AI Hallucinations: Why AI Makes Things Up (and How to Catch It)

⚠️ Affiliate Disclosure: This post is not an affiliate associated review. It may become one in the future and this disclosure will reflect that. This does not affect my reviews and I only recommend tools I genuinely believe in.

What Are AI Hallucinations?

AI hallucinations are answers an AI tool gives with total confidence that are simply wrong — a made-up fact, a book that was never written, a court case that never happened. The tool is not lying and it is not broken. It is doing exactly what it was built to do: produce the most likely-sounding answer.

That is what makes them hard to spot. A wrong answer from ChatGPT, Claude or Gemini looks identical to a right one: the same fluent tone, the same tidy formatting, often a confident citation attached. Every AI chatbot still does it in 2026. The newer models do it less, not never.

Why Does AI Hallucinate?

There are three main reasons, and none of them is a bug someone forgot to fix.

1. AI predicts words — it does not look up facts

A large language model writes by predicting the next most likely word, over and over, based on patterns in its training data. It has no built-in fact database to check against. When the pattern is strong — the capital of France — the prediction is right. When the pattern is thin — a niche product’s price, an obscure court ruling — the model still produces something with the shape of an answer. Plausible is not the same as true.

2. AI is trained to guess rather than say “I don’t know”

In September 2025, OpenAI published research arguing that the way models are tested makes hallucination worse. Most benchmarks mark an answer right or wrong, and “I don’t know” scores the same as a wrong answer. So a model that guesses scores higher than one that admits uncertainty — exactly like a multiple-choice exam with no penalty for wrong answers. Models learn to bluff.

3. AI knowledge stops at a cut-off date

Every model has a training cut-off. Ask about anything newer — a price change last month, a product launched last week — and a model without web access has nothing to go on. Rather than say so, it will often fill the gap with whatever used to be true, or with something that merely sounds right.

What Do AI Hallucinations Look Like?

They cluster in predictable places. Be most suspicious when an AI tool gives you:

• Sources and citations — AI hallucinations are most common here: journal articles, legal cases and links that look real but do not exist.
• Exact numbers — statistics, prices, dates and percentages, especially very precise ones.
• Quotes — words attributed to a real person that they never said.
• Details about small or new things — niche tools, local businesses, recent events, minor public figures.
• Software features and pricing — plans and limits change constantly, and models often describe an old version, or invent one.

That last one is personal for us. Every price and plan in our reviews is checked against the vendor’s own pricing page before we publish, because AI-generated summaries of software pricing are so often out of date or simply made up.

Real Cases Where AI Hallucinations Cost Someone

These are not hypothetical. Three court cases show what happens when nobody checks.

Case	What the AI invented	What it cost
Mata v Avianca (New York, 2023)	Six court cases cited in a legal brief, all fabricated by ChatGPT — which then confirmed they were real when asked	The lawyers and their firm were fined $5,000
Moffatt v Air Canada (Canada, 2024)	The airline’s website chatbot invented a bereavement refund policy	Air Canada was held responsible for its chatbot’s answer and ordered to pay about C$812
Ayinde v Haringey and Al-Haroun v QNB (High Court, London, 2025)	Fake case law in two separate cases — in one, 18 of the 45 authorities cited did not exist	Lawyers referred to their regulators, and a formal warning to the whole profession

The UK judgment is the bluntest. The High Court said freely available AI tools “are not capable of conducting reliable legal research”. The same holds, in smaller ways, for any research you hand to a chatbot and do not check.

Which AI Tools Hallucinate Less?

No AI tool is hallucination-free, and anyone claiming to rank them precisely is guessing — rates swing with the topic and the question. What does make a real difference is whether the tool can look things up.

• AI tools with web search and visible sources — Perplexity, and ChatGPT or Claude with search switched on, answer from pages they have just read and show you where. You can click through and check. They can still misread a source, but you can catch it.
• AI tools working from your own documents — paste in a report or upload a PDF and ask questions about it, and the model has the facts in front of it. This technique, known as retrieval, reduces hallucinations considerably.
• AI tools answering from memory alone — any chatbot with search turned off. Fine for drafting, brainstorming and explaining well-known ideas; risky for facts, figures and anything recent.

Our reviews of ChatGPT, Claude and Perplexity cover how each one handles sources in practice.

How to Spot an AI Hallucination

• Click every source the AI gives you. Fabricated answers often come with a link that goes nowhere, or a paper whose title you cannot find anywhere else.
• Search any precise number or quote — if it only appears in the AI’s answer, treat it as made up.
• Ask the same question twice, or in a new chat — real facts stay consistent; hallucinated details often change.
• Ask how confident it is, and why — not proof, but a good model will often flag its own shaky claims.
• Be wary of answers that are too perfect — a citation that exactly supports your argument, a statistic that is a little too neat.

How to Reduce AI Hallucinations

You cannot switch them off, but you can make them much rarer.

• Give the AI the facts to reduce hallucinations — paste the source text, upload the document or link the page, and tell it to answer only from that.
• Give it permission not to know — add “if you are not sure, say so” to your prompt. It sounds too simple to matter; it helps more than you would expect.
• Turn on web search for anything factual or recent, and ask for sources.
• Ask for quotes, not summaries, when accuracy matters — “quote the sentence that supports this” is easy to verify.
• Break big jobs into small ones — long, multi-step answers give errors more room to compound.
• Keep a human check on anything that leaves your hands: legal, medical, financial, or published under your name.

Frequently Asked Questions

Can AI hallucinations be fixed completely?

Not with current technology. Hallucinations come from how language models work — predicting likely text — so they can be reduced, not eliminated. Web search, answering from your own documents and better training have cut them noticeably, but every major chatbot still produces them.

Does ChatGPT still hallucinate?

Yes. Newer ChatGPT models hallucinate less than early versions, and web search lets you check its sources, but OpenAI itself describes hallucinations as a fundamental challenge for all large language models. Treat any fact, figure or citation it gives you as a lead to verify, not an answer.

Why does AI make up sources?

Because to the model a citation is just another pattern. It knows what a real reference looks like — author, title, journal, year — and can produce one that fits perfectly without the paper ever existing. That is why fake citations are the most common, and most damaging, kind of AI hallucination.

Is an AI hallucination the same as the AI lying?

No. Lying needs intent, and the model has none — it does not know the answer is false. It is generating what sounds most likely. The effect on you can be the same, which is why checking matters more than the label.

Do paid AI plans hallucinate less?

Only indirectly. You are not paying for honesty: paid plans usually unlock newer models and better search features, and those tend to hallucinate less. A paid plan used without checking is still a risk.

What does “hallucination” mean in AI?

It is the industry term for an AI output presented confidently but false, or unsupported by any source. The word is borrowed loosely from psychology — the model “perceives” something that is not there. Some researchers prefer “confabulation”, which is arguably more accurate.