Is ChatGPT Confidential? What Happens to Legal Info You Type In

Is ChatGPT confidential? The short answer is no, not by default. On ChatGPT Free and on Plus, OpenAI stores your messages as data to train the next model.

You can turn ai training off from a switch in your settings. OpenAI still holds your message log on its servers once you flip the switch. This page walks through the data flow and Rule 1.6, the bar rule that covers a lawyer's duty of care.

Is ChatGPT confidential by default?

ChatGPT is not private in the way a talk with a client is. Your session sits on OpenAI servers. OpenAI keeps the log until you delete it.

The default setting goes by the name "Improve the model for everyone" on a personal plan. On the free tier, Plus, and Pro, the switch is on from the start. You turn it off from the privacy panel in your settings.

OpenAI lays this out in a short FAQ for ChatGPT. The page names the default state and shows where to find the switch.

What OpenAI does with your ChatGPT messages

OpenAI holds every prompt and reply on its servers. On a consumer chatgpt plan, that content can feed the next training round. The FAQ says OpenAI can use the content to improve the model unless you opt out.

You stop that use with the toggle in your settings. The toggle sits in the data controls panel OpenAI built for ChatGPT. You open Settings, open the panel, and flip the switch to stop OpenAI from putting information into ChatGPT training.

Flipping the switch does not delete your past content. The FAQ says saved messages stay in history after the switch flips. Deleted and archived items also linger on OpenAI servers for a stretch.

The FAQ warns about one more pattern. A thumbs-up or thumbs-down on a reply ships the full session as data for training. OpenAI may use your data to improve the chat tool even when the switch is off.

Is a temporary chat really temporary?

A temporary chat works in a slightly different way. The FAQ says it does not sit in your chat history and does not add to memory. OpenAI also pledges not to use it for training.

But OpenAI still keeps the session for up to 30 days for safety reasons. So "temporary" is not the same as "gone the moment you close the tab." A court order can still reach the log in that window.

Can people see what I type into ChatGPT?

OpenAI staff can read parts of your log in a few cases. The FAQ says safety features may draw on a slim slice. Feedback you send pulls the full session to OpenAI for review and model use.

Courts can also reach your log. A legal hold can freeze the record so OpenAI must preserve it. A subpoena can order OpenAI to hand it over to a party or a court.

A chatgpt conversation is not private in the way a meeting with a client is. OpenAI stores the log and staff can read parts of it. A court can still reach it after you flip the switch.

Confidentiality under ABA Model Rule 1.6

ABA Model Rule 1.6 names a lawyer's duty of confidentiality to a client. Part (a) says a lawyer must not disclose client matters from a case, outside a short list of exceptions.

Part (c) of the rule adds a second duty on sensitive information. The lawyer must make "reasonable efforts" to stop unauthorized disclosure of client data.

The rule fits many tools. It covers email, cloud files, and a chat session with an ai tool the same way. The rule asks the lawyer to think about what a tool does with the data.

Is ChatGPT safe for confidential information?

On a free or Plus plan with the default on, no, not for a lawyer's confidential data. OpenAI treats prompts and replies as data used to train the next model, unless you opt out.

With training off on a personal plan, the privacy risks drop but data is still stored. Your prompt log sits on OpenAI servers under the privacy policy. A court can still reach it through a subpoena or a legal hold.

For client work, an enterprise plan is a better fit than a free or paid personal tier. The business workspace and the API skip model training by default. The same default holds for chatgpt enterprise and edu, and for ChatGPT for Healthcare.

ChatGPT and AI attorney client privilege

Data privacy is one issue. AI attorney client privilege is a second one, and the two do not line up neatly. Privilege is a rule of evidence that lets a client refuse to share certain talks with a lawyer in court.

For court rulings on ai attorney client privilege, see our page on the Heppner AI privilege ruling. A sibling piece covers the Assini work product ruling. Those pages lay out what each ruling held and what it left open.

Rule 1.6 is the lawyer's own duty. Privilege is a client's shield in court. The two can both apply to the same chat session, or neither can, depending on the facts and the tool.

Even a paid plan with training off does not make a session privileged. The data is still held on a third party server, and a court can reach it through a hold.

What should a working lawyer check?

The steps below map to what the OpenAI FAQ names. They are general steps for a working lawyer, not a program for one firm. The firm's own privacy policy and Rule 1.6 analysis still govern.

Start with the plan. Use an enterprise plan or the API for client work, not the free tier. Those options skip ai training by default and let a firm cap retention.

Flip the switch. Turn the training setting off on any personal plan tied to firm work. Keep the switch off for any task that touches a client matter or case fact.

Mind what you type. Do not use chatgpt without a plan that fits Rule 1.6. Do not send feedback on a reply that holds client facts, since that click ships the whole session to OpenAI.

What ChatGPT Enterprise and the API give you

The enterprise tier and the API serve firms and teams. The FAQ says OpenAI does not use Business, Enterprise, or API content to train its models by default. The paid plan also unlocks admin controls on top of that default.

A firm can set its own rules on retention and access under the business plan. The admin panel can cap how long records sit, lock who can read them, and turn memory on or off. These settings fit the kind of data handling a firm can describe in a contract to protect client information.

None of this makes a session privileged in court. It does move data practice closer to what Rule 1.6 asks for. A firm still owes a duty of care on every prompt it ships to the tool.

Zero data retention and processing agreements

Zero data retention is a separate option OpenAI offers through a data processing agreement for business accounts. The enterprise privacy page spells out how business data sits on the server. The processing addendum for chatgpt business names the terms of use for a firm that signs it.

A firm that needs tight rules can tie compliance with data protection regulations into the agreement. That path fits a legal team under a bar rule plus a data terms clause to keep client confidentiality. The company behind chatgpt also runs abuse monitoring on the account, so a firm must understand the privacy trade.

Does OpenAI share my data with third parties?

OpenAI can share your data with others in a few named ways. A court order can force disclosure. A staff review can touch the content for safety or feedback.

OpenAI does not sell your data to data brokers. On free and Plus plans, the content feeds training unless you opt out. The FAQ names this path used for training, a form of use, not a sale.

The point holds: your log is not just between you and the model. It also passes through OpenAI and its vendors, and it may feed future model work unless you flip the switch.

The bottom line

ChatGPT is not private by default. OpenAI holds every prompt and reply, and chatgpt may feed training on free and Plus plans. A court can reach the log through a hold even when training is off.

The right plan narrows the risk but does not erase it. The enterprise tier, the API, and the default for ChatGPT Business all skip training by default. A lawyer still owes the Rule 1.6 duty of care on every prompt.

None of this bans generative ai use for a lawyer. It asks for care around privacy and security. Read the FAQ, read the rule, and pick the plan that fits the work.