Iron Goo
Two colleagues review highlighted terms beside a laptop in an office overlooking a city at dusk

Does ChatGPT Use My Data for Training?

Atamyrat Hangeldiyev
Atamyrat Hangeldiyev
Systems Architect
AIBusiness
Table of contents
  1. OpenAI: the opt-out has a feedback exception
  2. What does “unless legally required” mean in practice?
  3. Anthropic: your training choice changes retention
  4. Google: reviewed Gemini chats can survive activity deletion
  5. Microsoft: the updated Copilot app changed its training wording
  6. Consumer account controls and surviving copies
  7. What to copy into your company rule

“When you use our services for individuals (Free, Plus, Pro), we may use your content to train our models.” That is OpenAI's own wording.

So, does ChatGPT use my data for training? Yes, on individual plans unless you turn off Improve the model for everyone under Settings > Data controls. ChatGPT Business, Enterprise, Edu and the API do not use your data for training by default. The vendor pages cited here were read on 1 November 2026.

The account's governing terms determine the answer. A paid personal subscription is still a consumer account, and a promise not to train on data does not mean the vendor keeps no copy.

Company documents take consumer and business account paths with different training and retention rules

The same documents follow different paths depending on the account terms. Consumer copies can survive deletion through feedback, review or safety retention; business and API terms can prohibit training while still allowing a retention window.

OpenAI: the opt-out has a feedback exception

OpenAI's individual account opt-out applies to new conversations, and turning it off does not delete or hide saved chats. If you want a saved chat removed, you have to delete it separately. Training and retention are different controls: one governs a use of the conversation, while the other governs how long a copy remains.

The feedback button changes the individual account rule. OpenAI says the entire conversation associated with a thumbs up or thumbs down rating may be used to train its models, even after you opt out. For ChatGPT Business, Enterprise and Edu, response feedback does not permit training on the associated conversation; workspace admins can turn feedback off under Workspace settings > Allow response feedback.

Codex has a separate Include environments setting. Changing the ChatGPT setting or using the Privacy Portal does not change that Codex control, so a company using both products needs to check both.

A temporary chat is not used for training while it remains temporary, but OpenAI may keep a copy for up to 30 days for safety. Deleted chats are scheduled for permanent deletion within 30 days, with exceptions that include security and legal obligations. The OpenAI API also has a retention distinction: data sent to it is not used to train or improve OpenAI models unless you explicitly opt in, while abuse monitoring logs are retained for up to 30 days by default.

The question of which ChatGPT plan a company should pay for has its own decision. Here, the useful check is whether the account holding company data is governed by the individual terms or the business terms.

What does “unless legally required” mean in practice?

OpenAI says a court order in the New York Times lawsuit required it to preserve deleted ChatGPT chats and API content that it would normally remove within 30 days. According to OpenAI's account of the data demands, the order affected ChatGPT Free, Plus, Pro and Team subscriptions and API use without a Zero Data Retention agreement.

OpenAI says ChatGPT Enterprise, Edu and API customers with Zero Data Retention agreements were not affected. Team was renamed ChatGPT Business on 29 August 2025. The distinction matters because a company could have used a paid workspace and still fallen within the affected account group.

In its update of 22 October 2025, OpenAI says its obligations under the earlier order ended on 26 September 2025. OpenAI says it is no longer required to indefinitely retain new user data going forward or any conversations originating from the European Economic Area, Switzerland or the UK, and that it still securely stores limited historical user data from April through September 2025 because the New York Times continues to demand it.

Engadget reported on 11 October 2025 that federal judge Ona T. Wang filed a new order on 9 October freeing OpenAI from preserving and segregating output logs that would otherwise be deleted going forward. Engadget reported that logs already saved under the previous order remain accessible and that OpenAI must still hold data related to ChatGPT accounts flagged by the New York Times.

The practical reading is that a deletion window can have an active legal exception, and the account terms can determine which customers fall within it. The cost of moving once data and workflows live with a provider is a separate decision covered in the vendor lock-in analysis.

Anthropic: your training choice changes retention

Anthropic's consumer terms give Claude users a choice about model training, and that choice changes how long data is retained. The consumer terms update of 28 August 2025 applies to Claude Free, Pro and Max, including Claude Code used through those accounts. It does not apply to Claude for Work, Claude for Government, Claude for Education or API use, including use through Amazon Bedrock and Google Cloud's Vertex AI.

If a consumer user allows model training, Anthropic may retain data in a de-identified form in its training pipelines for up to five years. If the user does not allow training, the existing 30-day retention period continues. Existing users had until 8 October 2025 to make their choice.

Turning training off later does not remove data from training runs that have already started or models that have already been trained. Incognito chats are never used for training, even when the user has otherwise allowed it.

Anthropic keeps some other material longer. For chats flagged by automated trust and safety systems as potential Usage Policy violations, inputs and outputs can be kept for up to two years and safety classification scores for up to seven years. Feedback is kept for five years.

Anthropic's commercial terms state: “Anthropic may not train models on Customer Content from Services.” API inputs and outputs are deleted from the backend within 30 days by default, subject to stated exceptions. Anthropic's commercial privacy page says a thumbs rating on business plans or the API can let Anthropic use the associated conversation for training and keep it for up to five years, de-linked from user and customer IDs. Unlike Anthropic's rule, OpenAI's business tier response feedback does not permit training on the associated conversation. Primary Owners and Owners of Team or Enterprise plans can turn off thumbs feedback through Organization settings > Data and Privacy > Rate chats.

