Contents
Claude and ChatGPT now put links under their answers. It looks like proof. It isn't. A citation gives you somewhere to look. It doesn't settle whether the page supports the answer. ChatGPT's own help pages say search results and citations can be out of date or wrong. I'd treat every tool the same way.
Three things to check at the source
The claim. Open the link and find the actual passage. Does it say what the answer says, or has the meaning been stretched on the way through?
The date. Find when the page was published or last updated. For anything that changes, you want the current official page, not an old announcement a search happened to surface.
The scope. Read the conditions around it. Does it apply to your state, your kind of agreement, your plan, your account type? A rule written for one situation gets quietly applied to another all the time.
Say a property manager asks whether a particular notice period applies and the answer arrives with a link to a government page. Open it. The passage might cover a different kind of agreement. The page might predate a change in the rules. It might say "generally" where the answer said "always". If the page doesn't establish the claim, the claim stays unconfirmed. In my business a notice period is a legal requirement, so that's the difference between a letter we can send and one we can't.
One habit that helps. Ask the AI to separate what the sources confirm from its own interpretation. Then check that separation yourself, because it's also an answer.
Five checks before a draft leaves your desk
Checking sources is one part of a wider review. Before anything AI-assisted goes to an owner, a tenant or a client, I run five checks.
Facts. Open the source and compare the names, dates and claims against it.
Completeness. Does the draft cover everything the brief asked for? Missing things are harder to spot than wrong things.
Audience. Read it as the person receiving it. Does it answer their actual question, in words they'd use?
Commitments. Underline every promise about timing, availability or results. Then confirm each one with the person who can authorise it. Promises nobody approved are the most common thing I catch. The AI isn't being reckless. It's completing a pattern. A confident email usually ends with a commitment, so it writes one.
Approval. Record who approved the version that went out. Not the draft, the version.
AI can run this checklist over its own work and it'll catch some things. Its review is another draft to inspect, not a sign-off.
What a person still does
Opens the source. Reads the draft as the recipient. Decides what we're prepared to promise. The final edit is yours, which is the whole point of having one.
If you'd like to see how this runs inside a real week, on invented examples, book a free call.