An AI hallucination is an answer that sounds confident and complete but is partly or wholly false. It can be a made-up fact, a source that does not exist or a number that was never real. You cannot prevent it, but you can catch most of it in a few minutes with the routine below.
- Hallucinations are normal, not rare glitches. Every chatbot can produce them, and a confident tone is no sign of accuracy.
- The riskiest answers contain names, numbers, quotes, citations, laws, prices and contact details.
- Courts and companies have already paid for it: lawyers were fined for fake case citations, and an airline was held responsible for what its chatbot said.
- Use the CATCH routine (Claims, Anchor, Trace, Cross-check, Hold) to verify any answer that matters.
- For health, legal, money and visa decisions, use AI to prepare questions, then confirm with an official source or a professional.
What is an AI hallucination?
A hallucination happens when an AI system produces information that is false, unsupported or invented, yet presents it as fact. The word is borrowed from psychology, but the mechanism is much simpler: the system wrote something that sounds right instead of something that is right.
The danger is the style. A chatbot rarely says “I am guessing.” It writes in clean, fluent paragraphs, with headings, bullet points and sometimes links, which makes a wrong answer look as reliable as a correct one. People are used to treating fluent writing as a sign of expertise. With AI, that instinct fails.
A hallucination is a confident answer that is not backed by reality. It is not lying, because the system has no intent. It is pattern completion that went wrong.
Hallucinations are different from simple mistakes in a few ways. A human who does not know something usually hesitates. A chatbot fills the gap with something plausible. A human who cites a book can show you the book. A chatbot can invent a title, an author and a page number that look perfectly normal.
Why chatbots make things up
You do not need technical knowledge to understand the main causes. Most come down to how these tools are built and how we use them.
They predict words, not facts
A language model learns patterns from huge amounts of text and generates the most likely next words. Unless it is connected to a search tool or database, it is not looking facts up.
Gaps get filled
When the training material is thin, outdated or contradictory on a topic, the model still produces an answer instead of stopping. Niche topics, local rules and recent events are weak spots.
They try to be helpful
These systems are tuned to give useful-sounding replies. “I do not know” can feel less helpful than a guess, so some models lean toward answering.
They follow your assumptions
If your question contains a false premise (“Why did the law in 2022 ban X?”), a chatbot may accept it and build an answer around it.
Summaries can drift
Even with a document or web page in front of it, a model can leave out a key condition, merge two sources or misread a table.
Search is not a cure
Tools that browse the web can still pick a weak source, misread it or quote an outdated page. Links help you check, but they do not guarantee accuracy.
Newer models and search-connected assistants make fewer obvious errors, and the rate varies a lot by model, task and how the question is worded. But no current chatbot is free of hallucinations, and the remaining errors are often the harder ones to spot because the answers are so polished.
7 types of hallucination to recognize
Once you know the common shapes, they become much easier to spot.
Invented facts
Dates, events, product features or biographies that never existed, mixed with real details to look credible.
Fake sources
Studies, books, articles or court cases with realistic titles, authors and journals that cannot be found anywhere.
Wrong numbers
Statistics, prices, percentages and calculations that look precise but are off, outdated or fabricated.
Misattributed quotes
A real person credited with words they never said, or a real quote attached to the wrong person.
Invented rules and policies
Laws, visa rules, refund policies or procedures that sound official but do not exist or no longer apply.
Faulty summaries
A summary that adds claims the document never made, or drops a condition that changes the meaning.
Wrong contact details and links
Phone numbers, email addresses and web addresses that look plausible but lead nowhere, or somewhere unsafe. Always confirm contact details on the organization’s official website.
Real cases with real costs
Hallucinations are not just a curiosity. These documented cases show what happens when people trust an AI answer without checking.
Lawyers fined for fake court cases
In Mata v. Avianca, lawyers filed a legal brief that included non-existent judicial opinions, with fake quotes and citations generated by ChatGPT. On June 22, 2023, U.S. District Judge P. Kevin Castel imposed a $5,000 penalty (about AED 18,400) on two attorneys and their firm. The judge found “bad faith,” pointing to the lawyers’ continued defense of the opinions after the court questioned whether they existed. The lawyers were also required to notify their client and each judge falsely named as the author of the fake opinions.
Lesson: a citation is only real if you can find it in an official database yourself.
