Using ChatGPT to Review a Contract? Strip Out the Personal Data First

Using ChatGPT to Review a Contract? Strip Out the Personal Data First

It usually starts with a reasonable question. A founder has a contractor agreement due back by Friday, or an HR manager is looking at a non-compete in an offer letter, and nobody on staff has the time to read eight pages of clauses. So the whole document gets pasted into ChatGPT, Claude or Copilot with a simple request: explain section 7 in plain English and tell me what looks unusual.

The tool is good at that job, which is why so many people now do it. In a Pew Research Center survey of 5,119 U.S. adults, conducted in February 2026 and published in June, 38% of employed adults said they use AI chatbots for tasks at work. In the same survey, 71% of Americans said they expect the growing use of AI to make their personal information less secure.

Both numbers describe the same office. The contract that goes into the chat window carries names, home addresses, salaries, bank details and sometimes a Social Security number, along with business terms the other side expects you to keep private. Here is what rides along when you paste a document, what U.S. rules and regulators actually say about it, and a simple routine that lets you keep using AI without handing it the data.

What rides along when you paste a contract into an AI chatbot

A blank template is generic by design. A filled-in one is not. An employment agreement names the employee, lists a home address, a start date and a base salary, and may attach a direct deposit form or benefits paperwork. A commercial lease names the guarantor. A client services agreement can list the client's contact person, billing details and the scope of a project the client has not announced yet.

Very little of that helps the AI answer your question. To explain how a non-compete works, a model does not need the employee's full name, the street she lives on or the salary on page two. It needs the clause, the defined terms and the jurisdiction. Everything else is simply along for the ride, because copying the whole document was faster than editing it.

Confidentiality agreements deserve extra care. An NDA, by definition, is a document about information someone asked you not to share, and the schedule attached to it may describe that information in detail. Pasting the full agreement, exhibits included, into a third party's tool is exactly the kind of disclosure many NDAs limit to employees and advisers with a need to know.

What happens to the text after you hit Enter

The answer depends on which product and which plan you are using, and it is worth reading rather than assuming. OpenAI's help center says that for its services for individuals, such as ChatGPT, it may use your content to train its models, and that users can opt out by turning off Improve the model for everyone under Settings and then Data controls.

The same page says that by default OpenAI does not use inputs or outputs from ChatGPT Business, ChatGPT Enterprise or its API to improve its models. It adds a detail many people miss: if you give feedback with a thumbs up or thumbs down, the entire conversation tied to that feedback may be used for training, even after you have opted out.

Training is only part of the picture. Whatever the setting, the text has left your computer and now sits with another company under that company's terms. In a small business, the person pasting the contract is often working from a personal account whose settings nobody has checked. That gap between company policy and what actually happens at someone's desk is the one worth closing first.

What U.S. privacy rules and regulators say

There is no single federal privacy law that covers every business, but the Federal Trade Commission sets the baseline. Its business guidance is direct: if your company makes privacy promises, expressly or by implication, the FTC Act requires you to live up to them. Even without specific promises, a business must keep security appropriate to the nature of the data it holds.

State law adds detail. California's privacy statute, which applies to businesses above certain revenue or data thresholds, treats a person's Social Security, driver's license, state ID or passport number as sensitive personal information, along with financial account numbers combined with access codes and information about citizenship or immigration status. Onboarding packets and employment files routinely contain several of those at once.

Lawyers have their own rulebook. In Formal Opinion 512, released on July 29, 2024, the American Bar Association's ethics committee said that a lawyer using generative AI must keep in mind the duty to keep all information relating to a representation confidential unless the client gives informed consent. If you send contracts to outside counsel, it is fair to ask how their firm handles that question.

NDAs, trade secrets and the clause you may be breaking

Before you paste a counterparty's document anywhere, reread the confidentiality section you signed. Many NDAs define who may receive the information and require you to protect it with at least the care you give your own secrets. An AI provider is rarely on the list. Whether a particular upload breaches a particular agreement is a question for a lawyer, but it is a question worth asking before the upload, not after.

Your own secrets face a related risk. Under the federal Defend Trade Secrets Act, information qualifies as a trade secret only if its owner has taken reasonable measures to keep it secret. A habit of dropping pricing models, customer lists or product plans into consumer chat tools makes those measures harder to demonstrate if you ever need to enforce your rights in court.

A practical workflow: strip, pseudonymize, then prompt

Start by deciding what the AI actually needs. For most contract questions, that is the clause, the defined terms and the state whose law governs. Then take stock of what else is in the file, as the FTC's guide to protecting personal information puts it, and scale down. Social Security numbers, account numbers, home addresses and salary figures can almost always come out before the text goes anywhere.

Contracts already give you the trick. They refer to "the Company" and "the Contractor" instead of repeating names, and you can do the same with everything else: the employee's name becomes [EMPLOYEE_1], the client becomes [CLIENT_1], the home address becomes [ADDRESS_1]. Because each placeholder stays consistent, the AI can still follow who owes what to whom, and you can swap the real details back into its answer.

Doing that by hand works for one document and falls apart at fifty, which is why tools for pseudonymization for AI exist. Nonimo, for example, is a desktop app for Mac and Windows that detects personal data such as names, Social Security numbers and account numbers and replaces them with placeholders on the computer itself, before the text reaches ChatGPT, Claude or Copilot. The AI does the same work without the real data.

Put the rule in writing

A routine only holds if it is written down. A one-page AI use policy can say which tools and plans staff may use, which documents never go into them in full, and that personal identifiers are removed before anything is pasted. Add a line on who reviews the output, because a confident summary of a contract is still not legal advice.

The goal is not to keep AI away from your contracts. Used well, it helps owners and HR teams understand what they are signing and spot the clause that needs a second look. The habit that makes that safe is small: send the AI the structure of the deal, keep the names and numbers at home, and check the settings of whatever tool your team actually uses.