Why AI Hallucinates
And five ways to catch it in practice
① Hallucination isn't the model "lying" — it's that no step in the pipeline checks whether an output is true. It only produces probable sentences.
② The weak spots reduce to three: numbers, citations, and recency. All three surface as "perfect form, wrong content", which is exactly what visual review cannot catch.
③ The fix isn't a better model, it's a verification procedure: demand sources → re-ask → invert → decompose → check externally.
① Hallucination is not a bug
Saying "the AI hallucinates" makes it sound like an occasional fault during otherwise correct operation. It's closer to the opposite. The model is always doing the same thing — that thing just isn't fact-checking.
As covered in how LLMs work, the model's only job is picking the highest-probability next token. Nowhere does it ask "is this actually right?" True statements appear often because the training data contained many true statements, not because the model prefers truth.
Ask for a paper that doesn't exist and the model builds a token sequence shaped like a paper title. The more typical the form, the higher the probability — so the most natural-looking fabrications are the ones it produces best. Structurally, undetectable falsehoods come out more easily than obvious ones.
② Three weak spots
Hallucination isn't uniformly distributed. In practice, trouble almost always comes from three places.
| Type | Why it's weak | Typical symptom |
|---|---|---|
| Numbers | Digits are predicted as tokens, never computed | Revenue, share, or year off by a plausible-looking amount |
| Citations | Author + title + year is a highly regular format, easy to recombine | Real author + real journal + fabricated paper title |
| Recency | Events after the training cutoff simply aren't in the data | Last quarter's results "inferred" from older patterns |
What they share matters: the form is flawless and only the content is wrong. No grammatical errors, no odd tone. Reading it over will never catch it.
Almost everything you'd use for an investment decision falls into these three types — earnings figures, report citations, recent news. Taking a number straight from "what was Nvidia's revenue last quarter?" is risky. Verify against primary filings, as in our Big Tech earnings guide.
③ Confident tone is not a signal
People try to read reliability from tone: assertive means right, hedging means wrong. That's an illusion.
Tone is the training data's writing style showing through. Trained on encyclopedic prose, it writes assertively. This has no relationship to the internal probability of the claim. Low-confidence guesses often come out more assertive, simply because assertive phrasing is more common in the training corpus.
④ Five ways to catch it
1. Demand sources alongside the answer
Instead of "tell me about X", ask: "tell me about X, cite a URL for each claim, and mark anything you're unsure about as uncertain." Two things follow. Claims with real sources become checkable, and claims without them either get dodged or get a fabricated URL. A fake URL dies on one click — that's the point. A wrong URL is far cheaper to verify than a wrong fact.
2. Re-ask in a fresh conversation
Pose the same question in a new chat. Because hallucination comes from probabilistic sampling, it often differs run to run. If both answers agree down to the numbers, confidence rises; if they diverge, at least one is wrong. Re-asking in the same thread is useless — the previous answer is still in the input and gets repeated.
3. Ask the inverse
If the model claimed "A is B", open a new chat and ask "give me the evidence that A is not B." If it's true, the model struggles to build a counter-case. If it was hallucinated, it produces an equally fluent story in the opposite direction. Fluency in both directions means the claim is unreliable.
4. Break the question into pieces
Broad prompts like "analyse this company's financials" invite hallucination: many facts get generated at once and early drift compounds. Asking "just the revenue trend", "just the debt ratio" keeps each answer short, verifiable, and free of accumulated error.
5. Always verify numbers and proper nouns externally
Even after the first four, numbers must be confirmed at the source — company IR pages or SEC filings for earnings, exchange data for prices, the issuing agency for statistics. No exceptions. What AI is good at is telling you where to look, not remembering what it said.
□ Does the answer contain numbers → don't use until checked at source
□ Any citations or reports → click the URL yourself
□ Post-cutoff event → use a search-enabled tool or discard
□ Is this a load-bearing conclusion → re-ask and invert in a fresh chat
□ Is the answer long and complex → decompose and ask again
⑤ How much do RAG and search help?
Modern tools often search the web or retrieve documents before answering — RAG (retrieval-augmented generation). It genuinely reduces hallucination, because the model reads a document instead of reaching into memory.
But it introduces two new failure modes. First, the retrieved source can itself be wrong. Grounding on an inaccurate blog yields a cited answer that's still false. Second, the answer can drift from the cited document — the source is correct but a number changes during summarization.
So even with citations, open the source and find the sentence. The presence of a citation is where verification starts, not where it ends.
⑥ Closing — it's a matter of posture
The most practical stance is to treat AI as a highly capable assistant who cannot remember sources. Structuring an argument, surfacing angles you missed, drafting quickly — excellent. But you wouldn't copy that assistant's numbers straight into a report.
In short: ask AI what to look at; ask it what's true only if you'll go check.
Bias is easy to confuse with hallucination. Both stem from the same structural gap — no verification step anywhere in the pipeline — but hallucination invents facts that don't exist, while bias faithfully reflects a real statistical skew in the data, so the causes and the fixes differ.
※ This article reflects information as of July 2026 and is not investment advice. If you use AI-generated information for investment decisions, verify it against primary sources such as official filings and exchange data. All investment decisions and their consequences rest with the investor.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.