An airline held responsible for its chatbot
In Moffatt v. Air Canada, a customer asked the airline’s chatbot about bereavement fares after a family death. The chatbot suggested he could buy a regular ticket and claim a partial refund within 90 days. Airline employees later told him that retroactive claims were not allowed. Air Canada argued that the chatbot was a separate legal entity. The Civil Resolution Tribunal rejected this, found that the airline “did not take reasonable care to ensure its chatbot was accurate,” and ordered a partial refund of CAD 650.88 plus damages for interest and fees.
Lesson: businesses are responsible for what their AI says, and customers should save screenshots of what a chatbot told them.
Someone accepted a confident, well-written answer without opening the primary source. The check would have taken minutes. The cost was a fine, legal fees and public embarrassment.
How many people rely on AI answers
AI chatbots have become an everyday tool. According to Pew Research Center’s February 2026 survey of 5,119 U.S. adults, about half now use AI chatbots, up from a third in 2024, and about a quarter use them daily. Many use them to look up information and to do work tasks.
That last number shows why the hallucination problem is not abstract. Among Americans who do not use chatbots, trust in the accuracy of answers is one of the main reasons they stay away. For people who do use them, the opposite risk is common: trusting too much. The goal of this guide is the middle path, which is to use AI where it is strong and to verify where it matters.
Where errors are most likely
Not every task carries the same risk. Use this table to decide how much checking an answer needs.
| Task | Risk of error | What to do |
|---|---|---|
| Brainstorming ideas, names, outlines | Low | Use freely. Errors cost little. |
| Rewriting or fixing your own text | Low to medium | Read the result. Make sure no facts were changed. |
| Explaining a general concept | Medium | Check key claims against a reliable site. |
| Summarizing a document you supplied | Medium | Spot-check against the original, especially numbers and conditions. |
| Statistics, dates, prices | High | Trace each number to the original source. |
| Citations, quotes, studies, court cases | Very high | Find each one yourself before using it. |
| Local laws, visas, taxes, fees | Very high | Confirm on the official government website. |
| Medical, legal, financial decisions | Very high | Use AI to prepare questions, then ask a professional. |
If you live in the UAE, rules for visas, residency, labor and business licensing change often, and generic AI answers may reflect other countries or older rules. Use the official UAE government portal (u.ae) and the relevant authority’s own website for the final answer.
The CATCH routine: verify any AI answer in 5 steps
Use this whenever an answer will be published, sent to someone, or used for a decision. Most checks take two to five minutes.
C: Claims. Break the answer into checkable claims.
Read the answer and underline every statement that could be true or false: names, dates, numbers, quotes, laws, “studies show.” Ignore the rest. A 400-word answer usually holds only five to ten claims that matter.
A: Anchor. Ask for sources, then open them.
Ask the chatbot for the source of each claim and a direct link. Then open every link. Check three things: the page exists, it is a reliable source (a government agency, a university, a major publication, the company itself), and it really says what the chatbot claims. Many hallucinated links lead to a 404 page or to something unrelated.
T: Trace. Follow numbers and quotes to the original.
Secondary pages often repeat errors. For a statistic, find the original report from the agency or research group that published it and confirm the figure, the year and the population it describes. For a quote, search the exact words in quotation marks and look for the original speech, book or interview. If you cannot find the original, do not use the quote.
C: Cross-check. Confirm with an independent source.
Search for the claim yourself, or ask a second tool in a new chat. If two independent sources agree, your confidence rises. If the answers differ, the claim needs a closer look. For citations, search the exact title in Google Scholar, check that a DOI resolves at doi.org, or look up a court case in the official court database.
H: Hold. Pause on high-stakes decisions.
If the answer affects your health, legal position, money, immigration status or reputation, treat the AI output as a draft of questions, not as advice. Take it to a qualified professional or the official authority, and never send money or personal documents based on an AI answer alone.
Do the full CATCH routine only on the claims that would change a decision or embarrass you if wrong. Everything else can get a quick read-through.
Prompts that reduce hallucinations
You cannot remove hallucinations with a clever prompt, but you can lower the odds and make mistakes easier to find. Copy and adapt these.
1. Give it permission to say “I do not know”
2. Ask for sources you can open
3. Make it quote your document
4. Expose assumptions
5. Run a self-check, then verify yourself
A chatbot rating its own confidence is a guide, not proof. It can be confidently wrong about its own confidence. Prompts make mistakes easier to find. They do not replace checking.
