Last verified: April 2026

AI Data Retention &
Training Opt-Out Guide

Step-by-step instructions for configuring retention policies and opting out of model training across Claude, ChatGPT, Gemini, and AWS Bedrock — for every plan type, interface, and surface.

Quick Comparison

Training defaults and retention across all four providers

ClaudeChatGPTGeminiBedrock
Consumer training defaultONONONN/A
Can opt out?YesYesYesN/A
Keep history after opt-out?YesYesNoN/A
Business/Team trained on?NeverNeverNeverNever
Custom retention policiesEnterprise onlyEnterprise onlyWorkspace adminCustomer-managed
Zero Data RetentionAPI / EnterpriseAPI / EnterpriseNoDefault
Feedback data retained?Up to 5 yearsYes (training eligible)Up to 72 hrs minNo
Admin can export user chats?Primary OwnerEnterprise (Compliance API)Vault + Data ExportCustomer owns all data
Paying for a "Pro" or "Plus" individual subscription does not protect your data from training on any platform. Only Team, Enterprise, or API plans provide contractual training exclusions by default.

Free · Pro · Max
Applies to claude.ai (web), Claude Desktop app, Claude Mobile app, and Claude Code when used with these plans.

1
Opt Out of Model Training

WebDesktopMobile

Open claude.ai (or the Claude Desktop/Mobile app) and sign in.

Click your profile icon (bottom-left on web/desktop, top-right on mobile).

Select Settings → Privacy (may appear as "Data Privacy Controls").

Find "Help improve Claude" (or "Model Improvement") and toggle it OFF.

Confirm the change. It takes effect immediately for new and resumed chats only.

Opting out only affects future conversations. Data already submitted for training (while the toggle was on) cannot be retroactively removed from models that have already been trained.

2
Use Incognito Mode for Sensitive Chats

WebDesktopMobile

Start a new chat and click the ghost icon (top-right of the chat area).

The chat is now Incognito — it will not be saved to history, memory, or used for training, even if you have Model Improvement enabled.

Close the incognito chat when done. Normal chats resume saving afterwards.

On Team and Enterprise plans, incognito chats are still included in admin data exports and follow organizational retention policies. Incognito only prevents memory and training use — not admin visibility.

3
Delete Conversations

In the sidebar, hover over a conversation and click the trash icon, or open the conversation and select Delete from the menu.

Deleted conversations are purged from back-end logs within 30 days.

Deleted chats are not used for future model training, regardless of your privacy setting.

Caveats & Risks — Consumer Plans
Thumbs up/down feedback: When you submit feedback via the 👍/👎 buttons, the entire related conversation is stored for up to 5 years — even if you've opted out of training. Avoid using the feedback buttons on sensitive chats.
Retention differences: Opted-in users → 5-year retention. Opted-out users → 30-day back-end retention. Neither is zero.
Connectors are excluded: Raw content pulled from Google Drive, MCP servers, and other connectors is not included in training data. However, if you paste that content directly into the chat, it is included.
Claude for Chrome: Data collected when using the Claude in Chrome browser extension follows the same consumer privacy settings. It is training-eligible if Model Improvement is on.
Safety flags: Conversations flagged by trust and safety classifiers may be retained for up to 2 years regardless of your settings. Safety classification scores can be retained for 7 years.

Team · Enterprise
Governed by Commercial Terms. Data is never used for model training by default — no opt-out needed.

Training Status: Automatically Excluded

No action required. Team and Enterprise conversations, projects, and Claude Code sessions are never used for training under Commercial Terms unless the org admin explicitly opts in via the Development Partner Program.

Configure Retention Policies (Enterprise Only)

WebDesktopEnterprise only

Pre-requisite: You must be a Primary Owner or Owner on the Enterprise plan.

Navigate to Organization Settings → Data and Privacy.

Set your desired retention period. The minimum is 30 days (each month = 30 days).

