Privacy and Data Controls in ChatGPT: Model Training, Temporary Chat, Memory and Business Plans

Version note: This post describes model training, retention, Temporary Chat, memory and business plan settings in ChatGPT as of 15 November 2026; the details were checked the same day in the Help Center's Data controls, Temporary chat, Chat and file retention and ChatGPT Voice articles (help.openai.com) and on the ChatGPT Work cloud security, Space and dots pages of the ChatGPT documentation (learn.chatgpt.com). This post is based on the official documentation. The KVKK section suggests a working practice and is not legal advice; older and newer versions of the Help Center articles differ slightly in wording, and the post points this out.

Guess First

Gül works in customer service at Bookstore, the online bookstore used as the shared example across this site's series. She prepares reply drafts in her own personal Plus account and pastes incoming complaint e-mails as they are: the customer's name, phone number, address and order number. Last week a friend told her about the "Improve the model for everyone" setting and she turned it off; now she thinks her old chats are out of training too. For a sensitive complaint she opened a Temporary Chat, because she believes nothing is stored in a Temporary Chat. Hakan, on the other hand, prepares institutional customers' quotes in Bookstore's Business workspace.

What is your guess? Which of Gül's two assumptions is right? Is Hakan's data handled differently from Gül's? The answer is in the Help Center's data controls article: the setting only affects new chats, a copy of a Temporary Chat may be kept for up to 30 days for safety, and the Business default differs from personal plans.

Four Separate Settings

Privacy-related settings in ChatGPT are not a single switch, and they do not replace each other. In the third post we looked at memory and Temporary Chat from the personalization side; here we look at the same settings from the data side:

Model Training: "Improve the Model for Everyone"

According to the Help Center article "Data controls in ChatGPT" (https://help.openai.com/en/articles/7730893, 15 November 2026), this setting is on by default in personal workspaces such as Free, Plus and Pro. On the web you turn it off under Settings > Data controls; on mobile you open Data controls from the profile icon in the sidebar; when you are signed in, the setting belongs to your account and applies on the web and mobile together. If you use Codex on a personal plan, the same setting covers your Codex tasks.

Gül's first assumption is wrong here. The article says that once the setting is off, new conversations will not be used to improve the models; there is no statement that content already shared is taken back. Turning it off does not delete or hide saved chats either. Two exceptions matter. If you give feedback on a response, such as a thumbs up or down, the entire conversation may be used to train OpenAI's models even when the setting is off. In Daybreak sessions on a personal account, which are security work under Trusted Access, the content of interactions may be used to improve models even with the setting off. Parent or guardian settings and workspace controls can limit whether you can change it.

Voice has a separate layer. As we saw in the ninth post, with the setting on, transcripts of voice conversations may be used for training; audio and video recordings are included only if you also turn on "Include your audio recordings" and "Include your video recordings" under Data controls.

Deleting, Retention and Temporary Chat

According to "Chat and file retention in ChatGPT" (https://help.openai.com/en/articles/8983778), a deleted chat leaves your account view right away and is scheduled for permanent deletion from OpenAI's systems within 30 days. There are two exceptions: the chat was already de-identified and disassociated from your account, or OpenAI must keep it for security or legal reasons. A deleted chat cannot be restored through ChatGPT or support. Files live separately: a file saved to Library or Space is not deleted when the chat is, so you need to delete it too. As we saw in the third post, saved memories are also kept separately from chats.

Gül's second assumption is not quite right either. According to "Temporary chat in ChatGPT" (https://help.openai.com/en/articles/8914046), a Temporary Chat does not appear in history, does not use or create memories and is not used to improve models while it remains temporary; but OpenAI may keep a copy for up to 30 days for safety purposes. The data controls article says the same thing more absolutely, "deleted after 30 days"; the difference between the two wordings is small but matters. If you save the chat, it becomes a regular chat and your account's model improvement setting applies. A Temporary Chat is not a safe place to paste customer data; the way to reduce personal data is masking.

Business, Enterprise and Edu

Hakan's data is handled with a different default. According to the data controls article and the "ChatGPT Work cloud security" page of the ChatGPT documentation (https://learn.chatgpt.com/docs/enterprise/chatgpt-work-cloud-security), OpenAI does not use data from Business, Enterprise and Edu workspaces to train its models by default; this data is encrypted in transit and at rest. According to the Help Center's Business articles, in Business each member's chat and Codex history is their own; other members do not see it automatically, and admin tools do not automatically grant access to private chats. In Enterprise and Edu, retention can be set at the workspace level, and audit records are available through the Compliance API.

