How to Summarise a Contract With AI, and What It Will Quietly Leave Out

Ask ChatGPT or Claude to summarise a contract and you will get an answer in about ten seconds: parties, term, a paragraph on payment, a paragraph on termination, done. It reads clean and confident. That confidence is exactly the problem. An AI summary is not wrong the way a bad photocopy is wrong, missing a page you can see is missing. It is wrong the way a very good student's exam answer is wrong: fluent, plausible, and quietly leaving out the one clause that decides who wins if this ever goes to a dispute. This guide is about getting a genuinely useful AI summary and knowing exactly where it stops being trustworthy. (This is published by Adira, which makes contract review and CLM software, so we have a commercial stake in you trusting a bare AI summary a little less, not more. The steps below work with any AI tool, paid or free, ours or not.)

Write a prompt that asks for specifics, not "summarise this"

"Summarise this contract" is a bad prompt because it lets the AI choose what matters. It will usually pick parties, term, and a payment figure because those are the easiest facts to lift and sound the most complete. It will often skip liability caps, indemnity, and auto-renewal, exactly the clauses most likely to hurt you, because they are harder to state cleanly and the model is optimising for a readable answer, not a risk report.

A good prompt names the categories you want and forces a structure. Ask for, at minimum:

• Parties and their roles (who is the buyer, seller, employer, licensor)
• Term: start date, end date, and how renewal works
• Payment: amount, currency, due dates, and what happens if payment is late
• Termination: who can end it, on what grounds, and with how much notice
• Liability: is there a cap, and what is it
• The three riskiest clauses for your side, specifically
• Anything unusual: a clause you would not expect in a standard contract of this type

Then add one instruction that changes everything: tell it to quote the exact sentence behind every point, and to say "not stated" rather than guess when something is missing. That single line is what turns a vague paraphrase into something you can actually check against the document.

Chunking a long contract without losing the thread

Most AI chat tools can technically accept a long document, but "accepts" and "reads carefully" are not the same thing. Contracts over roughly 30 to 40 pages, or contracts with dense schedules and annexures, are where summaries quietly get thinner: the model's attention to any single sentence does not stay even across a 60-page PDF. The risk is not that it refuses the document. It is that it answers just as confidently on page 45 as it did on page 3, while actually having weighted page 3 more heavily.

Three ways to handle this:

1. Split by structure, not by page count. Break the contract at its natural sections, definitions, commercial terms, boilerplate, rather than every 10 pages, which can cut a clause in half.
2. Keep the definitions section in every chunk. A term like "Confidential Information" or "Losses" is defined once and used everywhere. If you paste chunk 4 without the definitions from page 2, the AI is now guessing what the defined term means.
3. Ask explicitly what it could not check. After summarising a chunk, ask: "which cross-references in this section point to clauses you have not seen?" A contract that says "subject to Clause 14.3" in Clause 6 needs both clauses read together, and a chunked summary can easily miss that.

Cross-check the summary against the document, not against your memory

Once you have a structured summary, the actual verification step is simple and takes five minutes: open the source document and check the AI's most important claims against the real text, not against how the summary made you feel. Check the liability cap number, the termination notice period, and any date first. These are the most common places an AI summary quietly softens a number ("liability is limited" instead of quoting the actual multiple) or picks up the signature date instead of the effective date.

Bad prompt → better prompt

Bad: "Summarise this contract and tell me if it's okay to sign."

What is wrong: no structure, no requested categories, no demand for exact quotes, and "is it okay to sign" is a judgement call the AI is not positioned to make responsibly, it will usually answer anyway, which is worse than declining.

Better: "You are reviewing this contract for [my role, e.g. the vendor]. Answer using this exact structure, and for every point, quote the exact sentence from the document in quotation marks, with the clause number if there is one. If something is not stated in the document, say 'Not stated,' do not guess or infer.

