Imagine asking someone to recommend a good book on a topic. They confidently give you a title, author, publisher, and publication year — everything complete and convincing. And the book doesn't exist. Never did. They just made it up because you expected an answer.
Welcome to the world of AI hallucinations.
What is a hallucination?
The term sounds more dramatic than it is, but it does capture something real. An AI hallucination happens when a language model produces a statement that is factually wrong — but with the same confidence and style as a true one.
The model doesn't know it's wrong. There's no internal fact-checker running. It generates the statistically most plausible continuation of your input. And sometimes the most plausible continuation is, unfortunately, nonsense.
Why does this happen?
Here's the key misunderstanding: ChatGPT is not a fact database. It's a probability model.
During training, the model read enormous amounts of text. From that, it learned which words, sentences, and ideas frequently appear together. When you ask "Who wrote novel X?", the model doesn't look up the answer in a database. It calculates which name is most likely to appear in this context — based on patterns in the training text.
When those patterns are thin (because the book is obscure or the topic is complex), the model guesses. And it guesses with the same confidence as when it has actually learned a reliable answer.
What are the concrete risks?
Fake sources. This is the classic one: you ask for scientific studies on a topic, and ChatGPT gives you studies with real journal names, plausible author names — none of which exist. If you don't verify and cite them anyway, that looks bad. Or worse.
Wrong facts. Dates, numbers, names, historical events — all of this can be wrong. Especially for niche topics or recent events (the model has a knowledge cutoff and simply doesn't know anything that happened after it).
Wrong legal information. If you ask ChatGPT about laws, deadlines, or legal requirements and take the answer at face value, you can run into serious real-world problems.
Medical misinformation. Symptoms, diagnoses, medication dosages — here a hallucination can be genuinely dangerous.
Is ChatGPT deliberately lying?
No. This is important to understand. The model has no intentions. It's not optimizing for truth, because it has no concept of truth. It's optimizing for a plausible, fluent response. That's the difference between a liar and someone who has no idea but acts like they do.
"Hallucination" is actually a misleading word — the model doesn't experience anything. But it's the standard term, so we'll stick with it.
How do you deal with it?
Don't trust blindly. Anything you use for important decisions — medical, legal, financial, scientific — needs to be verified. Full stop.
Ask for sources, but verify them. ChatGPT will happily give you references. But those references need to be checked. Search the title in Google Scholar or the relevant journal and confirm it actually exists.
Use tools that connect to real sources. Perplexity AI, for example, provides actual links you can check directly. That significantly reduces hallucination risk — not to zero, but significantly.
Use AI for what it's actually good at. Structuring text, generating ideas, explaining code, writing summaries — hallucinations matter less there. Factual research without cross-checking, on the other hand, is risky.
Bottom line: powerful, but not trustworthy
ChatGPT is a powerful tool. But it's not an encyclopedia, a search engine, or an expert. It's a very convincingly-sounding language model that sometimes just makes things up.
That's not a reason to panic. It's a reason to stay alert. Understanding this lets you use AI sensibly — and still requires you to keep thinking for yourself. Which is actually the best thing you can say about a tool.
That wraps up our beginner series. What's next: practical use cases, step by step.