AI Hallucinations: Why Chatbots Make Things Up
Why do AI chatbots state false facts so confidently? Learn what AI hallucinations are, why they happen, and seven practical ways to spot and reduce them.
You ask a chatbot for a source, and it gives you a neat title, an author, and a year. It sounds perfect. Then you search for it, and the source does not exist. This is an AI hallucination: an answer that is fluent and confident, but false. Here is why it happens and how to protect your work from it.
What is an AI hallucination?
An AI hallucination is content generated by a model that is presented as fact but is not supported by reality or by the input it was given. Common examples include invented quotes, fake citations, wrong dates, and code that calls functions that were never written. The US National Institute of Standards and Technology calls the same problem "confabulation" in its AI Risk Management Framework work on generative AI.
Why do chatbots make things up?
1. They predict text, not truth
Large language models generate answers one token at a time by predicting what text is most likely to come next. If you want a refresher, read our guide on how language models read text as tokens. A model does not look facts up by default. When the right answer is rare or missing in its training data, it can still produce something that merely sounds right.
2. Guessing is often rewarded
OpenAI researchers argue in Why Language Models Hallucinate that common training and evaluation methods reward guessing over admitting uncertainty. It is like a multiple-choice exam: leaving a question blank scores zero, while a lucky guess can score points. The full paper is available on arXiv.
3. Vague prompts leave gaps to fill
When a question lacks context, the model fills the gaps with assumptions. The more specific facts it has to invent, such as names, numbers, and links, the higher the risk.
A real-world warning
In 2023, lawyers in the US case Mata v. Avianca submitted a court filing that cited cases invented by ChatGPT. The court sanctioned them. The lesson is simple: AI can draft, but people remain responsible for what they publish.
7 practical ways to reduce hallucinations
Give it the source. Paste the document, data, or notes and ask the model to answer only from that material.
Allow "I don't know". Say clearly that it is fine to answer "not enough information" instead of guessing.
Ask for quotes first. For long documents, ask for the exact supporting sentences before the summary, so every claim has evidence.
Use retrieval. Retrieval-augmented generation, or RAG, connects a model to a search index or knowledge base so it can ground answers in real documents.
Break big tasks down. Short, focused steps are easier to check than one huge answer. Our article on better prompts with context, format, and examples shows how.
Verify the risky parts. Always check names, numbers, dates, citations, legal or medical claims, and links. Open every URL before you share it.
Test code before trusting it. Run it, read the errors, and confirm that every library and function actually exists.
Treat an AI answer like a draft from a fast, helpful intern: useful, but always reviewed.
Hallucination is not always bad
In creative work, "making things up" is the point. Brainstorming names, story ideas, or game lore benefits from imagination. The problem appears only when invented content is presented as fact. Decide which mode you need before you prompt.
Keep exploring
Want to build reliable AI tools? Read about small AI agents that are actually useful, or discover what the Ryuartistry studio builds with AI, games, and software.