AI Safety
· 7 min read

Is uploading a contract to ChatGPT safe? The clause to check

By Amine Gaaliche · drafted with Claude, reviewed and edited by hand

For a template, or a draft with no counterparty details in it, yes. For a signed or received agreement with a real counterparty, usually not as it stands. The problem is the contract itself, which very likely forbids it.

The clause you may breach

Find the confidentiality section. Most commercial contracts, NDAs, employment and settlement agreements restrict who may see them, and often whether anyone may learn they exist. The people allowed to see them are usually employees, professional advisers and regulators.

Three things follow, even if the AI provider does nothing wrong.

  • Uploading is disclosure. Sending a copy to an outside service is a transfer to a third party, even if the recipient is a machine.
  • An AI provider is not your adviser. The adviser exception covers your lawyer and accountant, who owe you their own duties. A software vendor owes you none.
  • The breach is complete on upload. No leak is needed. The disclosure is the breach, which is why nothing seems to go wrong.

This is what makes contracts different. With a medical letter, you weigh a privacy risk. With a contract, you may breach the agreement you wanted help to understand. The other party only needs to know you uploaded it.

What a contract contains

A signed agreement holds data you would never publish:

  • Company names, entity numbers and places of incorporation.
  • Signatories, their titles, and often a direct email and phone number.
  • Rates, discounts, minimum commitments and termination fees.
  • Bank account and routing numbers in the payment schedule.
  • Service addresses and signing dates, which can identify the parties even without names.

Names, addresses and dates of real people are personal data, whatever the confidentiality clause says. So data protection law applies alongside the contract.

In the United States, the usual limits are the agreement itself and professional duties such as state bar rules. Health data adds HIPAA.

If the people in the file are in Europe or the UK, the GDPR applies too. You become the controller, and the provider becomes your processor. That normally requires a lawful basis and a written processing agreement.

Turning off training does not settle it

Switching off chat history or model training is worth doing. OpenAI says these settings control whether your conversations are used to train its models. They govern what happens after the file arrives. The disclosure happened the moment you sent it, and the copy now sits on systems you do not run.

Keep the terms, remove the identities

Almost nothing you want from a model depends on who the parties are. Obligations, unusual clauses and termination terms are questions about structure, and structure survives without names.

What the model needsWhat it never needs
The clauses, defined terms and cross-referencesThe parties' registered legal names
Payment structure, triggers and termination termsBank accounts and payment instructions
Dates, durations, notice periods and deadlinesSignatory names, titles and service addresses

Reversible redaction replaces each identifying value with a code: [P0ABCD] for a person, [O1WXYZ] for an organisation, [S2KLMN] for a Social Security number. The first letter shows the type, and its length follows the original's, with a minimum of six characters. The model reads a coherent contract and reasons about it normally.

Inside one folder, a value keeps its code. The buyer in section 4 stays the same party in section 19. Keep a matter's documents in one folder. A different folder uses different codes, by design.

Keeping a map back to the real values is pseudonymisation, not anonymisation. Under the European regime, pseudonymised data is still personal data. That is intentional: you need the originals back. Keep the map on your machine and know whether your data is pseudonymised or anonymised.

A workflow that works

  1. Encode on your machine. Names, entities, addresses, bank details and dates each become a reversible code before anything is sent.
  2. Review what it found. No local detector finds every identifier, and one missed name is enough. The file stays locked until you confirm what it found.
  3. Ask your questions. The coded text supports summaries, clause checks and comparisons.
  4. Decode the answer. The codes switch back to the real names on your own device.

promptShield keeps detection, replacement and the code registry on your machine. Disconnect from the network and the app still processes the document. The registry is a local file, deliberately not encrypted: the unencrypted contract is already on the same disk.

Frequently asked questions

What if the contract has no confidentiality clause?

Then the contractual bar may not apply. The signatory names and addresses are still personal data. Confidentiality may also be owed under an NDA covering the negotiation, an employer policy or a professional duty.

Is it safer to paste a single clause instead of the whole contract?

Only slightly. A single clause often names both parties, and the notice and payment clauses name them most clearly. An extract with names, company numbers or bank details is the same disclosure in fewer words.

I have already uploaded a contract. What should I do now?

You cannot take it back, so treat it as an incident. Note what was sent and when, delete the conversation, and check the agreement for a duty to report it. Tell the person who manages that client or partner, and strip the identities from the next contract before it goes out.

How can I use AI on a contract without disclosing it?

Replace the identifying details before the text leaves your machine, then restore them with reversible redaction. Within a folder, each party keeps one code, so the model keeps track of the same parties. A signed agreement is often a scan, which needs text recognition first.

In short

Is it safe to upload your contract to ChatGPT? The answer is usually in the contract, and it is usually no. The impulse makes sense: contracts are hard to read and models explain them well. Take the identities out before the document leaves your machine, and you keep the analysis.

promptShield is free to try on Windows. Download it here, or read the general case on pasting documents into ChatGPT.

Found this useful? Let us know.

Share