How to Request Removal of Personal Data from AI Services (2026)

There is no single switch that removes a person's information from every AI product, training dataset, crawler, search index, or source website. The correct action depends on where the information appears and how the service says it uses it.

This source-reviewed guide separates five different layers: a public source page, a crawler or archive, an AI product account, a model or search output, and a commercial data provider. A request at one layer does not automatically change the others. Provider controls, account tiers, regions, and legal exceptions also change, so read the current first-party instructions before submitting sensitive information.

Start with the exact exposure

Before filing a request, record what you actually observed:

• the provider or website name and current URL;
• the exact text, image, identifier, or account activity involved;
• whether the result came from a live webpage, a search retrieval, or an AI-generated answer;
• the date, account, product tier, region, and connected service involved; and
• the action you want: stop future use, delete a stored account item, correct a record, object to processing, or remove a source page.

An AI answer is not proof of its source. It may reflect a public page, a connected tool, an uploaded document, a user prompt, a search retrieval, or an inaccurate generation. Verify the underlying page before asserting that a company trained on a particular data-broker record.

The controls are different

Account training controls

An account setting may exclude future prompts, uploads, or interactions from a named model-improvement program. It usually does not mean that the chat is deleted, that a backup is erased, or that a deployed model has changed.

Stored activity and account deletion

Chat history, uploaded files, feedback, account details, security logs, and abuse-prevention records can have separate controls. Delete stored activity through the provider's documented route, and do not describe a training toggle as a deletion request.

Public web crawling

A website owner can publish `robots.txt` instructions for documented crawlers. Those instructions concern a crawler's future access to a path; they are not a universal deletion request, security control, or recall mechanism for copied data.

Source and downstream removal

Removing a source page can affect a later crawl or search result, but it does not prove that an AI model, archive, screenshot, or third-party dataset changed. Contact the source publisher and the relevant search or AI provider separately when their current routes fit.

OpenAI and ChatGPT

OpenAI's current data-use guidance describes a Data Controls setting for personal ChatGPT accounts. Review the explanation beside Improve the model for everyone and turn it off if you do not want future conversations used within that stated program. Product, region, account, and policy changes control the live scope.

For business, enterprise, education, or API work, use the applicable OpenAI business-data information and API data-controls documentation. Do not transfer a personal ChatGPT statement to an API or organization account.

If ChatGPT displays personal information about you, preserve the exact prompt, answer, account context, and any source links, then use the OpenAI Privacy Portal or the current help route. A request may concern a stored account item, a search result, or personal-data processing. Do not promise that a request will retrain a model or remove every related output.

Anthropic and Claude

Anthropic's Privacy Center guidance distinguishes consumer controls, Incognito conversations, and commercial products. Read the current explanation for the plan you use. A consumer setting does not automatically determine the policy for Claude for Work, an API account, an administrator-managed workspace, or a feedback submission.

If you need access, correction, deletion, or an objection, use Anthropic's verified privacy route and describe the exact account and data path. Send the minimum information needed to match the request. Do not claim that a privacy response proves an existing model has forgotten an identifier unless Anthropic explicitly says so.

Google Gemini and Google Search

Google's Gemini activity guidance describes Keep Activity and related deletion controls. Google explains that turning the setting off can change whether future chats appear in activity or are used to improve machine-learning technologies, while limited retention and service or safety use can still apply. The setting may also have a delay before it takes effect.

Use the Gemini Apps Privacy Hub for human-review, connected-app, and feature-specific details. Delete activity separately when the current control offers that option. If a Google Search result exposes personal information, use Google's current Results About You or other eligible removal route; a Gemini activity change does not remove a Search result.

Meta AI and public social content

Start with Meta's generative-AI privacy information and the current control shown for your country and account. Meta's rules can differ by product, content visibility, region, and the way the information was submitted.

Separate public posts, private messages, profile fields, uploaded media, interactions with an AI feature, and feedback. Changing visibility or deleting a post may affect a source page, while an AI training or objection control may concern a different processing path. Preserve the acknowledgement and read the stated scope and exceptions before describing the result as complete deletion.

