Before you forward it: the 90-second AI answer check

It is late, the memo is due in the morning, and the AI just handed you a paragraph that sounds exactly right. There is a small voice asking whether you should check it, and a louder voice saying it reads fine and you are tired.

Everyone who works with AI knows that moment. The forwarded answer that turns out to be wrong does not fail loudly at your desk; it fails later, in front of your boss, your client, or your professor, when someone asks where a number came from. The 90-second check exists so that the question "where did this come from?" is one you have already answered.

Ask the tool where the claim came from, and then actually open it. Not the tool's description of the source. The source.

This single step filters out the worst failures, because a fabricated reference has nothing behind the link, and a real reference that was misused starts to unravel the moment you see it in context. If the tool cannot show you a source at all, you have learned something even more useful: this claim is currently unverifiable, and it should travel with that label or not travel at all.

Read the sentence the AI wrote. Read the passage it points to. Ask one question: does the passage actually say this?

The subtle failures live here. A passage about one jurisdiction supporting a claim about another. A finding from 2019 presented as current. A cautious "may be associated with" upgraded into a confident "causes." None of these look like errors in the answer itself; the answer is fluent either way. Columbia's Tow Center found even the best AI search assistant cited sources incorrectly 37% of the time, and the errors it catalogued were mostly this kind: real sources, wrong support. The passage is where the truth lives, so the comparison is where the check lives.

This is the step that separates careful readers from everyone else, and it takes fifteen seconds. After confirming what the passage supports, say to yourself what it does not: it supports the rule, but not the exception you also need; it covers the survey data, but not the conclusion drawn from it.

That one habit changes your relationship with AI. You stop treating answers as verdicts and start treating them as evidence with edges. The people you work with will notice, because your version of the answer arrives with its limits attached, and limits are what competent work sounds like.

A habit survives only if it is nearly free. If checking a claim means searching for the document, finding the page, and scanning for the sentence, you will do it twice and then never again. The design answer is to work in tools where every claim is one click from its passage.

That is the principle Tatsulok is built on: answers assembled from your own documents and curated collections, each sentence linked to the exact passage behind it, and an honest "your sources do not contain this" when the evidence runs out. Inside a tool like that, the 90-second check costs seconds. But the ritual itself belongs to you, not to any product. Practice it everywhere, and no AI tool gets to spend your credibility for you.

Try it today, no signup needed: take the last AI answer you almost forwarded, and run the three steps. Open the source. Compare the claim to the passage. Say out loud what the passage does not prove. Time yourself. It will be under two minutes, and you will know something you did not know before.