Conversations and code you enter into ChatGPT or Codex are used to train models by default on Personal plans (Free, Plus, Pro). To stop this, go to ChatGPT's "Settings" → "Data controls" and turn off "Improve the model for everyone." Even with this turned off, the basic rule is not to enter valid API keys, passwords, or customer information. If you have entered a key, replace it.

Conversations and tasks

Stop it with one ChatGPT setting

The ChatGPT setting "Improve the model for everyone" also applies to Codex on Personal plans. It covers new conversations and tasks after you turn it off.

Codex environments

A separate setting exists

Officially, there is a separate setting called Include environments. Turning off conversations and tasks does not change this setting.

Storage and human review

Training off ≠ all data deleted

Storage and viewing for safety and support are separate. Stopping training and deleting stored data are separate actions.

Sources: OpenAI, Data controls in ChatGPT, How OpenAI handles data in consumer services. Checked October 4, 2026.

Go to the section you want to read (table of contents)

1. Default settings and plan differences

Default training settings differ between Personal plans and business contracts. Accounts that you turned off yourself stay off. Official: Keep your history on and stop training.

Personal Free, Plus, Pro

Data sharing is on by default

You can opt out in ChatGPT's data controls. Paid Pro is treated the same way.

On Personal plans, Codex conversations and tasks follow the same ChatGPT setting. Training on environments is a separate setting.

Business, Enterprise, Edu / API

Not used for training by default

Check the workspace you belong to and your organization's storage and access policies. For the API, also check whether your organization has enabled data sharing.

Not training on data and not processing or storing it are different things.

For the API, an organization owner can enable data sharing. Official: Use of each service for training.

2. Where to turn off training (ChatGPT and Codex)

For Codex on Personal plans, you also stop training in the ChatGPT settings of the same OpenAI account you use to sign in to Codex. If you use different accounts for ChatGPT in the browser and the Codex app, set each one separately.

  1. Check your account Confirm that the accounts and workspaces you use to sign in to ChatGPT and Codex are the same.
  2. Open settings Go to ChatGPT's "Settings" → "Data controls."
  3. Check the toggle Check the status of "Improve the model for everyone."
  4. Turn it off To stop training, turn it off and save. If it's already off, no change is needed.

Scope of this setting

Same account → applies across devices

ChatGPT New conversations after opting out
Codex on Personal plans New tasks after opting out

Settings for other accounts do not change. Include environments and the exception for ratings you submit (👍 and 👎) are handled separately.

This choice is tied to the account you are signed in to and applies across devices. You can also stop training on conversations and Codex tasks with "Do not train on my content" in the Privacy Portal (you don't need to do it in both places). Official: Settings steps and how they apply to Codex.

Codex environment settings (Include environments)

Include environments is a separate setting that decides whether Codex environments may be used for training, and the official help explains it (Official: Separate Codex setting). However, when we checked the screen of a personal Pro account on October 4, 2026, the official link led to the "legacy Codex Cloud" page, and this switch was not shown (Official: Cloud environments). Not seeing the switch, or having zero cloud environments, does not prove that training is off. If you are concerned, ask official support where the current setting is and what its status is on your account.

3. What not to enter, and ways to make inputs smaller

Even with training turned off, the content you send is processed and stored on OpenAI's servers. Valid keys, passwords, and private keys, customers' personal information, and contractually confidential information should basically not be entered, regardless of settings.

Your computer

Commands and file operations

Some processing runs locally.

Cloud models

Sending necessary code and context

Requests to the model are separate communication. Running locally alone does not mean nothing leaves your device.

The official security documentation also distinguishes between network controls for commands and the client's requests for models and authentication. Official: Approvals, security, and communication boundaries, Ways to use Codex.

Example of making it smaller before asking

Production data → reproduce with two dummy records

Storage bug Replace a real user's record with dummy input.
API integration bug Hide the key's value and describe the problem by error type, permissions, and configuration item name.

For design consultations, you can often still get useful judgment by replacing customer names, internal URLs, and unpublished figures.

  • What to give the conversation: the code needed to reproduce the problem, dummy input, and the expected result. Leave out real names and valid secrets.
  • What to put in the development environment: don't let it handle production credentials or private documents beyond what the task needs.
  • What to send outside: check each of the public repositories, shared links, plugins, and log submission destinations.
  • What contracts decide: follow your organization's usage policies and your agreements with customers. Turning off training does not make prohibited external sharing allowed.

For concrete examples, see Precautions when entering information into AI.

4. What to do if you entered sensitive information

First, decide how urgent this is based on what you entered. The top priority is any key or password that is still valid.

  • Valid keys, passwords, private keysCan directly lead to operations, billing, or data access
    Check usage records and permissions, then revoke or replace the affected credentials.
  • Customer information, confidential contract detailsImpact on individuals and organizations, and information management obligations
    Organize what was sent, and respond according to your contract and your organization's contact point.
  • Unpublished source code, plansImpact if design and business information becomes known
    Check training settings and the scope of sharing, and separate the data you use for future development to the necessary scope.
  • Family members' names, addresses, health, schedulesIdentification of a person, privacy of private life and third parties
    Check conversations, files, memory, and sharing destinations separately.
  • Published code, dummy samplesAdditional exposure is relatively small
    Check that no real secrets or personal information have gotten mixed into the samples.

Even with the same key, the urgency differs between one that has already been revoked and one that still has broad permissions. Even short fragments can identify a person when combined with a name, workplace, and schedule.