1. Parties: full legal names and roles.
2. Term: start date, end date, and renewal mechanism.
3. Payment: amount, due dates, and late payment consequence.
4. Termination: grounds for each party, and notice required.
5. Liability: is there a cap? Quote the exact number or formula, or say none is stated.
6. The three riskiest clauses for my role, with the exact sentence and why it is risky.
7. Anything unusual for a contract of this type.

Go straight to this structure. Do not summarise in prose first."

What changed: the AI has a role to reason from, a fixed structure it cannot skip parts of, and a standing instruction to cite rather than paraphrase and to admit gaps instead of filling them.

Why AI summaries are confidently wrong, the honesty core

Three failure modes hide behind one fluent paragraph, and they are not the same problem:

Selective summarising. The model is trained to produce a helpful, readable answer, not an exhaustive one. Given a 12-page services agreement, it will usually surface the clauses that are easy to summarise cleanly (term, payment, parties) and compress or drop the ones that are genuinely complicated to state in one sentence, an indemnity clause with a carve-out and a cap that only applies to certain claims, for instance. The summary is not lying. It is doing exactly what "summarise" asked for: making something shorter, which always means leaving things out, and the model does not reliably leave out the least important things.

Missed clauses. A buried auto-renewal ("this Agreement shall automatically renew for successive one-year terms unless either party gives 90 days' written notice") sitting in a boilerplate section near the end is exactly the kind of sentence a general-purpose summary skips, because nothing about its position flags it as important. The same goes for an indemnity clause with no stated cap, sitting quietly among other warranty language.

Hallucination. The sharpest failure. An AI model can state a fact, a number, a clause, an obligation, that sounds completely plausible and is not actually in the document. This happens more when you ask a leading question ("what's the liability cap?" on a contract with none can produce a confidently invented number) or when the model fills gaps with what a typical contract of this type usually says, rather than what this one actually says.

The test that catches most of this: make it quote the sentence

Here is a test you can run in under a minute on any AI summary, on any clause you are unsure about: ask the AI to "quote the exact sentence in this document that says X." Then open the document and Ctrl+F for a distinctive phrase from that quote.

Two outcomes matter:

• It cannot produce a quote, or hedges ("the document generally suggests…"). Treat the original claim as unverified. Do not act on it.
• It produces a quote that does not appear verbatim in the document, close but not exact, or plausible-sounding but invented. This is a hallucinated citation, and it is worse than no answer, because it looks like proof.

Only a quote you can find, word for word, in the actual document counts as verified. If it cannot survive this test, do not trust the summary on that point, full stop.

Red flags in an AI-generated contract summary

Normal	Red flag	Why it matters
Summary cites a clause number for each point ("Clause 14.2, Termination")	No clause numbers or headings anywhere in the summary	You cannot go back and check the real wording fast, so you are trusting the paraphrase blind
AI says "not stated" when a term genuinely is not in the document	AI gives a specific number or date for something the document is silent on	This is the classic shape of a hallucination: confident, specific, and invented
Indemnity and liability get separate, detailed lines	"Standard indemnity and liability provisions apply"	This exact phrasing is a template answer, not a read of your specific document, and it is where real exposure hides
Renewal terms are stated explicitly, even to say "no auto-renewal found"	Summary is silent on renewal and jumps straight to a risk score	Silence gets read as "nothing here," when it usually just means the extraction missed it
Summary length and detail scale with document length	A 50-page contract gets the same three-paragraph summary as a 5-page one	Long documents get chunked or truncated quietly; a short answer for a long contract is a warning sign
AI can quote the exact source sentence on request	AI paraphrases again, or says "the document broadly indicates," when asked to quote	If it cannot quote it, it may not have actually found it

The privacy problem: what happens to the contract you paste in

Before any of the above matters, ask a separate question: should this contract go into a public AI tool at all?

