Using ChatGPT for contract review
Working with documentsLast checked
Short answer
It is genuinely good at three things: telling you what a contract says in plain language, spotting what is missing compared with a typical agreement of that kind, and finding places where the two sides are treated differently. It cannot tell you what is enforceable where you live. Use it to prepare, not to decide.
Before you upload anything
On consumer ChatGPT plans, uploaded content may be used to improve model performance unless you have turned that off in the data controls. Business and Enterprise plans are excluded from that by policy. If the contract is under an NDA or contains personal data, check the setting first, or redact the names and figures before uploading. This is easier to do now than to explain later.
Make sure it read the whole thing
Contracts fail this more than most documents, because the important material tends to sit at the back, in schedules and annexes. Files past 2 million tokens get truncated with no warning.
Before we start: list every numbered section and schedule in this document, in order.
If the list stops short of what you can see in the PDF, it is working from a fragment, and the liability cap you were worried about is probably in the part it never saw.
The first pass: what does it actually say
Summarise this contract in plain language for someone who is not a lawyer. Cover: what each party must do, what it costs, how long it lasts, how it ends, and what happens if either side breaks it. Quote the clause number behind each point.
Clause numbers make everything after this faster, because you can jump straight to anything that looks wrong.
The second pass: what is missing
The highest-value question you can ask about a contract.
What would you expect a [service agreement / NDA / employment contract] to contain that this one does not? For each gap, say what risk it leaves open.
Absent clauses are invisible when you read a document front to back, because there is nothing there to notice. Missing liability caps, no termination for convenience, no data-protection terms, no governing-law clause. The gap is usually more expensive than anything actually written down.
The third pass: who does this favour
Go through this clause by clause and identify anything where the two parties are treated differently. For each one, say which side it favours and why.
Asymmetry is where the negotiating room is. One side can terminate on thirty days and the other on one hundred and eighty. One side's liability is capped and the other's is not. These are easy to miss and easy for a model to find.
Specific things worth asking about
Termination
How can each party end this, with how much notice, and what survives termination?
Money
Every payment obligation: amount, timing, what triggers it, what happens if it is late.
Liability
Is liability capped? For whom, at what, and what is carved out of the cap?
Automatic renewal
Does this renew automatically? When is the last date to give notice, and how must it be given?
That last one catches people constantly. Auto-renewal with a ninety-day notice window is easy to read past and expensive to miss.
Intellectual property
Who owns what is created under this agreement, and who owns anything either party brought into it?
Comparing a contract against your standard terms
Genuinely useful, and it takes two passes rather than one.
Do not upload both and ask for a comparison. Summarise each separately, then paste both summaries back and ask:
Here are summaries of our standard terms and the counterparty's draft. Where does theirs differ materially, and which differences increase our risk?
What not to rely on it for
Four things, and they are the four that matter most.
Enforceability. Whether a clause holds up depends on jurisdiction and case law. It does not know where you are and it is not tracking your courts.
What is normal for your industry. It has a general sense of what contracts contain. It does not know what is standard in your market this year.
What is worth fighting over. Negotiating leverage is a commercial judgement, not a text problem.
The final read. For anything that matters, a lawyer reads it. The point of all this is arriving with a marked-up copy and a specific list of questions instead of a blank stare, which is a real saving in both time and fees.
A sensible order
- Confirm it has the whole document, schedules included.
- Plain-language summary with clause numbers.
- What is missing.
- Where the parties are treated differently.
- Targeted questions on termination, money, liability, renewal, IP.
- Take the flagged list to someone qualified.
Steps one to five take about twenty minutes. Step six goes considerably faster than it otherwise would have.
Common questions
Is it safe to upload a contract to ChatGPT?
Depends on the contract and your plan. On consumer plans, content may be used to improve models unless you turn that off in data controls. Business and Enterprise plans are excluded from that by policy. For anything under an NDA, check your setting first, or redact names and figures before uploading.
Can ChatGPT replace a lawyer?
No. It is good at finding what a document says and what it does not say, which is genuine preparation. It cannot tell you how a clause is treated in your jurisdiction, what is unusual for your industry, or what is worth fighting over. Use it to arrive at the lawyer better prepared.
What does it catch most reliably?
Missing clauses, internal contradictions, and asymmetries between the parties. Those are pattern-matching problems over text, which is what it is good at. Asking what a contract of this type usually contains that this one lacks is consistently the most valuable question.
My contract is too long to upload. What now?
Split it and send the parts in order, telling it up front to hold its answer until the last one arrives. Reviewing a contract from a truncated copy is worse than not reviewing it, because the clause you needed is usually in the schedules at the back.
Keep reading
12 ways to get better answers out of ChatGPT about your documents
12 tips, and the first is a 30 second check, because long files are truncated with no warning. Then how to force quotes and stop invented answers.
How to summarise long documents with ChatGPT
Anything past 2 million tokens is cut off silently, so run the last section test before you read the summary. Then the prompts that stop it coming back bland.
How to upload a large PDF to ChatGPT: 4 ways that work
512 MB is the ceiling, but a 3 MB PDF can still be too long: text caps at 2 million tokens. How to tell truncation from rejection, and the 4 fixes.
How to get ChatGPT to cite pages and sections
Text extraction strips page breaks, so ChatGPT rarely knows the page. Ask it to quote the heading and the exact sentence, then search the PDF for that line.
How to compare two documents with ChatGPT properly
Uploading both files at once blends them into mush. Summarise each in its own chat, then compare the 2 summaries in a third, and attribution stays intact.
Summarising a legal document with ChatGPT
Summarise a contract, judgment or policy clause by clause, never in one sweep. What ChatGPT locates reliably, and why 'no such clause' is never proof.