AI Training Opt Out Settings Guide: Where to Switch Off Model Training in Popular AI Tools

Whether a prompt can end up in a provider's training data usually depends on which account the user signed in with and one setting buried in a privacy menu, not on the organisation's contract. Consumer and individual plans of the major AI assistants allow training on conversations unless the user opts out, while business plans generally exclude it by default. This guide sets out, tool by tool, where the training control sits, what it does and does not cover, and how long data is kept either way. It then turns that into an organisational approach: steering staff to business accounts, setting the controls admins own, and checking that it works. It is for security, privacy and IT teams writing AI use rules, and for anyone answering the question "does this tool train on my data?". Platforms are listed alphabetically, and order is never a ranking.

Key takeaways

• The account type decides the default: each provider covered here states that its business or commercial offerings are not used for training by default, while individual plans may be unless the user changes a setting.
• Opting out is forward-looking: the providers describe the setting as applying to new conversations; Anthropic states that data already used in training is not removed from trained models.
• Feedback and safety review are separate routes: thumbs up or down ratings and conversations flagged for safety review can be kept and used even where routine training is off, so opt-out is not the same as "nothing is retained".
• Retention continues after opt-out: short retention windows (72 hours to 30 days in the providers' own documentation) still apply to keep the service running and to monitor abuse.
• Settings are not a control on their own: an opted-out personal account still sends data outside your tenant, so the organisational answer is approved business accounts, identity-based access and monitoring, with the personal settings as a backstop.

Why training use is a data governance question

When a provider trains on a conversation, the content leaves the narrow purpose of answering the user and becomes part of the provider's model development data, retained for longer and handled by more people and systems. The direct risk of a model reproducing a specific confidential passage is hard to quantify, but the governance problems are concrete: confidential or personal data processed for a purpose the organisation never agreed, retention periods that outlast your own records policy, and no contractual protection because the user accepted consumer terms. The Samsung ChatGPT source code leak in 2023, where engineers pasted source code and a meeting transcript into a public chatbot within weeks of being allowed to use it, shows how quickly this happens once a tool is available and no approved alternative exists.

The same tool brand often has two different sets of terms. A user signed in with a personal account is on consumer terms; the same user signed in with the organisation's account on a business plan is usually on commercial terms that exclude training. Most of the practical work is making sure staff are in the second case. Wider embedded AI terms inside SaaS products are covered in the Embedded AI in SaaS Guide; this page focuses on the assistants people use directly.

Training controls at a glance

Each row is drawn from the provider's own help or privacy pages as of November 2026. Setting names and menus change, so treat the table as a starting point and confirm against the linked page.

Tool-by-tool settings and what they mean

Anthropic Claude

• Control: Anthropic updated its consumer terms on 28 August 2025, asking existing Free, Pro and Max users to choose a model training preference by 8 October 2025. The setting is now called Model Improvement in Privacy Settings and can be changed at any time.
• Retention: if the user allows training, data is retained for five years; otherwise the existing 30-day retention applies. The setting applies to new or resumed chats and coding sessions, and a deleted conversation is not used for future training. If a user later turns training off, Anthropic states that data already used in training or in trained models is not removed.
• Exceptions: feedback submitted with thumbs up or down stores the related conversation for up to five years, de-linked from the user ID; conversations flagged for safety review may be used to improve Usage Policy enforcement. Raw content from connectors such as Google Drive or MCP servers is excluded unless pasted into the chat.
• Admin route: on Team and Enterprise plans, Primary Owners and Owners can turn off member feedback with the "Rate chats" setting under Organization settings, Data and Privacy. See the Claude for Enterprise Security Guide.

GitHub Copilot

• Control: GitHub's documentation states that from 24 April 2026 it may use interactions on Copilot Free, Pro, Pro+ and Max plans for model training, and users can opt out with "Allow GitHub to use my data for AI model training" in their Copilot settings.
• Scope: the setting is per personal account. A developer with both a personal plan and an organisation-assigned licence should check which account their IDE is signed in to.
• Exceptions and checks: if the setting is missing, GitHub notes the user is probably signed in with a Business or Enterprise licence, where the Data Protection Agreement prohibits training use without customer authorisation.
• Admin route: assign Copilot Business or Enterprise seats to developers who work on organisation code. See the GitHub Copilot Security Guide.

Google Gemini

• Control: on personal accounts, Keep Activity governs both history and training. When it is on, chats, uploads, screens, audio and Gemini Live video are saved and used to improve Google services, including training generative AI models, with auto-delete after 18 months by default (3 or 36 months, or off, are options).
• Retention when off: chats are kept for 72 hours to respond, process feedback and protect users, and are not used for training unless the user sends feedback. Temporary chats follow the same 72-hour rule.
• Exceptions: conversations selected for human review are disconnected from the account and retained for up to three years, and deleting activity does not delete them. Some features, such as Tasks and some Connected Apps, need Keep Activity on, which creates pressure to turn it back on.
• Admin route: for qualifying Workspace editions, Google states that customer data, including prompts, is not used for model training without permission and is not human reviewed outside the domain. Admins control Gemini app conversation history (3, 18 or 36 months, default 18) and can turn the app off as it is a core service on by default. See the Gemini for Google Workspace Security Guide.

Microsoft Copilot

• Control: Microsoft's privacy FAQ describes a model training control for consumer Copilot and links to the Copilot privacy controls page for it. Opting out excludes future conversations, and users can opt back in later.
• Who is excluded anyway: Microsoft lists users signed in with an organisational Entra ID account, Microsoft 365 Personal and Family subscribers using Copilot in Microsoft 365 apps, users not signed in, signed-in users under 18, and users in Brazil, China (excluding Hong Kong), Israel, Nigeria, South Korea and Vietnam.
• Version change: Microsoft notes that an updated Copilot app was released on 18 August 2026 and that users of the older app must sign in to the updated app by 31 December 2026 or unmigrated content is deleted. Recheck the setting after migration.
• Admin route: make sure staff use Copilot with their work account so Enterprise Data Protection applies. See the Microsoft 365 Copilot Security Deep Dive.

