By Scott Sumner | AI at Work | Published October 4, 2026

An office worker files a document beside a training machine switched off, illustrating that stopping AI training does not mean a file is no longer stored.

Turning off AI training does not mean your conversation disappears. Before your team pastes customer information into ChatGPT, separate the training setting from storage, memory and workspace access.

Answer: AI training controls determine whether eligible content can help improve models. Retention determines how long content is kept. Turning off training does not delete saved chats, and temporary chats may still be retained for safety purposes.

The practical mistake is treating one reassuring toggle as permission to upload anything. You do not need a company-wide AI ban. You need a clear decision about which information belongs in which account.

Jump ahead

What turning off training actually changes

OpenAI’s data controls documentation says that turning off Improve the model for everyone prevents new conversations from being used to train its models. It also says those conversations can remain in your chat history.

Those are two different promises. One concerns how content is used. The other concerns whether it remains available.

Imagine a marketing manager uploading a customer interview to help write a case study. The training setting is off. That answers one question, but not whether the full interview should be stored in that account, whether the client approved that use, or whether the file contains details the task never needed.

This is a hypothetical example, not a report of a client incident. It is the kind of ordinary decision a useful AI policy should settle before the file reaches the upload box.

There is a further detail in the same documentation: if you choose to provide feedback, such as a thumbs-up or thumbs-down, the conversation associated with that feedback may be used to train models. Do not turn a general opt-out into an unconditional promise that covers every later action.

Why temporary does not mean never stored

OpenAI’s temporary chat documentation says a copy may be kept for up to 30 days for safety purposes. While a chat remains temporary, it is not used to improve models and does not create or update memories.

That is useful. It is not zero retention.

The documentation also distinguishes temporary chats from saved chats. Saving a temporary conversation makes it a regular chat that follows the account’s settings. In the version checked for this article, temporary chats can be personalized or unpersonalized; a personalized temporary chat can use existing memories. Do not assume that the word temporary settles personalization either.

Features and settings change. These statements describe OpenAI’s documentation checked on October 4, 2026, not a newly announced policy change and not a promise about every AI provider.

I would keep four questions separate in a business review:

  • Training: Can this content be used to improve models?
  • Retention: What is kept, for how long, and what happens when it is deleted?
  • Personalization: What existing context can influence the response, and what can become memory?
  • Access: Which people, workspace features or connected services can receive the information?

A favorable answer to one does not automatically answer the others.

Check the workspace, not just the brand

“We use ChatGPT” is not a complete description of a business setup. OpenAI says content from ChatGPT Business, Enterprise, Edu and Healthcare workspaces is not used for model training by default. Its documentation also notes that managed workspaces can have organization-level retention, memory, compliance and access settings.

The important word is workspace. A person can have a personal account and access to an organization’s workspace. Before discussing protection, check where the conversation is actually taking place.

OpenAI’s enterprise privacy page describes its business-data commitments. Read those alongside the settings and terms for the actual product your organization uses. Do not assume a feature available in one plan is available in every other plan.

Nor should the same rule automatically apply to every integration. If an assistant sends information to another service, review that destination too. The question is where the information goes, not just which logo appears over the chat box.

This is distinct from limiting what an AI assistant is allowed to do. Permissions restrict actions and access. Data settings describe other parts of the information’s lifecycle. Both need attention.

Give the team an upload rule they can use

A policy that says “use AI responsibly” leaves the employee to make the difficult decision alone. A short, concrete rule is more useful.

For example, a business could approve public website copy for drafting in its chosen workspace, while requiring a separate review before uploading unpublished client material. That is a proposed operating rule, not a universal standard or legal advice.

Specify the workspace by name. Name the permitted task. Explain what to remove before uploading. Identify the person who can decide when the file falls outside that rule.

For the case-study example, the assistant may only need an approved excerpt about the work and outcome. It may not need the interviewee’s email address, internal staffing discussion or unrelated customer names. Removing unnecessary information makes the task easier to review as well as reducing what you share.

Redaction needs a check. Changing a name to “Client A” does not necessarily remove identifying details elsewhere in the document. And do not upload an unapproved original to the same tool just to ask it to remove sensitive details.

When the goal is to test a workflow rather than produce a final case study, use a clearly fictional sample. You can learn whether the assistant follows the requested structure without giving it a real customer record.

A practical check before the next upload

Pick one recurring AI task, not the entire company. Write down the actual account or workspace, the minimum information the task needs, the current training and retention settings, and whether another service receives anything.

Then compare that description with what employees actually do. A written instruction to use a managed workspace is not much help if a saved browser bookmark opens a personal account.

Keep the review dated and linked to the provider’s documentation. Recheck it when you change plans, add an integration or materially change the kind of data involved. This is a more useful habit than treating a screenshot of one toggle as a permanent approval.

For website and marketing workflows, the same discipline belongs in the implementation brief: explain which customer information is needed and which information should never enter the AI step. If you need help reviewing that workflow, contact Scott Sumner with the task and tools involved, without attaching customer data to the initial inquiry.

Start here: Check the actual workspace and data settings used for one recurring customer-related AI task. Do not stop at the training toggle.

Sources