Website owners: using `robots.txt`

If you operate a website, publish crawler preferences only for documented user agents and only for paths you control. Google's Google-Extended documentation says the token can control Google's use of crawled site content for certain Gemini training and grounding purposes without affecting Google Search ranking.

For example:

User-agent: Google-Extended
Disallow: /

Do not add a vendor token merely because a blog post lists it. Check the vendor's current documentation first. A `robots.txt` rule can be ignored by an unknown or non-compliant crawler, does not control a user download or an archive, and does not erase content collected before the rule was published. To remove a page from a search engine, use the search engine's current process after addressing the source page where possible.

Commercial data sources and people-search sites

People-search providers may publish names, contact details, relatives, or other profile fields. That does not prove that a specific AI company bought the record, used it for model training, or can remove it from a trained model. A causal claim needs evidence about the named source and processing path.

If a commercial listing is the exposure you can verify, use that provider's current first-party opt-out or privacy route. Record the exact listing, request type, matching information, response, and later source check. OfflistMe can prepare browser-local, user-reviewed drafts for supported workflows; the user reviews and sends each request, and provider requirements vary. That workflow is source cleanup, not a model-untraining guarantee.

Also separate a data-broker record from a public filing, professional page, employer system, court record, social account, search result, or third-party copy. Removing one profile does not authorize changes to a different source.

A careful request workflow

1. Identify the layer. Decide whether the request concerns a source page, account data, a training-control setting, a search result, or a provider index.
2. Read the current first-party route. Confirm the provider, account tier, region, verification method, scope, and exceptions.
3. Minimize the disclosure. Provide only what is needed to match the account or result. Do not send a government ID or sensitive document to an unverified form.
4. Choose the right action. A training opt-out, account deletion, access request, correction, objection, copyright request, and search-result removal are different processes.
5. Save evidence. Keep the URL, request date, confirmation, response, stated scope, and any required follow-up.
6. Re-check the same layer. A changed search result is not proof of model change, and a provider acknowledgement is not proof that an unrelated source changed.

Depending on your location and the provider, privacy law may provide rights such as access, correction, deletion, objection, or restriction. Eligibility, identity verification, exemptions, and response periods vary. The EU GDPR text is one example of a legal framework; do not apply it automatically to every provider or resident.

What not to claim

Avoid statements that cannot be established from a current source, such as:

• “This data-broker record definitely trained ChatGPT.”
• “A robots rule deletes everything already scraped.”
• “Turning off training makes the model forget your name.”
• “A privacy request guarantees output suppression or model retraining.”
• “Every AI company buys verified files from commercial data brokers.”
• “One deletion request covers all products in a corporate group.”

The more defensible statement is narrower: a current control or request may change the provider process it describes, subject to the provider's terms, verification, technical limits, and legal exceptions.

Frequently asked questions

Can I remove personal data from a trained model?

There is no universal method or outcome. A provider may offer an account deletion route, a personal-data request, an output-reporting process, or a future-training control. Ask what the provider's response covers. Do not represent a privacy acknowledgement as proof that all model parameters or downstream copies changed.

Does turning off AI training delete my old chats?

Not automatically. Review the provider's separate history, activity, file, and deletion controls. Retention for security, abuse prevention, feedback, backups, or legal reasons may have a different scope.

Does `robots.txt` stop all AI scraping?

No. It communicates a preference to documented crawlers that follow the protocol. It does not bind every scraper, protect copied material, or erase earlier collection.

Should I submit my full name, address, or ID to an AI privacy form?

Only provide information needed for the verified request route, and first read why the provider needs it, how it protects it, and how long it retains it. A public support post or an unverified email address is not a safe substitute for the provider's official channel.

Will removing a people-search listing stop an AI answer?

Not necessarily. The answer may have a different source, a search retrieval, a copy, or an error. Remove the verified source where appropriate, then report the exact AI output through the provider's current route and keep both records separate.

For product-specific settings, see How to Review AI Training and Data-Use Controls, Remove Personal Data from ChatGPT, Manage Personal Data in Google Gemini, and the data-broker removal guide.