ChatGPT makes up believable facts because it generates text by predicting what words are most likely to come next, not by looking up verified information — so when it doesn't know something, it produces the most plausible-sounding answer instead of admitting the gap.
This is called a hallucination, and it's a built-in side effect of how large language models work, not a bug that gets patched away. The model isn't lying to you; it has no concept of "true" versus "false" in the way a search engine does. It only knows what a confident, fluent answer looks like. That's why a fabricated citation can look exactly as polished as a real one.
The mechanism is worth understanding because it explains the pattern of errors. During training, a model like ChatGPT reads enormous amounts of text and learns statistical relationships between words, phrases, and ideas. When you send a prompt, it generates a response one token at a time — each token chosen based on probability given everything that came before.
There's no internal fact-check step. If you ask about a real event, the model draws on patterns it absorbed during training, which may be outdated, incomplete, or blended with similar-sounding material. If you ask about something obscure or invented, the model still produces fluent output because fluency is the only target it was optimized for.
According to our AI tool database, ChatGPT is OpenAI's flagship assistant with a 1M context window and 700M+ weekly users, and it carries an editorial rating of 4.9/5 — but even a top-rated model has no retrieval guarantee unless it's explicitly connected to a search or database tool. Claude, Anthropic's safety-first model, has similar architecture and similar failure modes. The hallucination problem is structural, not brand-specific.
Here's a concrete example. Suppose you ask: "What year did the physicist Eleanor Voss publish her paper on quantum decoherence in the Journal of Modern Physics?" If Eleanor Voss doesn't exist, a model may still reply: "Eleanor Voss published her paper on quantum decoherence in 1987 in the Journal of Modern Physics, volume 14, pages 203–219."
Every detail is fabricated — the name, the year, the volume, the page numbers — but the sentence reads like a real citation. A beginner might copy that into a report. The way to catch it is to treat any specific name, date, number, or citation as unverified until you check it against a primary source: a DOI lookup, a library database, a company's official filing, or the original document.
If the model can't give you a source you can independently open and read, assume the detail is invented. This isn't about distrusting the tool; it's about knowing which outputs need verification and which don't.
A practical decision rule: trust the model for brainstorming, rephrasing, summarizing text you paste in, and explaining concepts you already partly understand. Verify anything involving names of people or organizations, dates, statistics, prices, legal or medical claims, citations, and current events.
The reason is simple — those categories require retrieval of specific facts, and the model's training data may be stale or the fact may never have been in it. The limits of this explanation are worth stating plainly: we can describe the mechanism, but we can't predict exactly when a hallucination will occur.
Some models now include web search or file retrieval that reduces the problem, but it doesn't eliminate it, because the model still has to decide what to search for and how to interpret what it finds. If you're new to this, a good next step is to understand what ChatGPT is and how it actually works before relying on it for anything factual.
And if you're using AI at work, be aware that passing along fabricated details can create real problems — see Can I get in trouble for using AI at work or in a job interview? for the practical side of that.