Click Save. All data that already falls outside the new window will be deleted immediately.

Irreversible on save: When you first set a retention window (e.g. 90 days), all conversations older than 90 days are deleted the moment you save. This cannot be undone. Plan carefully before enabling.
How it works: Chat retention starts from the last message. Project retention starts from the last update. Deletion happens at midnight UTC. All changes are tracked in audit logs. Default is indefinite retention if no policy is set.
Team Plan Limitations

The Team plan does not have custom retention controls. Data is retained indefinitely unless individual users manually delete their conversations. Retention policy configuration requires upgrading to Enterprise.

However, the Primary Owner of a Team plan can export all user conversation data via Organization Settings → Data and Privacy → Export Data.


Anthropic API · Claude Code (Commercial)
Data is never used for training. Additional ZDR options available.

Zero Data Retention (ZDR)

Pre-requisite: A signed ZDR contract/addendum with Anthropic.

Contact the Anthropic sales team to enable ZDR on your organization.

Once enabled, prompts and responses are not stored at rest after the API response is returned.

ZDR must be enabled per-organization — each new org needs separate setup.

ZDR limitations: Some features (like Message Batches, prompt caching beyond a session) may have qualified ZDR eligibility — meaning narrow, performance-required data may still be retained briefly. Check the feature eligibility table in Anthropic's docs. Standard (non-ZDR) API retention is 30 days.
Claude Code on Enterprise: ZDR applies to Claude Code when used through Claude Enterprise with ZDR enabled. Local session transcripts are cached in ~/.claude/projects/ for 30 days by default (configurable via cleanupPeriodDays).
Free · Plus · Pro
Applies to chatgpt.com (web), ChatGPT Desktop app, ChatGPT Mobile app, and ChatGPT in Chrome (browsing agent).

1
Opt Out of Model Training

WebDesktopMobile

Open chatgpt.com (or the ChatGPT app) and sign in.

Click your profile icon (bottom-left on web).

Select Settings → Data Controls.

Find "Improve the model for everyone" and toggle it OFF.

For voice mode, there's a separate toggle: "Voice mode audio" — turn that off too if applicable.

Unique advantage: ChatGPT is the only major provider that lets you keep full chat history indefinitely without consenting to model training. You don't have to sacrifice history for privacy.

2
Use Temporary Chat for Sensitive Sessions

Start a new chat and enable Temporary Chat (toggle in the top-right of the chat window).

Temporary chats are not saved to history and not used for training.

Content is still retained for 30 days for abuse monitoring — it's not zero-retention.

Caveats & Risks — Consumer Plans
Thumbs up/down feedback may explicitly opt specific conversations into training, even if you've turned training off globally. Be cautious with the feedback buttons.
30-day safety retention: Even with training off and Temporary Chat enabled, OpenAI retains data for 30 days for abuse detection. This is not configurable on consumer plans.
ChatGPT in Chrome: The browsing agent extension follows the same privacy settings as your main ChatGPT account.
Data export on consumer: You can export your own data via Settings → Data Controls → Export Data. This includes your chat history as a JSON zip file. The link expires in 24 hours.

Business (formerly Team) · Enterprise
Governed by Business Terms / Enterprise Terms. Training is excluded by default.

Training Status: Automatically Excluded

No action required. Workspace conversations are never used for training under the Business/Enterprise Terms. OpenAI does not train on your workspace data.

Data Retention & Admin Access

Business plan: Admins and owners cannot casually browse member chats. There is no org-wide data export available on the Business tier.

Enterprise plan: Admins can access user data via the Compliance API, which provides audit logging of conversations.

Enterprise admins can configure data retention policies and may access, export, audit, retain, and delete data tied to managed accounts.

Business plan limitation: No built-in org-wide data export exists. If you need compliance exports, you'll need to upgrade to Enterprise for the Compliance API.
Managed Account Notice

If you're on a managed Business or Enterprise account, your administrator may be able to: access prompts & outputs, export conversation data, configure retention, restrict features, suspend access, or delete accounts. OpenAI's standard consumer Privacy Policy does not apply while signed into a managed account.


