ChatGPT summarized your contract — but can you prove any of it?

I love a fast draft as much as the next developer. Paste a contract into ChatGPT, ask for a summary, get bullet points in thirty seconds — it feels like cheating in the best way.

Then someone on your team asks a very reasonable follow-up: "Where does it say that?"

And the room gets quiet.

The problem: speed without sources

Generic AI chat is good at sounding confident. It's much worse at giving you a citation you can verify in the original PDF.

You get output like:

• "Termination requires 60 days' notice."
• "Liability is capped at fees paid in the prior 12 months."
• "IP assignment includes work product created during the engagement."

Maybe that's right. Maybe it's a reasonable paraphrase of three different sections. Maybe the model stitched together language from an exhibit and a main body clause that actually conflict.

You don't know. The summary has no receipts.

I learned this the hard way when I used an AI summary to prep for a call and got challenged on a renewal term. I opened the PDF, searched for the phrase the model used, and... it wasn't there. The model had inferred the term from surrounding language. Close, but not something I could defend when money was on the line.

Why this gets dangerous fast

The risk isn't that AI hallucinates every time. The risk is that it hallucinates at the worst possible moment — and you can't tell which bullet points are solid.

A few patterns I keep seeing:

Confident tone, fuzzy grounding. The summary reads like a lawyer wrote it. That makes you trust it more than you should.

Merged clauses. Two sections that don't quite agree get blended into one clean statement.

Missing conditions. "You can terminate for convenience" — true, except there's a cure period, a minimum term, and a specific notice window the summary flattened into one sentence.

Wrong document context. Paste the MSA without the SOW, and the model fills gaps with assumptions.

In a negotiation, email thread, or internal approval, "the AI said so" is not a position anyone respects. You need page, quote, and context.

The fix: treat AI output like a draft, not evidence

My rule now: no bullet goes into a deck, email, or decision doc unless I can verify it in the source document.

That doesn't mean ditching AI. It means changing the workflow:

1. Upload the actual PDF (not a messy copy-paste).
2. Get structured output — scenarios, dates, obligations — not just prose.
3. Click through to the source for every item that matters.
4. Flag anything that can't be verified.

This is the gap I wanted to close with Geordi Scan. It's built for contract review inside a single document, with source links on the important outputs — page number and quoted passage, so you can confirm it in the PDF.

There's also a consistency pass that flags contradictions and unsupported claims. Still advisory, still not legal advice — but a useful second check before you trust the summary.

Want to see the difference? Compare a generic chat summary with the freelancer agreement sample. Click an obligation or date and watch it jump to the highlighted source. That's the workflow I wish I had the first time someone asked, "Where does it say that?"

What I still use chat for

I still use general-purpose AI for brainstorming questions, drafting emails, and explaining clauses in plain language after I've verified the source.

The order matters:

1. Verify in the document.
2. Then summarize for humans.

Not the other way around.

---

Next steps: Take a contract you've already summarized with a chat tool. Pick three claims from that summary and try to find the exact supporting language in the PDF. If any claim takes more than a minute to locate, that's your signal to switch to a source-grounded workflow.

Thanks for reading!