There is a comfortable belief among business owners that AI mistakes are a temporary glitch, and that the next model will simply stop making things up. It is an understandable hope. It is also wrong, and believing it is how small businesses end up sending a customer a price, a policy or a legal claim that never existed.
Newer models are clearly getting better. But the reason AI invents facts is built into how it works, so the risk shrinks rather than disappears. This guide explains what an AI hallucination is, why it happens, how often it still happens in 2026 and the simple routine that keeps it away from your customers.
What is an AI hallucination?
An AI hallucination is when an AI tool states something false, invented or unsupported as if it were true. It might quote a law that does not exist, cite a report nobody wrote, give a wrong opening time or calculate a total incorrectly, all in the same calm and confident tone it uses for correct answers.
The word "hallucination" is borrowed from psychology, but the AI is not seeing things. A better comparison is a very eager new employee who hates saying "I don't know". Ask them a question they cannot answer, and instead of checking, they give you their best guess in a confident voice.
The danger is not that AI gets things wrong. People get things wrong too. The danger is that a hallucination looks exactly like a correct answer, so it slips past anyone who is not checking.
Why does AI make things up?
AI makes things up because a language model is trained to produce the most likely next words, not to look up verified facts. When it lacks reliable information, it still produces a fluent answer that sounds right. Training and testing methods have also tended to reward a confident guess more than an honest "I am not sure".
Think of a multiple-choice exam with no penalty for wrong answers. A student who leaves a question blank scores zero, while a student who guesses sometimes scores a point. Over thousands of questions, guessing wins. Researchers at OpenAI made this argument in a 2025 paper, which concluded that common evaluation methods push models to guess rather than admit uncertainty.
Three situations make hallucinations more likely:
--- Rare or local facts. An AI has seen far more text about New York than about a specific street in Sham Shui Po. Questions about Hong Kong regulations, local suppliers or your own company are exactly where it has the least to go on.
--- Precise details. Names, dates, prices, phone numbers, case references and web links are easy to get slightly wrong, and slightly wrong is still wrong.
--- Leading questions. Ask "Which section of the Employment Ordinance covers X?" and the AI will tend to name a section, even if no section fits.
How often do AI models hallucinate in 2026?
AI models in 2026 hallucinate noticeably less than earlier versions, but the rate depends heavily on the task and is far from zero. For OpenAI's GPT-6.1 Sol, launched on 29 September 2026, reported factual error rates fell from 11.4% to 7.7% at low effort, roughly a one-third improvement.
Independent testing tells a more sobering story on hard questions. On its AA-Omniscience knowledge benchmark, Artificial Analysis found that GPT-6.1 Sol's hallucination rate fell from 60% to 54%. That figure measures how often a model gives a wrong answer, rather than admitting it does not know, on questions it cannot answer correctly.
Model makers publish these figures because accuracy is now a selling point. That is good news for buyers, but every figure comes from a specific test set, and none of those test sets is made of questions about your shop, your staff or your contracts.
Those two numbers are not in conflict. They measure different things:
--- Everyday questions: on common topics, a modern model is right most of the time, and errors are the exception.
--- Questions beyond its knowledge: when a model does not know, it still guesses more often than it admits uncertainty.
For a business, the second number is the one to remember. The questions that matter most to you, about your prices, your customers and Hong Kong rules, are often the ones the AI knows least about.
What can an AI hallucination cost a business?
An AI hallucination can cost a business money, reputation and legal standing, and courts have generally held the business, not the software, responsible. Three well-documented cases show the pattern: an AI states something false, a person relies on it, and the company that deployed the AI pays for the result.
The airline chatbot
In 2024, a Canadian tribunal ordered Air Canada to compensate a passenger after its website chatbot invented a refund policy for bereavement fares. The airline argued the chatbot was responsible for its own words. The tribunal disagreed, ruling that the company is responsible for everything on its website. The amount was small, around CA$800, but the precedent was not.
The lawyers and the fake cases
In 2023, a New York court fined two lawyers and their firm US$5,000 after they submitted a legal brief citing court cases that ChatGPT had invented. The cases had realistic names, dates and quotes. None of them existed.
The consulting report
In 2025, Deloitte Australia agreed to refund part of an A$440,000 government contract after a report it delivered was found to contain fabricated references and a made-up quote, produced with help from AI.
