AI Hallucinations
About this resource
A hallucination is not a glitch. It is a direct consequence of how these systems work: they generate plausible continuations rather than retrieving facts, and a fabricated citation is produced by exactly the same process as a correct one. That is why AI hallucinations arrive with the same confidence as accurate answers and cannot be spotted by tone.
This guide covers why it happens, the warning signs that raise the probability an answer is invented, the categories where fabrication concentrates, and a practical detection routine you can run in proportion to the stakes. It combines the warning-signs material with the detection checklist into one document, because knowing the signs without a routine produces anxiety rather than accuracy.
Fabrication comes from the same mechanism as correct output. There is no future version where confident wrong answers stop existing, which is why verification is permanent.
How to use it
Read the hot-spot list and check your own work against it this week. Most people verify what feels uncertain, which is exactly backwards: fabrication concentrates in specifics that read as authoritative, and the parts of an answer that feel most solid are frequently the parts most worth checking.
Inside the download
The mechanism, which is what makes the risk predictable rather than random.
What raises the probability that a given answer is invented.
The categories where fabrication concentrates, so checking is efficient.
A repeatable process scaled to how much the work matters.
Common questions
Because these systems generate statistically likely text rather than retrieving stored facts. A plausible-sounding citation is produced by the same process as a correct one, so there is no internal signal distinguishing them. This is a property of the architecture rather than a defect awaiting a fix.
Specifics. Citations and references, statistics and figures, direct quotations, proper names, dates, URLs, legal and regulatory detail, and anything after the training cutoff. General explanations of well-established concepts are far more reliable than any particular fact attached to them.
Not from tone, which is the central difficulty. Fabricated content is delivered with the same confidence as accurate content. What you can do is recognize the high-risk categories and check those specifically, rather than trying to sense which parts of an answer feel less certain.
Free. No email list, no follow-up sequence.
More from the resource library, chosen because they pair well with this one.
A five-step human-in-the-loop verification discipline you can run on any AI response in under sixty seconds.
The extended treatment of CHECK, with more depth at each of the five verification steps.
What these tools do genuinely well and where they reliably fail, so you can assign work correctly the first time.
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.