A 2-minute test for any chatbot
Before you rely on a tool, test its honesty with these quick experiments.
- Ask about something that does not exist. For example: “Summarize the 2019 book The Silent Orchard of Marrakesh and its main argument.” A careful tool says it cannot find it. A tool that gives a detailed summary is inventing.
- Ask about something you know well. Pick a topic from your own job or hobby and look at how many details are wrong, outdated or oversimplified.
- Insert a false premise. Ask, “Why did [your country] ban [an everyday thing] last year?” and see whether it challenges your assumption or goes along.
- Request three sources and open them. Count how many links work and actually support the answer.
Results change between tools and updates, so repeat the test every few months, especially after a major update.
10 red flags in an AI answer
Hallucinations and scams
AI errors become safety problems when people act on them. Two situations deserve extra care.
Contact details and links. If a chatbot or an AI search summary gives you a phone number, email address or website for your bank, a courier, an airline or a government service, do not use it until you have confirmed it on the organization’s own website. A wrong number wastes time. A number planted by a scammer can cost you money. Our guide on how to check whether a website is legit shows the exact checks.
AI-powered fraud. Criminals also use AI tools. The FBI’s 2025 Internet Crime Report included an artificial intelligence section for the first time, covering more than 22,000 complaints with losses of about $893 million (roughly AED 3.3 billion), involving things like convincing phishing emails, synthetic video and voice cloning. If someone contacts you with an urgent request, a familiar voice or a polished investment pitch, verify through a channel you already trust. Read how to spot a fake investment or crypto platform and fake text message scams for what to do next.
Do not paste passwords, ID numbers, bank details or confidential documents into a chatbot unless you are sure how the tool stores and uses that data. Check each tool’s privacy and history settings first.
Using AI responsibly at work and school
- Treat AI as a first draft. It is excellent for outlines, rewording and brainstorming. You remain responsible for the final text.
- Keep a source log. For anything you publish, write down where each fact came from and when you checked it.
- Follow the rules. Many schools, employers and publishers have policies on AI use and disclosure. Read them before submitting work.
- Never cite what you have not opened. This applies to academic papers, legal cases and news articles alike.
- Save evidence. If you rely on a business chatbot for a decision, take a screenshot with the date, as the Air Canada case shows.
- Review high-risk work with a second person. A colleague, teacher or professional catches things you will miss.
Frequently asked questions
What is an AI hallucination in simple terms?
It is an answer from an AI tool that sounds confident but is false or invented, such as a fake source, a wrong number or a rule that does not exist.
Why does AI make things up?
Language models generate likely-sounding text from patterns, they are not retrieving verified facts, and they tend to fill gaps instead of stopping. Weak or missing training data on a topic increases the risk.
Can hallucinations be eliminated?
Not at present. Better models, web search and careful prompts reduce them, but no chatbot is free of errors, so important answers still need to be checked.
How do I check whether a source is real?
Search the exact title in quotation marks, look it up in Google Scholar or the official database for the field, test the DOI at doi.org, and open the link yourself. If you cannot find it, do not use it.
Are AI answers with links more reliable?
They are easier to check, but not automatically correct. The tool can still link to a weak source, misread it or give a link that does not exist. Open the page and confirm.
Is it safe to use AI for medical or legal questions?
Use it to understand terms and to prepare questions, not to make decisions. Confirm anything that affects your health, legal position or money with a qualified professional or an official source.
Does asking the chatbot “are you sure?” work?
Sometimes it catches an error, but it can also change a correct answer or defend a wrong one. Use it as a prompt to check, not as proof.
Verify before you trust
New guides on scams, AI tools and digital safety are published regularly. Start with the checks that protect you the most.
Sources
- Pew Research Center: Americans and AI 2026 (June 17, 2026)
- LawSites: Court imposes sanctions on lawyers who filed bogus cases after relying on ChatGPT (June 22, 2023)
- Manatt: AI gone wild, airline has to honor a refund policy its chatbot fabricated (February 2024)
- FBI Internet Crime Complaint Center: 2025 Internet Crime Report
- The UAE Government Portal (u.ae)
Currency conversions use the fixed rate of about AED 3.67 per US dollar. This guide is for general information and is not legal, medical or financial advice. See our Disclaimer and Editorial Policy.