A Hong Kong SME is unlikely to face a headline like these. The everyday equivalent is quieter: a wrong delivery promise on WhatsApp, an incorrect MPF figure in a staff letter or a supplier comparison built on a price that was never quoted. Each one costs time, trust or money to fix.
What do people get wrong about AI hallucinations?
The biggest misconception is that a confident, well-written answer is more likely to be correct. Tone tells you nothing about accuracy. A hallucinated answer is usually just as fluent and detailed as a correct one, and sometimes more so, because the AI fills gaps with plausible detail.
Four other myths are worth dropping:
--- "If it gives a source, it must be true." AI tools can invent sources, or cite a real page that does not say what the AI claims. Click the link and read the relevant line.
--- "Asking 'are you sure?' fixes it." Sometimes the AI corrects itself. Sometimes it apologises and swaps one wrong answer for another. It is not a reliable check.
--- "Paid versions do not hallucinate." Stronger models make fewer errors, as the GPT-6.1 Sol figures show, but no model on the market has eliminated them.
--- "It only matters for facts." Hallucinations also appear in calculations, summaries of documents you provided and translations, where the AI may add or drop a detail.
How can you reduce AI hallucinations in your business?
You can reduce AI hallucinations by giving the AI your own reliable information, allowing it to say "I don't know", asking it to show where each fact came from and keeping a human check on anything a customer will see. None of these steps needs technical skills, and together they remove most of the risk.
Here is a six-step routine any small team can adopt:
--- 1. Give it the source. Paste or upload your price list, policy or contract and tell the AI to answer only from that material. This is called grounding, and our guide to AI grounding explains it in more detail.
--- 2. Give permission to not know. Add a line such as: "If the answer is not in the document, say you cannot find it." This simple instruction reduces guessing noticeably.
--- 3. Ask for the exact line. Ask the AI to quote the sentence it relied on. If it cannot point to one, treat the answer as unverified.
--- 4. Check the five risky details. Before anything goes out, verify names, numbers, dates, legal references and links. These are where hallucinations hide.
--- 5. Keep a human on customer-facing work. Let AI draft replies, quotes and letters, and let a person approve them, at least until you have a track record.
--- 6. Test with questions you already know. Ask your AI setup ten questions with known answers every month and count the mistakes. Our explainer on AI evals shows how to turn this into a habit.
Frequently asked questions about AI hallucinations
Business owners usually ask whether hallucinations will disappear, which tasks are safest and who is responsible when AI gets something wrong. The short answers are no, routine drafting on your own material, and your business.
Will AI hallucinations ever disappear completely?
Most researchers expect them to keep falling but not reach zero, because the models generate likely answers rather than retrieve guaranteed facts. Plan for a lower error rate, not an error-free tool.
Does letting AI search the web stop hallucinations?
It helps, because the AI can check live pages instead of relying on memory. It does not solve the problem. The AI can still pick an outdated page, misread a table or blend two sources into one wrong claim, so the links still need a quick look.
Do AI tools hallucinate more in Chinese than in English?
Often, yes, for detailed or local questions, because far less training text exists about Hong Kong businesses and regulations than about larger English-speaking markets. Cantonese expressions and Traditional Chinese place names are also common sources of small errors.
Does a "thinking" or higher-effort mode reduce hallucinations?
Usually a little. Reasoning modes give the model more time to check its own work, which tends to catch simple slips. They are slower, cost more on some plans and are still not error-free.
Which business tasks are safest to give to AI?
Tasks where you supply the facts and a person reviews the result: drafting emails from your notes, summarising a document you uploaded, reformatting a list or brainstorming ideas.
Who is responsible if an AI tool gives a customer wrong information?
In the cases decided so far, the business that deployed the tool. Treat AI output as if a junior staff member wrote it under your company name. This is general information, not legal advice.
Conclusion: trust the speed, check the facts
AI hallucinations are not a bug that the next update will remove. They are a side effect of how language models work, and in 2026 they are rarer but still real, especially on local, specific and high-stakes questions. The businesses that benefit most from AI are not the ones that trust it blindly. They are the ones that give it good sources and check what matters.
Build that habit once and AI becomes a fast, dependable assistant instead of a confident guesser. UD stands with you, making AI human.
Reviewed by the UD AI team.
Getting value from AI without the made-up answers comes down to setup: the right sources, the right limits and a human check where it matters. Talk to us about AI staff built around your own documents and rules, and we will walk you through it step by step, from the first task to a review routine your team can keep.
