Why does AI make things up? Hallucinations explained

You ask ChatGPT a question. It gives you a confident, specific answer with names, dates, and citations. You feel good. You go to verify one detail and discover the citation is fake, the date is wrong, or the person it mentioned does not exist. This is called a hallucination. It happens with all AI chatbots and it is the most important thing to understand if you are going to use them for anything important.

What "hallucination" actually means

"Hallucination" is the word AI researchers use for when an AI chatbot generates information that sounds true but is not. It is not lying (the AI is not trying to deceive you). It is also not a bug in the usual sense. It is a side effect of how these systems work.

An AI chatbot like ChatGPT was trained by reading enormous amounts of text. When you ask it something, it predicts what words should come next, based on patterns from training. Most of the time those patterns produce correct answers because the training data contained correct information. Sometimes the patterns produce wrong answers because:

• The training data had errors
• The training data did not include the answer
• The question is about something newer than the training cutoff
• The AI is filling in a gap with what "sounds right" given everything else

Why does AI sound so confident even when wrong?

Two reasons:

1. The training data was confident. Books, articles, and websites are written with confidence. The AI learned that style.
2. Hedging gets penalized during training. Early versions of these systems rewarded "decisive" answers. Saying "I don't know" was treated as a worse answer than something specific.

Newer models have been trained to say "I'm not sure" more often, but the bias toward sounding confident is still there. Claude does this best; ChatGPT and Gemini are improving.

What kinds of things does AI hallucinate?

Fake citations and sources

The most famous category. Ask AI to cite sources, and it will sometimes invent papers, articles, court cases, or books. The titles sound real. The authors sound real. They do not exist. A New York lawyer was famously fined in 2023 for submitting a court brief with citations entirely made up by ChatGPT.

Numbers and statistics

AI will sometimes invent statistics like "73% of Americans..." that sound real but trace back to nothing. If you need a real number for a presentation, find the actual source.

Quotes from real people

AI will sometimes generate a "quote" attributed to a famous person that the person never said. Especially dangerous because the wording sounds plausible.

Dates and timelines

AI confuses years, gets dates of historical events wrong, and may misremember when something happened (especially anything close to or after its training cutoff).

Real people, fake details

If you ask about a real person, AI may mix in details from other people with similar names, or invent details that fit the pattern.

Code that does not work

AI confidently writes code referencing libraries or functions that do not exist. This is the version programmers see most often.

Math with multiple steps

AI is better at math than it used to be, but multi-step math problems often have one wrong step that throws off the whole answer.

When AI is most likely to hallucinate

• Specific recent events. AI's knowledge cuts off at a certain date. After that, it is guessing.
• Niche topics. Less training data on a topic means more invention.
• Anything that asks for exact sources or quotes. Citations are the most invented thing.
• Math that requires precise multi-step reasoning.
• Real people who are not famous.
• Numbers where the AI cannot calculate (most likely just predicts).
• "What did X say about Y?" questions.

When AI is least likely to hallucinate

• General how-to questions ("how do I unclog a sink?")
• Writing tasks (drafting emails, editing your text)
• Brainstorming ideas
• Summarizing text you provide directly
• Translating between common languages
• Explaining well-known concepts
• When it has just searched the web and shows you the source

How to reduce hallucinations

1. Turn on web search

If the AI just searched and cited sources, it is much harder for it to invent. Most current AIs do this automatically when needed. You can force it: "Search the web for [X] before answering."

2. Use Perplexity for factual questions

Perplexity (perplexity.ai) is built around "always cite sources." Every answer comes with the web pages it pulled from. You can verify in one click.

3. Paste your source in

Instead of asking "What does the IRS say about HSAs?" copy the actual IRS page and paste it. Then ask. The AI now has the source in front of it; less room to invent.

4. Ask for uncertainty

Add to your prompt: "If you are not sure about something, say so. Do not guess." Or: "Rate your confidence in this answer from 1 to 10." This nudges the AI to be more honest about gaps.

5. Cross-check critical claims

Before you act on any AI-generated number, name, date, or quote that matters, Google it. If it is real, you will find it in a few seconds. If it is made up, you will find nothing or only AI-generated content that looks suspicious.

6. Ask the same question to two different AIs

If ChatGPT and Claude give you the same answer, it is probably right. If they disagree, one of them is wrong (or both). Worth knowing.

The biggest mistakes people make with AI accuracy

• Trusting medical answers. AI can be a great starting point for "what does this term mean" but should never replace a doctor's input. Especially for medications and dosing.
• Trusting legal answers. AI can explain general concepts but it invents case names and laws often enough that you should never rely on it for an actual legal decision.
• Using AI quotes in important writing. Always verify the source before publishing anything with a quote AI gave you.
• Trusting AI math without checking. Multi-step math is especially error-prone. Verify with a calculator.
• Assuming AI knows about anything recent. If a topic is from the last few months, AI is likely guessing unless it searched the web.

A real test you can run

Ask any AI: "Give me 3 academic papers about [niche topic you know well]. Include the authors, year, and journal." Many AIs will produce a confident list. Then Google the papers. There is a good chance one or more is invented.

The lesson: just because AI cites something does not mean it exists.

Will hallucinations go away?

They are getting less common with each new model generation, but probably not zero anytime soon. The underlying way these systems work (predicting plausible text) is fundamentally pattern based, not fact based. New techniques like "retrieval" (where the AI looks up real sources before answering) help a lot, which is why web-search-enabled AIs hallucinate much less.

The honest takeaway: AI is a remarkable assistant for many tasks, but it is not a fact lookup tool. Use it like a smart conversational helper. Verify anything that matters.

The questions you should never just trust AI on

• Medical diagnosis or medication doses
• Legal advice for your actual situation
• Tax obligations (the rules are recent and complicated)
• Financial decisions over $1,000
• Anything that names a specific real person
• Direct quotes from books, papers, or speeches
• Court cases, laws, regulations (always invented citations are common)
• Software code that you cannot test before relying on it

For these, AI is fine as a starting point. Always verify with a professional or a real source.

The bright side

Most everyday AI uses are not high-stakes:

• Drafting an email
• Suggesting dinner ideas
• Explaining something you can sanity check
• Brainstorming names, ideas, gifts
• Summarizing your own documents
• Editing your own writing

For these, hallucinations rarely matter. You read what came back. If it looks good, you use it. If something is wrong, it is obvious.