How to Compare Two Versions of a Document With AI
Short answer: paste both versions into one message, label them clearly, and ask for a numbered list of every difference. On documents that are short and clause-numbered, this works well — I planted six changes in a supply agreement and both tools I tested found all six. What it will not do is tell you which of those six actually matters, and that is the part you cannot outsource.
I wanted a test with a right answer rather than a vague impression, so I wrote a fictional eight-clause wholesale supply agreement for an invented bakery and made a second version with exactly six changes in it: four edited values, one clause deleted, one clause added. Then I pasted both into ChatGPT and Claude with the same instruction and counted what came back.
What was planted, and what came back
| Change planted in version B | ChatGPT | Claude |
|---|---|---|
| Delivery days: Tue/Thu → Mon/Wed/Fri | Found | Found |
| Price per case: $42.00 → $44.50 | Found | Found |
| Payment terms: Net 30 → Net 45 | Found | Found |
| Termination notice: 60 days → 30 days | Found | Found |
| Seasonal price review clause deleted | Found | Found |
| Missed-delivery credit clause added | Found | Found |
| Not planted: clause renumbering | Reported anyway | Reported anyway |
Both found all six. Both also volunteered a seventh item I had not asked for: because the deleted clause shifted the numbering, the shortage-reporting clause moved from number 7 to number 6. That is accurate, and on a real contract it is the kind of thing you would want flagged — but notice that it arrives in the list looking exactly as important as the payment terms changing.
ChatGPT quoted both values for every change, which makes the list longer but means you can check it without going back to the documents. Claude gave the same seven findings in about a third of the words.
The harder test: one changed digit
Finding six changes in a document that obviously changed is not much of a challenge. The realistic version is the document that looks identical and is not, so I ran a second pair: two copies of the same eight clauses where the only difference was the liability insurance figure, $1,000,000 in one and $2,000,000 in the other.
It came back with one item — the insurance figure — and the sentence "That is the only difference." No padding, no invented extras. That is the result you want, and it is the one people most often assume these tools will get wrong.
The instruction that does the work
Most of the quality here comes from three words in the prompt. This is what I used:
Here are two versions of the same short contract. List every difference between them. Number each one. Do not comment on anything that is unchanged.
"Number each one" is what turns a readable paragraph into something countable, and countable is what lets you check the work. "Do not comment on anything that is unchanged" is what stops it padding the answer with a summary of the parts you did not ask about. Without that second sentence you tend to get a polite recap of the whole agreement with the differences buried in it.
Two more things worth adding when the documents are longer:
- Label the versions yourself — VERSION A and VERSION B, in capitals, on their own lines. Left to guess where one document ends and the next begins, it will sometimes treat a heading partway through as the boundary.
- Ask it to quote both values for every change it reports. It costs you nothing and it means you can verify the list against the originals without reading both documents again.
When the documents are too long to paste
Both tools take file uploads, and for a normal contract that is the simplest route — attach both files in one message and give the same instruction. The pasting approach is worth keeping for anything you want to strip first, since editing text in a chat box is easier than editing a file.
For documents that run to many pages, compare a section at a time rather than asking for the whole thing at once, and keep a running list yourself. There is a specific trap with chunking: if a clause was moved from one part of the document to another, the chunk it left shows a deletion and the chunk it arrived in shows an addition, and neither one tells you it is the same clause. If the two versions have different section counts, that is a sign to check for moved text before trusting a section-by-section list.
What a clean result does not tell you
This is the part worth being careful about, and it has nothing to do with whether the tool finds the changes.
A list of seven differences is not a list of seven problems. In my test, "payment terms move from Net 30 to Net 45" is fifteen days of your money, "termination notice drops from 60 days to 30" halves your warning if they walk away, and "the shortage-reporting clause is now number 6" means nothing at all. They arrive in the same font, in the same list, in no particular order of importance. The tool has no idea which of them would cost you money, and it will not rank them unless you ask — and even then it is guessing at your situation.
So treat the output as a complete inventory rather than a verdict. The value is that nothing got past you. Deciding what to do about each item is still yours, and for anything with legal weight it belongs with a person who is qualified to read it.
Two other limits from this test specifically. My documents were short, numbered, and lined up clause for clause — the easy case. I did not test what happens with long unstructured prose, documents that have been reformatted, or clauses that moved rather than changed, and I would not assume the same accuracy there without checking. And a clean run on one pair tells you nothing about the next pair: this is a tool to check your reading against, not a reason to stop reading.
Before you paste a contract anywhere
A real agreement usually has two parties' names on it, and often terms someone expects to stay private. Strip the names and any figure that identifies the deal if you can still do the comparison without them — the tool does not need to know who the parties are to spot that Net 30 became Net 45. If it is under an NDA, the answer is that it does not go in the chat box at all. I went through where that line sits in more detail in the guide on pasting work documents into AI tools.
Where this earns its keep
Redlines you were sent without track changes. A supplier's "minor update" to terms you already agreed. Two drafts of your own where you lost track of which was current. A policy document that got reissued and you need to know what moved. In each case the job is the same: produce the complete list in thirty seconds so that the twenty minutes you spend afterwards go on the three items that matter rather than on hunting for them.
Related reading
- How to Summarize a Long PDF With ChatGPT, Claude or Gemini — for when the problem is length rather than two versions.
- How to Turn Rough Notes Into a Polished, Proofread Document With AI — turning scribbles into a clean document, then proofreading it.
- How to Analyze a Spreadsheet and Find Insights With AI — the same checkable-constraint trick, applied to numbers.
Written by Mitch, a software analyst who tests software for a living. Every guide here comes from actually using the tool on a real task — including the parts where it falls over. Tested on ChatGPT Plus, Claude Max and Gemini (free). More about this site · Corrections: [email protected].
Get the next guide by email
A new tested how-to when there is one — usually a couple a month. No hype, unsubscribe any time.
Tested in real accounts. No affiliate links.
