How AI chatbots work and why they "hallucinate"
Chatbots such as ChatGPT, Claude or Gemini are not machines that "know" — they are language models that predict the next word by probability. This neutral reference explains, on the basis of sources, how they work, why they sometimes answer with confident falsehoods (hallucination), and how to use them critically.
A large language model (LLM) — the "heart" of chatbots such as ChatGPT, Claude or Gemini — is in fact a giant statistical prediction machine. It learns from text containing billions of words (books, articles, websites, code) which words tend to follow which, and on that basis repeatedly predicts the next word (more precisely, the next "token"). In other words it does not reason — it generates text that fits the patterns it learned. (Source: IBM — What Are Large Language Models.)
It does not "know" — it picks what "fits"
When the model chooses the next word, it computes not the truth about the world but which word best fits its pattern (highest probability). In IBM's words, it "optimizes for plausibility within the pattern, not for correctness about the world." That is why a chatbot's text can read fluently and convincingly while the information in it is not necessarily correct. (Source: IBM — What Are AI Hallucinations.)
What a "hallucination" is
An AI "hallucination" is when the model produces an answer that sounds plausible but is actually wrong, irrelevant or entirely made up. This can take the form of fake "facts," nonexistent studies, or invented, non-working links (URLs) — all presented as if factual. The reason is simple: the model has no concept of "true" or "false," only of "what fits the pattern"; when data is sparse or ambiguous, it fills the gap with a plausible but false detail. (Source: IBM — What Are AI Hallucinations.)
Why it does not say "I don't know"
According to research OpenAI published in September 2025, one reason models hallucinate is that the way they are trained and evaluated rewards guessing. In many benchmark tests an honest "I don't know" scores zero, just like a wrong answer; so the most "useful" strategy for the model becomes never to abstain but always to guess. That is why it answers confidently even when uncertain. (Source: OpenAI — Why language models hallucinate, 2025.)
Knowledge is "frozen": the knowledge cutoff
A model's knowledge is "frozen" at the time it was trained — this is called the knowledge cutoff. Events after that date simply do not exist for the model. By default a chatbot does not search the internet — it answers only from what it was trained on; to obtain current or new information it must be connected to a separate search/retrieval system (such as web search or document lookup). (Source: Wikipedia — Knowledge cutoff.)
A short history
The first widely used chatbot — ChatGPT — was launched by OpenAI on 30 November 2022 as a free "research preview" running on the GPT-3.5 model. The term "hallucination" itself was borrowed from earlier computer-vision work and came to describe text output "unfaithful to the source"; it spread widely after ChatGPT appeared. (Source: Wikipedia — ChatGPT; ACM Computing Surveys — Survey of Hallucination in NLG, 2023.)
How to use a chatbot properly
The expert advice is simple: treat a chatbot's answer not as final truth but as a draft that must be checked. Re-verify any important number, date, quote, or medical/legal/financial advice against a reliable primary source. Ask the model for its sources — then open them yourself to confirm they actually exist, because the model may invent a nonexistent address. The more serious the field (health, law, money), the more important the caution. (Source: IBM — What Are AI Hallucinations.)
Conclusion
A chatbot is a powerful assistant, but it does not "know" — it confidently writes text that is "probably how it goes." Its strength is fast, fluent text; its weakness is that it can state falsehoods with confidence. So the golden rule is one: trust, but verify. This article is general information, and technology changes fast — check the capabilities and limits of specific models in their official documentation.