OpenAI ChatGPT

• Control: "Improve the model for everyone" under Settings, Data Controls, on web (including logged-out use), iOS and Android. It is account-wide, and once off, new conversations are not used for training while history remains visible. OpenAI also offers a privacy portal request ("do not train on my content"), and earlier opt-outs made through support remain honoured.
• Retention: Temporary Chats are not used for training and are deleted from OpenAI's systems within 30 days, reviewed only when needed to monitor for abuse.
• Exceptions and checks: conversations from before the opt-out are not described as removed, so opt out before first use. Export from the same Data Controls menu lets a user see what is stored.
• Admin route: OpenAI states it does not train on content from its business offerings, naming ChatGPT Team, ChatGPT Enterprise and the API Platform, and ChatGPT Enterprise admins control retention. See the ChatGPT Enterprise Security Guide.

Availability notes: Anthropic, GitHub, Google, Microsoft and OpenAI each state that their terms differ by plan or account type, as summarised in the admin routes above. Confirm which plan a user is actually on before relying on any default.

What opt-out does not cover

An organisational approach in five phases

1. Discover current use. Use secure web gateway, CASB or browser telemetry to find which AI domains are used and whether sign-ins come from corporate or personal identities. The Shadow AI Discovery Guide covers the methods.
2. Provide approved business accounts. For each tool people actually need, buy the business tier whose terms exclude training and connect it to single sign-on, so the work identity is the easy path. Record the contractual training and retention terms in the supplier register.
3. Set the admin controls you own. Retention periods (Gemini app history, ChatGPT Enterprise retention), member feedback (Claude "Rate chats"), and app availability in Workspace or Microsoft 365. Document each value and who changed it.
4. Write the personal-account rule. In the AI Acceptable Use Policy Template, state that work data goes only into approved business accounts and that, where personal accounts are tolerated for non-confidential tasks, training opt-out and temporary modes must be used. Give staff the setting paths from the table above.
5. Enforce and test. Block or warn on consumer sign-in to tools where a business tenant exists (tenant restrictions or login-domain controls in your proxy), add generative AI DLP rules as described in the Generative AI Data Loss Prevention Guide, and pilot with one department before wider rollout. Roll back by moving a control from block to warn if business-critical work breaks, rather than removing it.

Decision guidance

• Small organisations: one approved business tool, SSO, and a written personal-account rule with the opt-out steps usually covers most risk. Checking settings on individual devices does not scale and is not verifiable.
• Large or regulated organisations: combine tenant-restricted business accounts, inline DLP on prompts and uploads, and a block on consumer AI for regulated data classes. Personal settings are a residual measure only.
• Developers: code assistants are often signed in with personal accounts inside the IDE. Assign organisation seats and check the signed-in account as part of developer onboarding; see also AI Coding Tools and Secrets Sprawl.
• Personal data: if staff enter personal data into a tool that may train on it, the organisation has a purpose and processor question under data protection law, not only a security one. See the AI and GDPR Guide.

Detection and measurement

• Log: proxy or gateway records of AI domain use with the identity domain at sign-in, DLP events on prompts and file uploads to AI services, and admin audit logs for changes to AI retention and feedback settings.
• Review: quarterly, re-read each approved provider's privacy and training pages, because defaults and plan names change (GitHub's April 2026 change and Microsoft's August 2026 app update are recent examples).

Common failure patterns

• Trusting the brand, not the account: the organisation buys the business tier, but staff keep using a personal login in another browser profile, under consumer terms.
• Opting out after the fact: users change the setting once they hear about it, but conversations already shared are not withdrawn from use.
• Forgetting feedback: a user who is opted out rates a response, sharing the full conversation through the feedback route.
• Feature pressure: a feature that needs history or activity on (for example Gemini Tasks) leads users to turn the setting back on without realising it also controls training.
• Stale guidance: internal instructions name menus that have since moved or plans that have been renamed, so staff cannot find the control.

Framework mapping

Practitioner checklist

• List every AI assistant in use and whether staff sign in with corporate or personal identities.
• Provide business-tier accounts with SSO for each tool with a real business need, and record their training and retention terms.
• Set admin retention, feedback and availability controls in each business console and document the values.
• Publish the opt-out paths from this guide in the AI use rule for any tolerated personal use.
• Tell staff that feedback ratings and safety review are not covered by opt-out.
• Block or warn on consumer sign-in where a business tenant exists, and add DLP for prompts and uploads.
• Include the signed-in account check in developer onboarding for code assistants.
• Re-read provider training pages every quarter and update internal guidance when menus or plans change.

Standards and references

• Anthropic: Is my data used for model training? (consumer)
• Anthropic: Is my data used for model training? (commercial)
• GitHub Docs: Managing Copilot policies as an individual subscriber
• Google: Gemini Apps Privacy Hub
• Google Workspace: Generative AI in Google Workspace Privacy Hub
• Microsoft: Privacy FAQ for Microsoft Copilot
• OpenAI: Data Controls FAQ
• OpenAI: How your data is used to improve model performance

Related NHI Mgmt Group resources: Embedded AI in SaaS Guide · Shadow AI Discovery Guide · Generative AI Data Loss Prevention Guide · AI Data Retention and Deletion · AI Acceptable Use Policy Template