Does Claude Train on Your Chats? What to Redact First

Claude consumer users can control whether new chats and coding sessions are used for future model training. Anthropic says turning off Help Improve our AI models stops that use for future training runs, while Incognito chats are not used for model improvement. Commercial products follow different terms. None of these settings make live credentials or unnecessary personal data safe to paste.

The short answer for Claude consumer accounts

Anthropic's current consumer model-training guidance says chats and coding sessions may be used to improve Claude when the user allows it, when a conversation is flagged for safety review, or when the user otherwise explicitly opts in.

The company also states that Incognito chats are not used to improve Claude even when model improvement is enabled. Safety review and policy enforcement can still involve separate processing, so “Incognito” does not mean the service never receives or handles the conversation.

How to turn model improvement off

For Claude's consumer products, Anthropic documents the following browser workflow:

1. Open Claude and select your name or account menu.
2. Open Settings.
3. Select Privacy.
4. Turn off Help Improve our AI models.

According to Anthropic's privacy-settings documentation, disabling the setting prevents new chats and coding sessions from being used in future training runs. It does not remove information from models already trained or a training run already underway.

Consumer Claude and commercial Claude are different

Privacy discussions often become inaccurate because they combine consumer subscriptions with commercial services. Anthropic explicitly separates Claude Free, Pro, and Max from products covered by its commercial terms, including Claude for Work and the API.

Before a company approves Claude for customer data or confidential documents, someone should verify:

• The exact Claude plan and authentication method
• The contract and data-processing terms that apply
• Retention and deletion controls
• Whether feedback or safety-review exceptions apply
• Which integrations, connectors, or tools can receive context

A personal Pro account used for work should not be assumed to inherit an employer's commercial protections.

What to redact even with training disabled

Training is only one stage of data handling. Your prompt must still be sent to Claude for inference. It may also appear in conversation history, feedback workflows, local session records, connected tools, or organizational audit systems depending on the product.

Remove values that the model does not need:

• API keys and tokens: replace them with typed labels such as [GITHUB_TOKEN].
• Database credentials: keep a safe URI shape but use a fake username, host, and password.
• Customer and employee data: substitute names, emails, phone numbers, account numbers, and identifiers.
• Health and financial information: remove it unless your approved workflow and agreement explicitly cover it.
• Internal infrastructure: generalize private domains, IP addresses, bucket names, and project codenames.
• Proprietary code: share the smallest function or reproducible example rather than an entire repository.

Be careful when sending feedback

Feedback buttons can have different retention implications from an ordinary chat. Anthropic says that when users provide thumbs-up or thumbs-down feedback, the related conversation can be stored for up to five years. Read the current notice before submitting feedback on a conversation containing sensitive material.

This is easy to overlook: a developer sees an incorrect answer, attaches feedback, and unintentionally includes the surrounding customer example or code context.

Does Claude Code change the risk?

Claude Code can work with far more context than a manually pasted chat: repository files, commands, generated output, and tool results. That makes it useful, but it also makes accidental exposure less obvious.

Before starting a coding session:

1. Confirm which account and terms apply.
2. Review the model-improvement setting for consumer use.
3. Keep .env, credentials, and private keys out of context.
4. Use scoped development secrets instead of production credentials.
5. Inspect files and command output before deliberately attaching them.

If your immediate question is about environment files, see our practical .env sanitization guide. The same principle applies regardless of which model receives the prompt.

A safer prompt still preserves context

Instead of this:

Customer Jane Smith cannot connect.
DATABASE_URL=postgres://admin:real-password@prod-db.internal/customer
Please debug this.

Send this:

Customer [CUSTOMER_1] cannot connect.
DATABASE_URL=postgres://example_user:[PASSWORD]@db.example.test/customer
Please debug the URI structure and connection handling.

Claude still has the protocol, field names, and task. The credential, customer identity, and internal host add risk without improving the answer.

Settings are a baseline, not a data-minimization strategy

Turn off model improvement if that matches your preferences, use Incognito chats when appropriate, and use commercial products under reviewed terms for organizational work. Then apply the more durable rule: do not send data the model does not need.

OmniShield masks matching credentials and PII locally at send time on supported Claude and other AI-chat pages. You can also clean a prompt manually with the free browser sanitizer. Neither replaces reviewing your account settings and organizational policy.

Product settings and privacy terms change. This article reflects Anthropic's public documentation reviewed on September 3, 2026 and is not legal advice.