Published 2026-10-27
Hallucination
A model stating something false with the same confidence as something true. In quant work it shows up as invented numbers, fields and sources.
What it is
A hallucination is a statement a model produces with full confidence that is simply not true.
The important part is not that it is wrong — models are often wrong. It is that the wrong answer is formatted exactly like a right one. Same tone, same precision, same structure. There is no tell.
Why it is worse in quant work
In most domains a hallucination is caught by reading it. In quant work it is caught by checking it, and most people do not check.
A model that invents a closing price to two decimal places produces something that looks like data. Nothing about 1683.20 announces itself as fiction. To tell whether it is real you need a second source — which is precisely the work the model was supposed to save you.
This is why rule 5 of the standards puts the model in one role and the code in another: the code computes, the model narrates.
Three forms
| Form | Example |
|---|---|
| Invented number | A price, a return, a ratio that no source contains |
| Invented field | A field name the API does not return, described as if it does |
| Invented citation | A paper, author or URL that does not exist |
The third is the most dangerous, because a fabricated citation looks like verification. It is the one that turns a wrong answer into a wrong answer with apparent evidence.
How to check
One rule: every number must come from a tool response, not from the model.
The model may summarise, rank and explain. It may not produce the figure. If a number cannot be traced to a source you can open, treat it as invented until proven otherwise — including when it is right.