Ask a language model for a company's revenue and it may give you a figure that is wrong but written with total confidence. Ask it for a market statistic and it can produce a number that looks real and never existed.
These fabricated outputs are called hallucinations, and they are the biggest danger of using AI in financial research: the mistake is not the rare obvious error, but the fluent plausible fact.
A hallucination is a confident, fluent output that is not grounded in fact. Language models produce them because they are trained to generate plausible text, not to verify facts. In financial research the risk is a fabricated number or source that looks real, so every AI-produced figure and citation must be checked against the underlying data.
What a Hallucination Is
A hallucination is an output that is fluent and confident but not grounded in fact. The model produces text that sounds right because it is built to sound right, not because it checked the claim against reality.
It is different from a normal error because it is hard to catch: the fabricated fact looks and reads exactly like a real one.
Why It Happens
A language model is trained to predict the next word in a way that produces plausible prose. It has no built-in step that verifies the claim against a database or a document.
That is why the confidence of the answer tells you nothing: the model is equally fluent whether the underlying claim is true or invented.
| Type | Example | Why it is dangerous |
|---|---|---|
| Fabricated number | A revenue figure that never existed | Looks real, easy to quote |
| Invented source | A study or report that does not exist | Feels like evidence |
| Wrong date or quantity | A plausible but incorrect detail | Slips into a summary |
| Confident false claim | A 'well-known' market fact that is false | Sounds authoritative |
Why It Matters in Financial Research
Financial research runs on facts: a revenue number, a price level, a statistic, a source. If any of those is fabricated and you build an analysis on it, the conclusion inherits the error.
The danger is not the obvious mistake. It is the plausible fact that passes every reading check and changes what the research concludes.
| Item | Figure the AI gave | After checking the official filing |
|---|---|---|
| Company X's full-year revenue | $960 million | $840 million |
| Price/sales, at a $5.0 billion market cap | ≈ 5.21x | ≈ 5.95x |
Market cap is unchanged; only the revenue input was swapped for the real figure, and the resulting price/sales moved from 5.21x to 5.95x.
A 14.3% overstatement in revenue pulls the price/sales ratio down by about 0.74x. If a later judgment about whether the stock looks cheap or expensive rests on that understated multiple, the conclusion is wrong from the start — and the error began with a single unverified number.
A model produces confident prose because it is trained to. The fluency of the answer is not evidence that the underlying claim is true.
How to Check for Hallucinations
Verify every number against the actual data or document. If the model cites a figure, find the figure in the source and compare. If it names a report or study, confirm the report exists.
The habit is to treat any AI-produced fact as unverified until traced. A figure that cannot be found in the source should be discarded, not quoted.
Frequently Asked Questions
Why does AI fabricate facts so confidently?
Because it is trained to generate plausible text. Confidence is a property of the writing, not evidence the claim was checked.
Can hallucinations be eliminated?
Not reliably. The practical defense is verification: check every figure and source against the underlying data.
Are hallucinations more likely with specific topics?
They are more dangerous with specific, checkable facts like numbers and citations, because the fabricated version looks identical to a real one.
