The CHECK Framework
About this resource
AI produces wrong information with total confidence, and it does so most reliably around the things people are least likely to question: citations, statistics, quotations, proper names, dates, and anything recent. That is a property of how these systems work rather than a defect awaiting a patch, which means verifying AI output is a permanent step rather than a temporary precaution.
This free one-page handout takes each of the five CHECK steps and gives you the specific question to ask at that step, in language you can apply to a real piece of work rather than to a principle. It covers identifying the claims being made, spotting what could plausibly have been invented, testing whether the reasoning actually holds or merely sounds like it does, judging whether the model had the context it needed, and recognizing the knowledge gaps where an answer should be treated as unreliable by default. CHECK is a widely used community mnemonic rather than something we invented.
Once it becomes habit, running CHECK on a piece of AI output takes under a minute. It is the single highest-return behavior in everyday AI use.
How to use it
Teach it once, then reference it by name. The value of a mnemonic is not that it contains new information but that it survives time pressure at four o'clock on a Friday, which is exactly when unverified work tends to leave a building. Pair it with a stated rule about which categories of work require it, so nobody has to decide how careful to be while the deadline is running.
Inside the download
Why naming what the output asserts has to come before defending or dismissing it.
Which categories of detail models invent most often, so you know where to look first.
How to tell logic that holds from logic that merely reads smoothly.
Whether the model had what it needed, or whether the answer is generic where it should be specific.
Common questions
Work through five things. Identify the factual claims the output makes. Consider whether any citation, statistic, quotation, name, or date could have been invented. Test whether the reasoning actually holds rather than merely reading well. Ask whether the model had the context it needed or produced something generic. Finally, check whether the topic is time-sensitive, proprietary, or industry-specific in a way the model could not know.
Specifics. Citations, statistics, quotations, proper names, dates, URLs, legal and regulatory detail, and anything recent are the categories where fabrication is most common. General explanations of well-established concepts are considerably more reliable than any particular figure attached to them.
Under a minute for most work, once the habit is formed. The effort should scale with the stakes: a first draft you will rewrite anyway needs a quick read, while anything reaching a client, a filing, or a system of record needs every factual claim checked against a source.
Free. No email list, no follow-up sequence.
More from the resource library, chosen because they pair well with this one.
Two practical mnemonics that anchor everyday AI use: better prompts in, verified answers out.
Seven habits that make AI writing obvious, and one copy-and-paste instruction that switches all of them off.
Why fabrication happens, where it concentrates, and a detection routine scaled to the stakes.
Tell us a little about yourself and this document is yours. We don't have an email list. We collect this so we know who's finding our resources useful, and that's it.
You're downloading
this document
Your information is kept confidential and never shared with anyone.
