It can. Most NDAs let you use the other side's information only for an agreed purpose and share it only with certain people, and a chat in ChatGPT, Claude or Gemini does not fit either rule easily. Whether your paste was a breach depends on the wording of your NDA, the plan you used and what exactly you pasted. A handful of clauses decide it.
The examples below quote oneNDA, a free standard NDA written and reviewed by lawyers and used by many companies. Your NDA will use different words, but it will almost certainly have the same clauses. How a court reads them depends on the governing law named in the contract.
What counts as Confidential Information
Modern NDAs define it broadly. oneNDA covers any information disclosed by one party to the other "in connection with the Purpose", whether or not it is marked. Some older NDAs protect only information marked "confidential" or confirmed as confidential in writing. If yours is one of them, an unmarked spreadsheet may fall outside it.
Then read the exclusions. Information that is already public, that you knew before, or that you developed on your own is usually not covered. Everything else in the document you received is usually in: prices, volumes, technical details, plans, names of customers. Some NDAs also cover notes and analyses you make from it, so check whether yours does.
What you may use it for
Nearly every NDA names a purpose, such as evaluating a possible deal or preparing a quote, and lets you use the information only for that. oneNDA says the receiver must "only use the Confidential Information for the Purpose".
Asking an AI tool to summarize a target's financials so you can evaluate the deal is arguably use for the purpose. Asking it to turn the same material into a pitch for another client is not. In most cases, though, the trouble is in the sharing clause rather than the use clause.
Who you may share it with
NDAs list who may see the information, usually employees, advisers and sometimes contractors, and add two conditions: they need to know it for the purpose, and they are bound to keep it confidential. In oneNDA, "Permitted Receivers" include "agents and subcontractors", but only if they "have agreed to keep it confidential and restrict its use to the same extent that the Receiver has". The receiver is liable for what they do.
An AI provider processes your text on its own servers. Whether that makes it a third party you disclosed to, or a service provider you were allowed to use, is not settled for NDAs in general. It depends on the wording, and on the terms you accepted with the provider:
- Business plans. Anthropic's Commercial Terms say that inputs and outputs are the customer's Confidential Information and that Anthropic "may not train models on Customer Content". OpenAI's enterprise privacy page says it does not train on data from ChatGPT Business, Enterprise and the API by default, and limits who at OpenAI can access it. Terms like these are close to what an NDA asks of a subcontractor.
- Personal accounts. Free and personal paid plans run under a privacy policy, not a confidentiality agreement with you. Our overview of AI data settings shows what each vendor's help pages say about training and retention as of October 2026.
One US court has already looked at a consumer AI chat and asked whether it was confidential. In United States v. Heppner (S.D.N.Y., February 17, 2026), Judge Jed Rakoff held that a defendant's exchanges with the consumer version of Claude were protected neither by attorney-client privilege nor as work product. One reason: Anthropic's privacy policy said it collects inputs and outputs, uses them to train Claude, and may disclose data to third parties, including governmental regulatory authorities, so the defendant "could have had no 'reasonable expectation of confidentiality'". That was a privilege ruling in a criminal case, not an NDA dispute. It still shows how a court may read consumer terms. Debevoise & Plimpton noted that a court may well take a different view of enterprise tools.
How you must protect it
Most NDAs require you to keep the information secure. A common formula is the same degree of care you use for your own confidential information, and no less than reasonable care. oneNDA says to "keep the Confidential Information secure and confidential".
A simple test: would your company let you paste its own price list or source code into a personal AI account? If its AI policy says no, it will be hard to argue that doing it with someone else's information was reasonable care.
Return or destroy, and your chat history
When the talks end, or when the other side asks, most NDAs require you to return or destroy the information. oneNDA gives thirty days after a request to "take reasonable steps to destroy or erase" it, with exceptions for back-ups and legal retention. Some NDAs require a written certificate that you have done it.
