Hallucinations tend to cluster in predictable places. If you know where they show up, you know where to look first. There are four forms to watch for.
The four forms#
Each form looks believable on the page, and each needs a different check.
Fake citations and sources#
This is one of the most dangerous forms, because it looks authoritative. Ask for a reference and you may get a realistic author name, a believable title, a plausible journal and a year, none of which exist.
It is hard to spot because every part looks right. The cause is the same next-token prediction loop: the model writes what a citation usually looks like, not one it looked up. To check, find the source yourself and confirm that the author, title and year match a real entry.
Invented facts and numbers#
Be suspicious wherever specific figures appear: statistics, dates, prices and version numbers. The model has no calculator and no database, so any specific number it produces is a prediction, not a lookup.
A wrong figure reads just like a right one, so the tone gives you no warning. To check, trace each number back to a primary source, and recompute any arithmetic yourself.
Made-up code#
The model may suggest a function, a library or an API that does not exist, or that existed in an old version and has since changed. The code looks plausible and may even be syntactically correct. It just will not run.
That makes it easy to miss when you only read it. To check, run it, and compare the names it uses with the documentation for the version you have installed.
False premises accepted#
This form is subtler. Ask "Why did the 2019 policy fail?" about a policy that never existed, and many models will explain the failure instead of pushing back. This comes from the habit of pleasing the user, called sycophancy, which is covered in training.
The model aims for a plausible continuation of your question. Explaining a failure looks more plausible than saying it never happened. To check, confirm the premise yourself before you trust the explanation, or ask the model whether the event happened at all.
Short line#
Hallucinations cluster in fake citations, invented numbers, made-up code and accepted false premises. Check each one against a real source, a real run or the actual facts.
Key takeaways#
- Fake citations look authoritative: realistic author, title, journal and year, but no real source.
- Any specific figure (statistic, date, price, version) is a prediction, not a lookup.
- Made-up code may be valid syntax and still call a function, library or API that does not exist or has changed.
- Models often explain a false premise instead of challenging it.
- Each form needs its own check: find the source, verify the number, run the code, confirm the premise.