You ask ChatGPT, Claude or Gemini a question and get a clear, confident, well-written answer. Then you check, and part of it is simply wrong: a fake quote, a date off by ten years, a research paper that doesn't exist. This is called an AI hallucination, and knowing how to spot it is one of the most useful AI skills you can have.
What is an AI hallucination?
An AI hallucination is any output that sounds right but isn't supported by reality. It isn't a bug in one app. It's a side effect of how all large language models work. Common types:
- Invented facts: events, people or statistics that don't exist.
- Fake citations: realistic-looking papers, books, court cases or links that were never written.
- Wrong details: real facts with the wrong date, number or name attached.
- Mash-ups: two true facts combined into one false one.
- Going along with you: repeating a false claim in your question as if it were true.
Can you spot them? Try it
Here are three answers in the style real chatbots produce. Each mixes true and false claims. Find the fakes:
🕵️ Spot the hallucination
Round 1 of 3Tap every underlined claim you think the AI made up. Then press Check.
Notice the pattern: the made-up parts sit right next to true ones and sound just as confident.
Why does AI make things up?
In how ChatGPT works you saw that a language model writes by predicting the next word, over and over. It picks words that are likely given everything it read during training. It doesn't, by default, look anything up or check it.
“The first person to walk on Mars was …”
So hallucinations are most likely when the model's patterns are thin or misleading:
- Rare or niche topics it saw little about during training.
- Recent events after its training cut-off.
- Exact details such as figures, dates, page numbers and quotes.
- Leading questions that assume something false (“Why did India win the 2015 World Cup?”).
- Pressure to answer: models are trained to be helpful, and an answer often feels more helpful than “I don't know”.
When hallucinations matter most
A made-up birthday wish is harmless. A made-up medicine dose is not. Sort these prompts by risk:
⚖️ High risk or low risk?
6 promptsHow bad would it be if the AI made something up here?
“What are the side effects of this medicine?”
“Write 5 fun title ideas for my college fest.”
“Which sections of the IT Act apply to my case?”
“Summarise this article I pasted in.”
“Give me 3 research papers to cite in my assignment.”
“Explain recursion with a simple example.”
7 ways to avoid AI hallucinations
- Check the risky five. Names, numbers, dates, quotes and citations are where made-up details hide. Verify these first.
- Ask for sources, then open them. A link or reference you haven't opened is not a source. Fake citations look perfectly real.
- Give it the facts. Paste the article, notes or PDF and say “answer only from this text”. That's the idea behind RAG.
- Allow “I don't know”. Add: “If you're not sure, say so instead of guessing.” Models follow this surprisingly well.
- Use search or tools for recent info. Built-in knowledge stops at a training cut-off. Ask a version with web search for news and prices.
- Ask “how sure are you?”. Follow up with “What in your answer might be wrong?” It often flags its own weak spots.
- Cross-check important answers. Health, money, law, exams: confirm with an official site, a textbook or a real expert.
A prompt that bakes several of these in:
Answer using ONLY the notes below. If the notes don't contain the answer, say "Not in the notes" instead of guessing. After your answer, list which line of the notes you used. Notes: <paste your notes here> Question: What are the three causes of the 1857 revolt?
For more techniques like this, see prompt engineering for beginners, the next part of the course.
How AI companies are reducing hallucinations
- Retrieval (RAG): the AI looks up relevant documents first and answers from them. It has its own part in this course.
- Web search and citations: answers link to the pages they came from, so you can check.
- Training models to say “I'm not sure”: rewarding honesty over confident guessing.
- Self-checking: having the model review or verify its own answer before showing it.
Where to go next
The best protection against hallucinations is asking well. Continue to prompt engineering for beginners.
Key takeaways
- ✓An AI hallucination is a confident answer that is false or invented.
- ✓It happens because models predict likely words; they don't check facts by default.
- ✓Watch names, numbers, dates, quotes and citations most closely.
- ✓Ask for sources and open them; paste your own source text when accuracy matters.
- ✓For health, money, law and exams, always verify with a trusted source.
Frequently asked questions
What is an AI hallucination?
An AI hallucination is when a chatbot or other AI tool gives an answer that sounds confident and fluent but is false or made up: an invented fact, a fake quote, a wrong date or a research paper that doesn't exist. The AI isn't lying on purpose; it is producing text that looks plausible without checking whether it is true.
Why does ChatGPT make things up?
Language models are trained to predict the next most likely words, based on patterns in huge amounts of text. They don't look facts up in a database by default, and they are rewarded for giving an answer rather than saying “I don't know”. When the patterns are weak (rare topics, recent events, exact numbers, citations), the most plausible-sounding continuation can be wrong.
How common are AI hallucinations?
Modern models hallucinate much less than early chatbots, especially on well-known topics, and tools that search the web or read your documents reduce it further. But no model is hallucination-free. Errors are most likely with niche facts, very recent events, exact figures, legal or medical details, and citations.
How can I tell if an AI answer is a hallucination?
Be most careful with names, numbers, dates, quotes and citations. Ask the AI for its sources and open them, search for key claims yourself, compare with a second trusted source, and ask the AI how confident it is or what could be wrong. If a reference or quote can't be found anywhere, treat it as made up.
Can AI hallucinations be completely fixed?
Not completely, with today's technology. Techniques like retrieval-augmented generation (RAG), web search, better training and letting models say “I'm not sure” reduce them a lot. For anything important, a human should still check the answer.
🧠 Quick check: AI hallucinations
1. What is an AI hallucination?
2. What is the main reason language models hallucinate?
3. Which of these is most likely to be hallucinated?
4. What's the best way to reduce hallucinations when asking about a specific document?
Checking your access…