Usage & Prompting
Hallucination
The model produces plausible text, even where it knows nothing. That yields invented sources, statutes, and numbers, delivered in the same confident tone as everything else. This is not a bug but the flip side of how it works: a system that generates the most probable continuation has no built-in mechanism that prefers not knowing over a fluent answer. The term itself is misleading, because it sounds like a rare lapse. In fact it is the same operation as for every correct answer, just without backing in the data. It becomes dangerous because the errors do not sound absurd but exactly the way a correct statement would have to sound. What helps against it is grounding, mandatory sourcing, and checks at the points where an error has consequences.
An example
A researcher has a literature list compiled on a topic. She gets twelve references: plausible titles, existing journals, consistent volumes, correctly formatted DOIs. Four of them don't exist. A fifth title exists but is by different authors. Nothing about it looks wrong: the model has learned how a good reference sounds, and that is exactly what it produces when it doesn't know the real one.
A closer look
The model has no state for “I don't know.” In every case it produces the most probable continuation, even when the data basis is thin. Invented evidence therefore looks just like real evidence: formally correct, free in content.
Why it matters
The decisive point in practice: hallucinations are not an error that can be trained away but the flip side of how the system works. A model that always produces the most probable continuation has no mechanism for preferring “I don't know.” The build rule follows from this: don't trust the model, supply the source and make the answer checkable against it. And the process rule: build the check in where the error becomes expensive, at numbers, names, statutes, and references. All of it is cheap to check and expensive to miss.
Commonly misunderstood
Hallucinations can't be prompted away. They decrease through grounding, enforced sourcing, and verification steps, and remain a residual risk that the workflow has to catch.
Where this term does its work in the magazine.
3 articles use the term or take up its principle.
On Stage
The Demo and the Tuesday
A good demo makes a claim memorable. It doesn’t make it true.Read the articleGlass Lab
AI for Researchers: The Honest Guide
The most useful boundary is simple: only delegate what you can verify, disclose, and defend.Read the articleFrom Science