Givebutter, PrivacyApplies to: Givebutter Magic, AI ConnectorsChecked 2026-10-08	When you use AI, Grantable sends the AI provider relevant portions of your documents, organizational context and funder information needed for the request. Content generated through AI features remains your property. Source for Grantable, PrivacyApplies to: GrantableChecked 2026-10-08 Grantable says your full database is never transmitted to the AI provider, just the relevant portions needed for each request. Source for Grantable, PrivacyApplies to: GrantableChecked 2026-10-08 Account data is kept while the account is active and deleted when it is closed, subject to legal requirements. User content is usually kept for one year after closure unless you ask for earlier deletion. You can request deletion at hello@grantable.co. Source for Grantable, PrivacyApplies to: GrantableChecked 2026-10-08 Custom data retention is an enterprise feature. Source for Grantable, PrivacyApplies to: Grantable (Enterprise)Checked 2026-10-08 Grantable's pages name different AI providers. Its sub-processor list (last updated 2026-09-30) says Anthropic is the only AI provider that reads what you write or upload, and that OpenAI is not listed because the app sends it nothing. Its AI model tiers doc (updated 2026-03-26) names only Claude Sonnet and Claude Haiku. Its privacy policy (last updated 2025-05-16) and trust page name Anthropic and OpenAI, and its pricing-page FAQ names Anthropic, OpenAI and Google. The sub-processor list says data is stored in the United States (Supabase on AWS N. Virginia) and Anthropic requests may be processed in other regions. Source for Grantable, PrivacyApplies to: GrantableOther pages: https://grantable.co/docs/ai-automation/model-tiers, https://www.grantable.co/privacy, https://www.grantable.co/trust and https://grantable.co/pricing (FAQ "What AI models do you use?").Checked 2026-10-08

Training on your data: Frontier model providers
Claude	ChatGPT	Gemini	Grok	Llama
Free, Pro and Max: your chats are used to improve Claude only if you allow it in your settings (or send feedback, or a chat is flagged for safety review). Incognito chats are never used. Source for Claude, Training on your dataApplies to: Claude (Free, Pro and Max)Checked 2026-10-07 To opt out on Free, Pro or Max: Settings > Privacy > turn off "Help Improve our AI models". Data already used in past training is not removed. Source for Claude, Training on your dataApplies to: Claude (Free, Pro or Max)Checked 2026-10-07 Team, Enterprise and API: by default Anthropic does not train on your inputs or outputs, unless you send feedback or choose to opt in. Source for Claude, Training on your dataApplies to: Claude (Team, Enterprise and API)Checked 2026-10-07	Free, Go, Plus and Pro: OpenAI may use your chats to train its models unless you opt out. Temporary chats are not used. Source for ChatGPT, Training on your dataApplies to: Free, Go, Plus and ProChecked 2026-10-07 To opt out: Settings > Data controls > turn off "Improve the model for everyone". Feedback you send (thumbs up or down) may still be used. Source for ChatGPT, Training on your dataChecked 2026-10-07 Business, Enterprise, Edu and API: by default OpenAI does not use your inputs or outputs to improve its models, unless you opt in. Source for ChatGPT, Training on your dataApplies to: Business, Enterprise, Edu and APIChecked 2026-10-07	Personal Google accounts: when "Keep Activity" is on, your chats are used to improve Google services, including training AI models. Temporary chats are not used. Source for Gemini, Training on your dataChecked 2026-10-07 To opt out on a personal account: turn off "Keep Activity" in Gemini Apps Activity. Chats are still kept for 72 hours, and feedback you send may still be used. Source for Gemini, Training on your dataChecked 2026-10-07 Workspace accounts: Google does not use your prompts, files or responses to train AI models without your permission. Source for Gemini, Training on your dataApplies to: Google WorkspaceChecked 2026-10-07	Personal accounts: xAI may use your prompts and Grok's replies to train its models. Private Chat is not used for training. Source for Grok, Training on your dataChecked 2026-10-07 To opt out: Settings > Data (website) or Settings > Data Controls (app) > "Improve the model". Feedback you send may still be used. Source for Grok, Training on your dataChecked 2026-10-07 Business and API customers: xAI says it does not use business prompts or outputs to train its models (it may offer free credits in exchange for permission). Source for Grok, Training on your dataApplies to: Business and API customersChecked 2026-10-07 x.ai is run by SpaceXAI. Some older claims below say "xAI", the name the same pages used when we first read them.	From the model developer (Meta) Training on your data: Not found in reviewed sources: Whether Meta uses prompts sent to a Llama model that you or a vendor host to train its models Source for Llama, Training on your dataThe Llama license and Llama pages do not address it.Checked 2026-10-07 From the host: Amazon Web Services (Amazon Bedrock) Example host, Amazon Bedrock: data you use to customize a model is not shared with the model maker and is not used to improve the base model. Source for Llama, Training on your dataApplies to: Amazon BedrockChecked 2026-10-07 Open-weight: the developer's license and the host's terms are separate. Whoever hosts the model is responsible for its security.