Credentials are not protected just by turning off settings

If you entered a key, don't paste it into the AI again. Instead, check the permissions, usage records, and expiration in the service. OpenAI also advises replacing API keys that may have leaked and checking usage. Official: API key safety best practices.

For a normal replacement, you can limit the impact of the stop in this order:

  1. Create a new key Prepare it with the required permissions and expiration.
  2. Update where it's used Change the reference in your app or service.
  3. Verify it works Check that it works with the new key.
  4. Revoke the old key Stop use of the old key.

If unauthorized use has already happened, first stop the damage by revoking the key and suspending permissions. OWASP: Revoking and renewing secrets.

Record the scope of what was entered and any signs of harm separately

  • Scope of input What you provided, to which service and account, and roughly when
  • Settings record When you checked, and whether each setting was on, off, or unknown
  • Signs of harm Any suspicious activity, usage, or charges

You do not need to copy the secret itself into a new consultation message or a public comment. When you contact support or the Privacy Portal, ask what procedures apply and what measures can be taken. Deleting content on request has not been publicly stated to remove anything from a model that has already been used for training. Official: Privacy policy and request contacts.

5. How dangerous is it if training was on?

Having the training setting on does not by itself mean that "every conversation has been trained on" or that anything "leaked to third parties." A practical way to assess the risk is by the impact if the content you entered became known to others (see the table in section 4).

Human review of content is officially permitted

Access, to the extent needed, is limited to a small number of authorized staff and service providers bound by confidentiality. OpenAI says access is controlled and logged, but that does not mean no one ever looks at the content.

If you opt out of training

Excluded from use for model improvement

Opting out only stops the use of your content to improve models.

Purposes that still apply

Safety, support, and legal compliance

Necessary content review, for example for abuse or security investigations, does not go away.

Official: Purposes and limits of human access.

Privacy protections exist for removing personal information, but they are not complete

Role of Privacy Filter

A tool that identifies and masks ≠ a guarantee that all sensitive information is removed

What it targets Names, addresses, email addresses, API keys, and so on.
Remaining limits It can miss things because of context or language. Unpublished designs and business plans are not all removed.

OpenAI has stated that, when processing conversations where training is enabled, it uses an internal version of Privacy Filter, a tool that identifies and masks personal information. Official: How ChatGPT protects privacy, Official: Scope and limits of Privacy Filter.

Research on extracting training data (2023)

A 2023 study showed that some text and code that appear to have come from training data could be extracted from ChatGPT as it was at the time (Original paper: Scalable Extraction of Training Data from (Production) Language Models). However, the matches the researchers confirmed were against text from the public web. This is not evidence that private users' chats were extracted (Researchers' explanation of their verification method).

6. Limits of turning off, deleting, and temporary chats

Ratings (👍 and 👎) have an exception

Even with training off, if you submit a rating, the entire conversation tied to that rating may be used for training. Do not submit ratings for conversations that contain sensitive information.

Official: Feedback after opting out.

Deleting conversations and deleting files and memory are separate

Conversations

Deleted ≠ archived

After deletion, they are removed from the system, usually within 30 days. There are exceptions, such as for safety or legal reasons.

Files

Check the storage location separately

Files in Library are managed separately from conversations. Files may remain even after you delete a conversation.

Memory

May remain even after you delete conversations

Saved memory is information used for personalization. Check it separately in the management screen.

Storage exceptions, where things are placed, and memory references within the same account

When you delete a saved ChatGPT conversation, it is removed from your account immediately and is usually deleted from the system within 30 days. However, there are exceptions, such as when it has already been anonymized or disconnected from your account, or when it is kept for safety or legal reasons. Archiving is not deleting.

Files saved to Library are managed separately from conversations. Files in projects and custom GPTs are also handled according to where they are placed. Official: Retention of conversations and files in ChatGPT.

Memory is a feature for personalizing responses and is separate from training settings. Even if you delete a conversation, saved memories created from it are not deleted automatically. If several people share the same account and someone else's schedule shows up in a response, the first thing to suspect is the history and memory references within that account, not leakage through training. Official: Managing memory and how it differs from training.

Temporary chats change how they are handled once you switch to saving

Stays a temporary chat Not used for training, and not kept in regular history or new memory. It may be kept for up to 30 days for safety purposes.
Saved as a regular conversation After switching, it follows that account's settings.

Even in a temporary chat, if you choose personalization, existing memory and other information may be used. It is not a "mode that sends nothing." Official: How temporary chats are handled.

FAQ

If I'm on Pro, are my conversations and code not used for training?

No. Personal Pro has data sharing enabled by default. This is different from Business, Enterprise, Edu, and API, where training is off by default.

If I turn it off in ChatGPT, is everything in Codex turned off too?

On a Personal plan with the same account, training on new Codex conversations and tasks also stops. However, Include environments is a separate setting, and there is an exception for ratings you submit.

If I can't see Include environments, can I assume it's off?

No. The reason it isn't visible and its default value have not been published. If you are concerned, ask official support.

Once I realize it was on, should I delete all my chats?

First, organize the content and the impact. Deleting everything at once would also lose records you need. Decide separately on future training settings, replacing credentials, and organizing the relevant conversations, files, and memory. Deleting content has not been publicly stated to cancel past training.

Basis of this article: OpenAI's public materials (checked October 4, 2026). From the published policies, we cannot tell whether a specific conversation was actually selected for training.