A chat that contains the document is a copy, and so is a file you uploaded. On ChatGPT, deleting a chat removes it from your account at once and schedules it for permanent deletion within 30 days, unless OpenAI must keep it longer for security or legal reasons. Deleting a chat does not delete files saved separately in the Library. Archiving a chat does not delete anything. On business workspaces, admins set how long chats are kept.
If you will have to certify destruction, know where every copy went before you paste, not after.
NDAs that mention AI
Newer NDAs often deal with AI directly, and such a clause overrides everything above. Read it first. The published oneNDA template does not mention AI, but many company templates now do. In a December 2025 article, the law firm Smart & Biggar described the range: a full ban on using AI tools with the other side's information, use only with consent case by case, or use under conditions such as named tools and settings, no model training, inputs and outputs treated as confidential, and deletion rules.
A July 2026 Carta article quotes a typical broad version: "Without the Discloser's consent, no Recipient shall use, input, or incorporate any Information into any artificial intelligence or machine learning models, including but not limited to large language models (such as but not limited to ChatGPT), whether for training, fine-tuning, inference, or any other purpose." Under a clause like that, a paste into any chat, business plan or not, needs consent. The same article reports a move towards clauses that allow enterprise tools and forbid training.
What placeholders fix, and what they do not
Replacing names with placeholders such as Person_1 and Company_1 keeps personal data and the parties' identities out of the chat. When what is confidential is who is involved, such as the fact that two companies are talking, or a list of customers, that does most of the work.
An NDA protects information, not only names. If what is confidential is the deal itself, a placeholder copy still contains it. "Company_1 will pay Company_2 a 4% royalty on net sales from March" is still the royalty rate. A process description with the names removed is still the process. Anyone who knows the market may also guess who Company_1 is from the details.
So placeholders reduce what you expose. They do not, by themselves, make a use permitted that the NDA forbids. Figures, formulas, technical details and plans have to be removed or generalized by hand, or kept out of the chat. Our list of what to remove before a document goes to AI covers the details that identify someone in combination.
How to use AI and stay inside the NDA
- Read the AI clause, if there is one. If it forbids AI use without consent, stop here and ask.
- Ask the other side. A short email often settles it: "We'd like to use [tool], under a business plan that does not train on our data, to summarize and compare the documents you sent under our NDA. Please confirm you have no objection." Keep the reply with the NDA.
- Use a plan with contractual terms. If the NDA lets you use service providers bound by confidentiality, a business plan whose terms say so is far easier to defend than a personal account. Check with whoever owns the contract on your side.
- Keep AI to the parts that are not confidential. Your own draft reply, the structure of a memo, a general question about how earn-outs usually work: none of these needs the other side's document.
- Strip what you send. Take only the passage you need. Replace names and contact details with placeholders, and replace or remove the figures and technical details by hand. ClientSafe does the first part on your PC: it replaces names, companies, addresses, phone numbers, email addresses, account and ID numbers.
This article explains which clauses to read; it is not legal advice. When the wording is unclear, ask the lawyer who drafted or negotiated the NDA. For lawyers' own professional duties, which sit on top of any NDA, see using AI as a lawyer without breaching client confidentiality. For the general workflow with client documents, see how to use ChatGPT, Claude or Gemini with client documents.
Frequently asked questions
Is it a breach if nobody at the AI company ever reads it?
Most NDAs restrict sharing itself, not only the harm that follows. Whether anyone read the chat may matter to what the other side could claim, but it does not decide whether the clause was kept.
Does a business plan make it allowed?
It helps, because its terms usually treat your inputs as confidential and exclude training. It is enough only if your NDA allows service providers on those terms and has no AI clause that says otherwise.
I already pasted the document. What now?
Delete the chat and any uploaded files, and note when you did it. Then read the NDA: some, like oneNDA, require you to tell the other side promptly if you become aware of a breach. Talk to your lawyer before you write to them.
ClientSafe replaces names, companies, addresses and other personal data in documents with placeholders on your PC, and puts the real names back into the AI's answer. It does not decide what your NDA allows. Detection is not perfect, so review the list before you copy.