What a hallucination is
A hallucination is a confident answer that does not match reality: an invented quote, a nonexistent book, a link to a page that isn't there, a wrong number carried out to three decimal places.
The cause sits in the very nature of the model from the first lesson: it generates a plausible continuation of text. If the right answer isn't in the weights, or is blurry, the model doesn't stop — it fills the gap with the most plausible option. The form comes out perfect: even style, confident tone, convincing details. That is exactly why hallucinations are dangerous — they are indistinguishable from the truth by appearance.
Where the risk is highest
Memorize the high-risk zones — check every fact in them:
- Exact numbers: statistics, prices, dates, device specs, dosages.
- Quotes and sources: the model readily "quotes" things a person never said and cites nonexistent articles.
- Links: URLs can look real and lead nowhere.
- Narrow niches: the fewer texts on a topic there were in training, the bolder the model's invention. About Python it rarely errs; about the API of an obscure service, constantly.
- Recent events: anything past the knowledge cutoff (previous lesson).
- Calculations: the number of letters in a word, arithmetic with large numbers — the model "estimates" rather than computes.
Legal, medical, and financial decisions based on the model's answer without checking with a specialist or the primary source are a straight road to expensive mistakes. The model is useful as an assistant in preparation, not as the final authority.
How to reduce the risk
Hallucinations can't be removed entirely, but their frequency drops several times over with simple techniques:
- Give it material. Attach a document, an article, data — and ask it to answer strictly from them. A model with a source in front of it hallucinates an order of magnitude less.
- Allow "I don't know." The phrase "if you're not sure, say so" really works: the model gets a legitimate way out instead of an obligation to answer something.
- Demand sources and check them. "Give links to sources for each claim" — then open two or three links. Broken links expose fabrication instantly.
- Turn on web search for factual questions — an answer grounded in found pages is more reliable than one "from memory."
- Re-ask in a new chat. Pose the same question fresh, with no hints. Two independent answers that agree are worth more than one.
Where the model is strong and where it's weak
Strong zones — where you can assess the result yourself: texts, ideas, structure, explanations, code (you run it, you check it), work with attached documents. Here hallucinations are either rare or immediately visible.
Weak zones — where the answer can't be checked at a glance: exact facts from memory, numbers, quotes, niches. Here the model's answer is a draft by default, requiring verification.
Practical rule: the higher the cost of an error and the harder it is to check, the less you should trust an answer "off the top of the model's head."
Key takeaways
- A hallucination is a plausible fabrication; by style and tone it is indistinguishable from a correct answer.
- Risk zones: numbers, quotes, links, narrow topics, recent events, calculations.
- Reducing risk: give material, allow "I don't know," demand sources, turn on search, double-check in a new chat.
- The more expensive the error, the more mandatory the check against the primary source.