Two Indian legal angles make this more than a comfort question. First, most commercial contracts contain personal data: signatory names, email addresses, sometimes salary or bank details in an employment or vendor annexure. Under the Digital Personal Data Protection Act, 2023, if your company decides why and how that data is processed, it is a "Data Fiduciary" for it, and Section 8(2) requires that a processor be engaged "only under a valid contract":

"A Data Fiduciary may engage, appoint, use or otherwise involve a Data Processor to process personal data on its behalf for any activity related to offering of goods or services to Data Principals only under a valid contract." Source: Section 8, Digital Personal Data Protection Act, 2023

Pasting a document with someone else's personal data into a consumer AI account you have no data processing agreement with sits awkwardly against that requirement. Second, for contracts carrying sensitive personal data, financial details, health information, passwords, biometric data, the older Information Technology Act, 2000 still applies in parallel until the DPDP Act's transitional provisions fully take effect. Section 43A makes a body corporate liable for negligent handling of such data:

"Where a body corporate, possessing, dealing or handling any sensitive personal data or information in a computer resource which it owns, controls or operates, is negligent in implementing and maintaining reasonable security practices and procedures... such body corporate shall be liable to pay damages by way of compensation to the person so affected." Source: Section 43A, Information Technology Act, 2000, read with the IT (Reasonable Security Practices and Sensitive Personal Data or Information) Rules, 2011.

There is a third angle that has nothing to do with data protection statutes: the confidentiality clause in the contract you are about to paste in. Many NDAs and commercial agreements define disclosure to any third party, including a cloud AI service, as a breach unless the recipient is a named category (advisors, affiliates) or consent is obtained. Summarising your own draft is usually fine. Summarising a signed agreement with a live confidentiality clause, without checking that clause first, can itself be the breach you are trying to avoid.

Practically: check your AI tool's data controls before pasting anything confidential. OpenAI, for one, lets you turn off "use my content to train models" and offers a Temporary Chat mode that is not retained for training (see OpenAI's Data Controls FAQ), and most major providers offer something similar, but check the current setting rather than assuming. For a genuinely confidential contract, the safer route is one where the document never leaves your machine. You can mark up and flag clauses locally, in your browser, without uploading anything, using Weave, before deciding whether you need a cloud AI summary at all.

For a fuller answer on whether uploading is safe for a specific tool, see Is It Safe to Upload a Contract to AI?. For the deeper review workflow once you have a summary you trust, see How to Review a Contract With AI.

FAQ

Is it safe to paste a contract into ChatGPT? It depends on the contract and the tool's data settings, not on ChatGPT as a category. Check whether training is switched off, use a temporary or non-retained session for anything confidential, and check your own confidentiality clause before you paste a signed agreement. See Is It Safe to Upload a Contract to AI? for the fuller answer.

Can an AI summary miss a clause that is actually in the document? Yes. This is common with long contracts, dense boilerplate, and clauses phrased in unusual language. An auto-renewal buried in a "Term" section near the end, or an indemnity with no cap sitting among general warranties, are the two most commonly missed.

Can AI invent a clause that isn't there? Yes, this is hallucination, and it gets worse, not better, when you ask a leading question about a term the contract does not actually address. Always run the "quote the exact sentence" test on any specific number or clause you plan to rely on.

Does asking the AI to quote the exact sentence always work? It is the single best test available, but it is not foolproof; a model can occasionally produce a quote that sounds exact and is not. The test only works if you actually check the quote against the document with Ctrl+F, not if you just trust that it said "quote" this time.

How long can a contract be before AI summaries get unreliable? There is no fixed page count, it depends on the tool and how dense the document is, but summaries of documents over roughly 30 to 40 pages, or documents with heavy cross-referencing between sections, are meaningfully more likely to miss or blur something. Chunk by structure rather than trusting a single pass over the whole thing.

Should an AI summary replace having a lawyer look at a contract? No. Treat it as fast triage, a way to know what to ask about and where to look closely, not as the review itself. For anything with real money or real risk attached, a summary is the start of the process, not the end of it.

This guide gets you to a usable, checkable AI summary and shows you where that summary quietly stops being reliable. It does not tell you whether a specific clause in your contract is enforceable, or whether it is safe to sign, that depends on facts and drafting a lawyer needs to see, and none of this is legal advice. Talk to a lawyer before you rely on an AI summary for a decision that matters.