Training on your data: Other general AI assistants
Microsoft Copilot	Perplexity	Mistral Le Chat
Prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs. Optional feedback may be used to improve the product but not to train the foundation LLMs. Source for Microsoft Copilot, Training on your dataApplies to: Microsoft 365 Copilot (now named Microsoft Copilot), paid work licenseChecked 2026-10-08 Prompts and responses are not used to train the underlying foundation models. Source for Microsoft Copilot, Training on your dataApplies to: Microsoft 365 Copilot Chat (now named Microsoft Copilot Chat)Checked 2026-10-08 Prompts, responses and file contents are not used to train foundation models. Source for Microsoft Copilot, Training on your dataApplies to: Microsoft Copilot app, new version (from 2026-08-18)Checked 2026-10-08 Microsoft uses Copilot conversation activity, including uploaded files, for AI training unless the user opts out or is in an excluded group (Entra ID work accounts, Microsoft 365 Personal or Family users in apps, signed-out users, under-18s, and users in some countries). Opting out applies to future conversations. Source for Microsoft Copilot, Training on your dataApplies to: Microsoft Copilot app, older version (before 2026-08-18)Checked 2026-10-08 Prompts, responses and file contents are not used to train foundation models. Source for Microsoft Copilot, Training on your dataApplies to: Copilot in Microsoft 365 apps for home (Microsoft 365 Personal, Family, Premium, Pro)Checked 2026-10-08	Free, Pro and Max: "AI Data Retention" is on by default, so data can be used for AI training. Turn off the "AI data retention" toggle in Account > Preferences. The opt-out covers only data collected after you opt out, and already-collected training data cannot be removed. Source for Perplexity, Training on your dataApplies to: Perplexity (Free, Pro, Max)Checked 2026-10-08 Enterprise data is never used for AI training. If an organization downgrades to a non-Enterprise plan, the consumer default (training on) applies again. Source for Perplexity, Training on your dataApplies to: Perplexity Enterprise (Enterprise Pro, Enterprise Max)Checked 2026-10-08	Input and output are used by default to train Mistral's models unless you opt out. On Teams, an administrator can turn training off for the whole organization. To opt out: Admin panel > Manage > Vibe > Privacy, turn off "Allow your interactions to be used to train our models" (mobile: Settings > Data & Account Controls > untick "Enable data sharing"). Feedback (thumbs up or down with a comment) may be used on any plan. Source for Mistral Le Chat, Training on your dataApplies to: Vibe (Vibe (Free, Pro) and Vibe (Teams))The feedback sentence is on https://help.mistral.ai/en/articles/347617-do-you-use-my-user-data-to-train-your-artificial-intelligence-models.Checked 2026-10-08 Enterprise customers are opted out of training by default. The opt-in toggle is managed by admins. Source for Mistral Le Chat, Training on your dataApplies to: Vibe (Enterprise)Checked 2026-10-08 API (Mistral Studio): Free mode may be used for training (opt-out available), and Pay-as-you-go customers can opt out at any time. The Vibe and API opt-out toggles are separate. Source for Mistral Le Chat, Training on your dataApplies to: Mistral Studio (Free mode, Pay-as-you-go)Checked 2026-10-08 Data accessed through Connectors is never used to train or fine-tune models, on any plan. Source for Mistral Le Chat, Training on your dataApplies to: Vibe Work (All plans)Checked 2026-10-08

Training on your data: Tools built for nonprofits
Givebutter	Grantable
Givebutter's agreements with Anthropic, Google and OpenAI are designed to prohibit use of data submitted through the platform for training their models. Givebutter says it does not collect, use or sell personal data to train large language models. Source for Givebutter, Training on your dataApplies to: Givebutter MagicChecked 2026-10-08 Givebutter says it does not currently fine-tune or post-train third-party AI models using customer data. Source for Givebutter, Training on your dataApplies to: Givebutter MagicChecked 2026-10-08 Givebutter trains its own predictive ("homegrown") models on aggregated, de-identified platform data by default. Fundraisers can opt out by contacting Givebutter Support. Source for Givebutter, Training on your dataApplies to: Givebutter homegrown modelsChecked 2026-10-08 Caveat stated by Givebutter: for its AI infrastructure, observability services and internal AI tools, it will take reasonable steps to seek contractual limits on AI training, but in some cases data may be used for AI model training. Source for Givebutter, Training on your dataApplies to: GivebutterChecked 2026-10-08 Caveat for AI Connectors: some consumer-tier accounts on the connected AI platforms may allow training on the user's data. Source for Givebutter, Training on your dataApplies to: Givebutter AI ConnectorsChecked 2026-10-08	Grantable says your content is never used to train AI models, its own or anyone else's. Its AI providers are contractually barred from training on customer data. No opt-out is needed because training is off for all plans. Source for Grantable, Training on your dataApplies to: Grantable (All plans)Checked 2026-10-08

Questions to ask any provider

Ask these about privacy and training data before you commit. See the full printable checklist.

PrivacyHow long do you keep what we type in, and can we change that?Data that is kept longer has more time to be leaked or requested by others.Do you use our inputs to train your models, and how do we opt out?If your inputs train a model, pieces of them could shape answers given to other people.Where is our data stored, and in which countries?Some funders and laws limit where data about your clients may be held.How do we delete our data, and how long until it is really gone?Deleting from your screen is not always the same as deleting from the provider's systems.Will you sign a business associate agreement (BAA) for the plan we would use?If HIPAA applies to you, a vendor needs a signed BAA before it handles protected health information for you.Will you sign a data processing agreement, and where will our data be stored or transferred?If the GDPR applies, the vendor's work must be set out in a contract, and moving data out of the EU needs an approved route, such as an adequacy decision or standard contractual clauses.

New to this? Read lesson 3 first.