Google: reviewed Gemini chats can survive activity deletion

Google's Gemini app uses the Keep Activity setting to determine whether activity on a personal account can be used for model training. With the setting on, Google says it uses activity to improve services, including training generative AI models, with help from human reviewers. Google asks users not to enter confidential information they would not want a reviewer to see.

The Gemini Apps Privacy Hub distinguishes ordinary activity from reviewed chats. Gemini Apps Activity has an 18-month auto-delete period by default, but chats reviewed by humans are disconnected from the account and can be kept for up to three years, even after the user deletes the activity.

With Keep Activity off, future chats are not used to train AI models unless the user sends feedback, but Google still keeps chats for 72 hours. Temporary chats are also kept for 72 hours. Feedback sent with Keep Activity off attaches the last 24 hours of chats, so a rating can carry more context than the single answer on screen.

The account changes the terms. When someone uses the Gemini app with a Google Workspace account, the Workspace agreement applies: content is not human reviewed or used to train generative AI models outside the domain without permission. A personal Google account used at work remains a personal account for this purpose.

The Gemini API has a separate paid versus unpaid distinction. For unpaid services, including AI Studio and free quota, Google may use prompts and responses to improve products and human reviewers may read them; its terms say, “Do not submit sensitive, confidential, or personal information to the Unpaid Services.” Paid service prompts and responses are not used to improve Google products. Google says paid data-use terms also apply to unpaid use in the European Economic Area, Switzerland and the UK.

Microsoft: the updated Copilot app changed its training wording

Microsoft's updated Copilot app says personal-account prompts, responses and file contents “aren't used to train foundation models.” The updated app became available on 18 August 2026, according to Microsoft's Copilot activity history page.

Our reading is that “foundation models” is narrower wording than “any model.” The page does not establish whether other Microsoft systems learn from those conversations, so the distinction belongs in a company's account decision.

Users who have not updated remain under the older app's FAQ. Under that rule, Microsoft uses signed-in personal-account conversations for training unless the user opts out, subject to listed exceptions, and conversation activity is stored for 18 months by default. Optional feedback in the updated app may be used to improve Copilot, though Microsoft says it is not used to train foundation models.

A work or school account using Microsoft Entra receives enterprise data protection in Copilot Chat. Microsoft says prompts and responses are not used to train foundation models, and IT administrators can view logged prompts through audit tools. A work account therefore changes both the governing terms and who at the company can inspect the records.

For Azure Foundry Models sold by Azure, including Azure OpenAI, Microsoft says prompts are not available to OpenAI and are not used to train generative AI foundation models without permission or instruction. Azure's data privacy documentation says abuse monitoring can send flagged samples for human review unless the customer has been approved for modified abuse monitoring.

Consumer account controls and surviving copies

Consumer account controls govern different uses of data, and none of the four rows makes the delete button a complete inventory of retained copies.

Vendor and consumer productTraining controlWhat can outlive your delete or opt-out
ChatGPT Free/Go/Plus/ProTurn off Improve the model for everyoneFeedback-rated conversations may be used for training after opt-out; temporary chats can be kept up to 30 days for safety; legal or security holds can apply
Claude Free/Pro/MaxChoose at signup or in Privacy SettingsUp to five years in training pipelines if training is allowed; feedback for five years; flagged inputs and outputs up to two years and safety scores up to seven years
Gemini app on a personal Google accountKeep ActivityHuman-reviewed chats can remain up to three years after activity deletion; chats can be kept 72 hours with the setting off
Copilot app on a personal Microsoft accountUpdated app: prompts, responses and file contents are not used to train foundation modelsOn the older app, conversation activity is stored for 18 months by default and training applies unless the user opts out, subject to exceptions

Zero Data Retention requires prior approval from OpenAI or Anthropic; Microsoft's modified abuse monitoring requires eligibility and an application. Anthropic still retains User Safety classifier results under its approved arrangement. The AI readiness guide's retention sheet lists the business and API tiers product by product.

What to copy into your company rule

Our recommendation is to make the account contract visible to staff before company data enters a prompt. The following clauses are our opinion, based on the published terms:

  • Put company data only in accounts governed by business or API terms, and name the approved plans.
  • Have an admin switch off feedback buttons where the vendor provides a control; otherwise, do not rate chats containing company data.
  • Treat a deleted chat as unconfirmed deletion, with a stated retention window and possible legal holds.
  • Treat Gemini or Copilot used through a personal account as consumer use, with no company data.
  • Record Zero Data Retention only when the vendor has approved it in writing, along with the approval date.

The AI governance and risk guide covers the full AI policy, data classes and vendor questions.

Iron Goo's operations service builds and runs AI automations and chatbots on API or business accounts in the company's own name. This is a reading of vendor pages published and read on 1 November 2026, not legal advice, and contracts signed with a vendor can differ.

Free ebook

Put your next steps in order

Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.

By signing up and confirming your address, you consent to receive the free ebook plus one email every two weeks with 3 or 4 new Iron Goo videos, guides and articles, and nothing else; unsubscribe any time with one click.

More posts