Business plans do not remove the work of protecting data; they only change the default. The Work cloud security page states a few boundaries clearly. A cloud Work task does not automatically inherit the files, browser sessions or company network on your computer; but it can use everything you type into the conversation, upload, put in a project or let it retrieve from a connected app. Deleting a chat, removing a Library file, deleting a memory, disconnecting an app and clearing the cloud browser's data are separate actions; none of them covers the others. On the admin side, internet access for Work's code and shell commands, the cloud browser and plugins can each be restricted; on the user side there is the Settings > Data controls > Work network access setting.

New surfaces bring their own rules too. As we saw in the eighth post, content copied onto a page in Space is visible to everyone who can see the page, and each user's own training and memory settings apply. For dots, ChatGPT's data controls apply to conversations with the dot and the work it carries out; the documentation says OpenAI does not train models directly on the dot's proactive research or its private notes. Connected apps are where data flows out: the Space admin guide reminds you that even read-only queries can disclose data to external services and that pages, files and tool results can contain prompt injections.

KVKK Practice: What Goes In, What Does Not

This section is not legal advice; it is a working practice a software team can apply in daily work. From a KVKK point of view, a customer e-mail pasted into ChatGPT means transferring personal data to a service provider, often abroad. The Data Privacy and KVKK for Developers series post "Cross-Border Transfers and Breach Readiness: Cloud Regions, LLM Providers, Standard Contracts and the 72-Hour Notification" covers the legal basis and how to assess transfers to LLM providers; make the decision together with your company's legal team. On the developer side three rules are a solid enough start: do not process company data in a personal account, never send personal data that is not needed, and mask the rest before sending it.

Masking usually does not affect the quality of the answer; ChatGPT does not need the customer's name or phone number to write a good reply to a refund complaint. The text Gül pastes should look like this:

Before (do not paste):
  From: Ada Lovelace <ada.lovelace@example.com>, +90 5XX XXX XX XX
  Order BK-2026-104233, delivered to 12 Sample Street, Izmir.
  "Two of the three books in my order arrived damaged. I want a refund."

After (paste this):
  From: [CUSTOMER]
  Order [ORDER_ID], delivered to [CITY].
  "Two of the three books in my order arrived damaged. I want a refund."
  Context: order total 1,250 TRY, delivered 3 days ago, first complaint.

Once the reply draft arrives, you add the customer's name and order number in Bookstore's own support system; the personal data never goes to ChatGPT. The same rule applies to screenshots, exported CSV files and log files: look at what is inside a file before uploading it.

Bookstore: A One-Page Rule List for the Support Team

For Gül and her team at Bookstore we wrote the following list. It is not a policy but a list of habits; once the company's legal team approves it, it goes at the top of the support wiki:

• Use only Bookstore's Business workspace for work; no company data goes into a personal ChatGPT account.
• Customer names, e-mails, phone numbers, addresses, order numbers and payment details do not go to ChatGPT; use placeholders.
• ID numbers, card details and special categories of personal data such as health information are never uploaded, not even masked.
• Take screenshots in the test environment, or do not upload them without blurring personal data.
• Temporary Chat is not a deletion guarantee; it does not replace masking.
• Before giving feedback on an answer, make sure the chat contains no personal data.
• Do not open pages with customer information in Space; anything copied onto a shared page is visible to everyone who can see it.
• If you share personal data in a chat by mistake, delete the chat, check the related memory and tell your team lead.

In Production

Write the default down. If it is unclear who on the team uses which account, the data is handled with an unclear default too. For a team that started with personal Plus accounts, moving to a Business workspace matters for the admin controls as much as for the training default.

Track setting changes. The versions of Help Center articles differ in wording, and new surfaces (dots, Space, Sites) come with their own data rules. Compare the rule list with the official articles every few months.

Know the limits of deletion. A deleted chat is scheduled for permanent deletion within 30 days, but saved files, memory, exported copies and data sent to connected systems live separately.

Common Pitfalls

Thinking that turning the setting off takes the past out of training too. The setting is for new chats; it is not a retroactive deletion.

Thinking feedback is harmless. Giving a thumbs up or down on an answer can open the whole conversation to training even with the setting off.

Treating Temporary Chat as a private mode. It does not appear in history, but it may be kept for up to 30 days for safety.

Thinking a business plan is enough. In Business, data is not used for training by default; but sending unnecessary personal data is still a transfer.

A Small Experiment

★☆☆ Open Settings > Data controls and note the state of "Improve the model for everyone" and the audio and video recording options in your account.

★★☆ Pick five texts your team pasted into ChatGPT in the last week and mark whether each contains personal data. If it does, rewrite it following the masking example in this post and see whether the answer quality changes.

★★★ Write a one-page ChatGPT data rule for your own team and run it past your legal team: which account, which data types, which surface, and what to do after an accidental share.

The series is nearly over. In the last post we will compare the same jobs with Claude: Projects, skills, connectors, file creation, agentic work and privacy defaults, asking "for which job" instead of "which is better".