OpenAI API Platform
API data is NOT used for training by default. Standard 30-day retention for abuse monitoring.

Zero Data Retention (ZDR)

API data is not used for training by default — no opt-out needed.

Standard retention: 30 days for abuse monitoring.

For Zero Data Retention, contact OpenAI sales for an Enterprise API agreement. ZDR eliminates the 30-day safety retention window.

Free · Gemini Advanced (Google One AI Premium)
Applies to gemini.google.com (web), Gemini Mobile app, and Gemini in Chrome.

1
Opt Out of Model Training (Turn Off Gemini Apps Activity)

WebMobileChrome

Open gemini.google.com and sign in.

Click the hamburger menu (≡) in the top-left corner.

Select "Activity" from the menu.

Next to "Gemini Apps Activity", click "Turn Off".

Choose "Turn Off and Delete Activity" to also remove existing stored data. Click Next → Delete.

Critical trade-off: Turning off Gemini Apps Activity means you lose all conversation history. Once you leave a conversation, you cannot see it again. You cannot keep history while opting out — unlike Claude and ChatGPT. Google also states conversations may be reviewed by human reviewers when Activity is on.

2
Alternative: Disable Web & App Activity (Broader)

Go to myaccount.google.com → Data & Privacy.

Under "Web & App Activity", toggle it OFF.

This prevents Google from storing your activity across all Google services, including Gemini. Note: this affects Google Search, Maps, Assistant, etc.

Caveats & Risks — Consumer Plans
72-hour minimum retention: Even with Activity turned off, Google retains conversations for up to 72 hours to provide the service. This is not configurable.
No bulk export: Gemini chat content is not available through Google Takeout. You can view activity at myactivity.google.com/product/gemini but cannot download it in bulk.
Gemini in Chrome & mobile: Usage across gemini.google.com, Gemini mobile app, and Gemini in Chrome all share the same Activity setting and count toward the same usage limits.
Human review: When Gemini Apps Activity is ON, your conversations may be reviewed by human reviewers to improve Google's products. Avoid entering confidential or sensitive information while Activity is enabled.

Google Workspace with Gemini (Business / Enterprise)
Governed by Google Workspace Terms of Service. Data is never used for model training.

Training Status: Automatically Excluded

When you use Gemini through a paid Google Workspace account, your data is treated like any Workspace data (Gmail, Drive, etc.) — private and never used for training. No opt-out needed.

Configure Retention Policies (Workspace Admin)

Admin ConsoleSuper Admin required

Pre-requisite: You must be a Super Administrator with the Gemini Settings admin privilege.

Open the Google Admin Console → Menu → Generative AI → Gemini for Workspace.

Go to Conversation history & deletion.

Conversation history: Toggle ON or OFF for users/groups/OUs. When OFF, users cannot access past conversations.

Auto-deletion: Choose retention: 90 days (~3 months), 540 days (~18 months), 1080 days (~3 years), or indefinite. Default is 18 months.

Manual deletion: Choose whether to allow users to delete their own conversations.

Apply settings per organizational unit or configuration group for granular control.

🔍
Admin Access: Google Vault (eDiscovery)

Open Google Vault (vault.google.com).

Create a Matter, then use Search to find Gemini conversations by user, date range, or keyword.

Export relevant prompts and responses for compliance or legal review.

Most granular admin access: Google Vault lets admins search specific conversations by keyword — not just bulk export. This is significantly more powerful than Claude or ChatGPT's admin tools. Users should be aware their employer can search their Gemini prompts.
Additional Admin Tools

Data Export Tool: Admins can export Gemini app and NotebookLM data for archive/backup from the Admin Console.

Usage Reports: Admin Console → Generative AI → Gemini Reports shows per-user usage levels, active users, and feature adoption.

