The generation of lawyers who don't use AI because they won't do what they should know best how to do: read a contract.

View on LinkedIn

«But I can't use it because it's not confidential... or is it? I mean, are they going to use my data to train their models? What guarantees me that the information won't be leaked?»

Laura asks at an AI talk for lawyers.

Laura is a corporate attorney. She reviews contracts every day. She negotiates confidentiality clauses for her clients. She identifies risks in the fine print of agreements with technology vendors. She is, by professional definition, someone who knows how to read a contract.

Laura's concern is valid. It is fair. It is exactly the kind of due diligence a lawyer must exercise before using any tool that touches client information.

But it has something curious about it: it is cured by reading.

Specifically, by reading the Terms and Conditions of ChatGPT, Claude, and Gemini — the contracts that Laura has the perfect tools to interpret. She just forgot that she is not merely a user. She is a lawyer. And if there is one thing an attorney is good for, it is reading and interpreting contracts.

Let me explain, Laura.

I have spent three years giving talks, workshops, and courses on artificial intelligence for lawyers. In addition to my law degree, I studied a degree in software design and development, and I live in Silicon Valley — which placed me, by luck or by timing, at the epicenter of the generative AI explosion right when it began. In three years everything changed: the models, the applications, the prices, the regulation. But your question has not changed. Not at all. Not once.

In social psychology, this has a name. It is called an availability cascade. Someone says something alarming. Another person repeats it. A panelist cites it with an air of authority. And a claim that no one verified becomes accepted truth — not because it is true, but because it is everywhere. "AI keeps your information" became guild dogma. Except no one went to verify the source.

I did read it. What follows are your four questions, Laura, answered with the Terms and Conditions in hand.

Do they use my data to train their models?

No. Not if you use the correct plan or disable the training option. The enterprise and API plans of ChatGPT, Claude, and Gemini are excluded from training by default. Not by the user's choice. By design. OpenAI's commercial terms say it unambiguously: they will not use customer content to develop or improve their services unless the customer explicitly authorizes it. Anthropic and Google confirm the same for their commercial plans.

What about a personal plan? Consumer terms permit training unless the user disables it. In ChatGPT it is under Data Controls. In Claude, under Privacy Settings. It is disabled in seconds. With the toggle off, your data does not train anything. They are not keeping your information, Laura. That is what the contract says.

Can they read my conversations?

On enterprise plans, no. OpenAI's commercial terms state that they will only use customer content to provide the service, comply with the law, and enforce their policies. Google confirms for Workspace that business content is not reviewed by humans or used for training. On consumer plans, the platforms reserve the right to access conversations for security and abuse detection — but these are automated systems. Human reviewers intervene only in exceptional cases. And there is something more fundamental: what you generate with AI as part of your professional practice is work product. It is protected by attorney-client privilege, just like a draft in Word or research in a legal database. The tool does not determine the protection. The nature of the work does.

How long do they keep my information?

With training disabled: thirty days at Anthropic and OpenAI. On OpenAI enterprise plans, content is deleted within thirty days following contract termination. If training is enabled at Anthropic, retention increases to five years — but only for new conversations, and deleting a conversation excludes it from training. Five years sounds alarming out of context. In context — reading the contract — it is an option the user chooses to activate.

What guarantees me that the information won't be leaked?

On an enterprise plan: OpenAI's commercial terms include infrastructure behind VPNs, multi-factor authentication, least-privilege access, independent audits, Data Processing Agreements, and compliance with SOC 2, ISO 27001, and GDPR. On a personal plan with training disabled: thirty days of retention, no training, no sharing. On a personal plan left unconfigured: yes, there is risk. But that risk was created by the lawyer who did not read the contract or configure their tool.

And enterprise plans include intellectual property indemnification clauses. If the output infringes third-party rights, the platform is liable. How many of the other tools your firm already uses offer that?

The 4 questions vs. what each contract says

And if you are thinking about the cases that made the news, read them in full.

In South Korea, police used a 21-year-old woman's ChatGPT history to upgrade charges to premeditated murder. She had asked "what happens if you take sleeping pills with alcohol?" The headline is alarming. But the conversations were obtained through forensic analysis of a seized phone. Not through a request to OpenAI. The same mechanism by which police access WhatsApp history or any app on a confiscated device.

In New York, Judge Rakoff ruled in United States v. Heppner that an executive's conversations with Claude were not protected by attorney-client privilege. It sounds frightening. Until you read the facts: Heppner was not a lawyer. He used Claude on his own, on a personal plan, to research his own criminal case. His attorneys admitted they had never asked him to do so. The judge determined that an AI is not a lawyer, does not hold a license, and that the terms themselves permit disclosure of inputs to authorities.

None of these cases involved a legal professional using the tool within their professional practice with the proper configurations. They are ordinary individuals who gave the chat information relevant to an investigation. The difference between Heppner and a lawyer with an enterprise plan, training disabled, and a signed Data Processing Agreement is the same difference as sending an email from a personal Gmail account versus using a law firm's encrypted system. The tool may be the same. The legal framework is not.

The headlines feed the cascade. Read them in full: none involve lawyers using AI professionally.

What you should not do.

Do not use free or personal plans for work involving client information. Free and personal plans were designed for individual use. If you are a lawyer and you use a Free or Pro plan to review your clients' contracts, you are operating under consumer terms. And consumer terms do not give you the protections you need.

And do not forget to review the training settings. Because the platforms do not tell you in those words. You will not find a button that says "Do you want us to use your data to train our models?" What you will find is something like "Would you like to help improve the model?" or "Help improve Claude." It sounds harmless. It sounds generous.

It is a polite way of saying: you want us to use your conversations to train our models.

Disable it. In ChatGPT, it is under Settings, Data Controls. In Claude, under Privacy Settings. In Gemini, under Gemini Apps Activity. It takes seconds. And if you use the tool for professional work, do it from an enterprise plan. It is not a luxury. It is the digital equivalent of having a locked office instead of working in a coffee shop with an open case file on the table.

Three years.

In three years, technology took an unprecedented leap. But in lawyers' forums, the conversation remains where it started. The same question. The same tone. The same fear built on information that no one went to verify. There is an entire generation of lawyers being left behind. Not for lack of intelligence. Not for lack of opportunity. They are being left behind for not doing what they should know best how to do: read a contract, interpret it, and decide based on what it says — not on what is said about it.

The answer was always right there, Laura. In the kind of document you have spent your entire career learning to read.

You just have to open it.

T&C in full: openai.com/policies · anthropic.com/legal/consumer-terms · support.google.com/gemini