Gemini as Additional Service: If Gemini is enabled as an "additional service" (not a core Workspace service), the Gemini Apps Privacy Notice applies — conversations may be reviewed by human reviewers. Admins should verify whether Gemini is running as a core or additional service.

Amazon Bedrock (Claude on AWS)
AWS-managed infrastructure. No consumer plan exists — all usage is enterprise/API. Anthropic never receives or stores your data.

Training Status: Never Trained On

Anthropic does not use any data sent through Bedrock for model training — ever. Your prompts and outputs stay within your AWS account. AWS also does not use your content to train its own AI models (unless you explicitly opt in to specific AWS programs).

Bedrock provides the strongest default privacy posture of any way to access Claude. No opt-out needed, no admin toggles, no risk of accidental training inclusion.

Configure Retention & Logging

Default: Bedrock does not store your prompts or model outputs. Zero data retention is the default behavior.

Optional logging: You can opt in to model invocation logging via the Bedrock Console → Settings → Model invocation logging.

If enabled, logs are sent to your own S3 bucket and/or CloudWatch — you own and control this data entirely.

Use AWS CloudTrail for API-level audit logging (who called what, when) separate from content logging.

Configure S3 lifecycle policies or CloudWatch retention settings to manage how long logged data is stored.

🔒
Additional Security Controls

VPC Endpoints: Use AWS PrivateLink to keep all Bedrock traffic within your VPC — no data crosses the public internet.

IAM Policies: Control which users/roles can invoke specific models using fine-grained IAM permissions.

Guardrails for Bedrock: Configure content filters, denied topics, PII redaction, and sensitive information filters at the API level.

KMS Encryption: All data is encrypted at rest with AWS KMS. You can provide your own customer-managed keys (CMK) for additional control.

Also applies to: Google Cloud Vertex AI

If you access Claude through Google Cloud Vertex AI, similar principles apply:

Anthropic does not use Vertex AI customer data for training. Google does not train foundation models on your Vertex AI inputs.
Logs are stored by Google Cloud for approximately 30 days by default. You can configure shorter periods using Google Cloud Logging. Both Bedrock and Vertex support Private Service Connect (PSC) / PrivateLink for regional data residency.

Cross-Platform Considerations

🔑 Pre-Requisites Checklist

Know your plan type. "Pro" and "Plus" are consumer plans — they don't protect your data. Verify whether you're on a consumer or commercial plan in your account settings.

Know your role. Retention policies require admin/owner privileges. Regular users cannot change organizational settings.

Check before you save. On Claude Enterprise and Gemini Workspace, changing retention settings can trigger immediate, irreversible deletion of older data.

Export before changing. Always run a data export before modifying retention policies. Once data is deleted, it's gone permanently.

Audit your feedback. On both Claude and ChatGPT, using the 👍/👎 buttons may commit conversations to long-term storage or training — regardless of your other settings.

⚖ Risks & Caveats Summary

RiskAffectedMitigation
Opting out doesn't remove data from already-trained modelsAll consumerOpt out early; delete sensitive chats
Feedback buttons override privacy settingsClaude, ChatGPTAvoid 👍/👎 on sensitive conversations
Incognito chats visible to Enterprise adminsClaude Team/Ent.Treat incognito as memory-only protection
History lost when opting out of trainingGemini (consumer)Export important chats before toggling
Retention changes delete data immediatelyClaude Ent., Gemini WSExport data before setting new policies
Safety-flagged content retained for yearsAll platformsUnavoidable — by design for safety
Local caches on device (Claude Code, Desktop apps)Claude Code, all desktopsConfigure cleanupPeriodDays; clear app data
Connector data pasted into chat = training-eligibleClaude consumerDon't paste raw connector data into chats
Google Vault enables keyword-level admin searchGemini WorkspaceUsers should assume full visibility
No bulk export for consumer GeminiGemini Free/AdvancedUse extensions or manual copy