How to reduce AI Hallucinations

AI tools sometimes state things that are not true, in the same calm tone they use for things that are. You cannot switch that off, but you can lower how often it happens, and you can make the remaining errors far less likely to reach a customer. This article explains why it happens and then sets out the steps that make a difference in a small business.

It is built around seven checks. They cover the causes, the prompts and source material that help most, whether retrieval tools fix the problem, which tasks to keep away from AI, and a short routine your team can run without a quality department.

Key takeaways

Seven checks cut the number of wrong answers that reach a customer. Each one is covered in a section below.

1. Allow “I do not know”. Tell the AI it may admit uncertainty instead of guessing.
2. Restrict it to your documents. Tell it to answer only from the text you supply.
3. Ask for the quote. Make it show the passage behind each claim, and drop any claim it cannot support.
4. Compare more than one run. Different answers to the same prompt are a sign of guessing.
5. Check the sources. Keep your documents current, and open every citation.
6. Match the task to the risk. Keep AI on tasks where it reshapes text you give it, and keep a person on anything that commits the business.
7. Name who checks the result. A person reads it before it reaches a customer.

Even with all seven in place, hallucinations are reduced, not removed. The AI makers’ own guidance says as much.

What is an AI Hallucination?

A hallucination is an answer that sounds right and is wrong. The AI states a date, a price, a policy, a quotation or a source that does not exist, or misreports one that does, and it does so with no sign of doubt. Some researchers prefer the word confabulation, but the effect is the same: fluent text that is not backed by anything true.

The problem is not that the tool makes mistakes. People make mistakes too. The problem is that the mistake looks identical to a correct answer. A clumsy human draft often signals its weak points through clumsy writing. An AI draft does not. A made-up figure sits in a tidy sentence beside correct ones, and nothing on the page tells the reader which is which.

That is why the reduction work below has two halves. The first is making errors less likely. The second is making sure a person catches the ones that remain, because they will not go to zero.

Why do AI Models hallucinate?

A large language model works by predicting which words are likely to come next, given the words before them. It is very good at producing text that fits the pattern of a good answer. It does not look facts up in a database of things it knows to be true unless it has been connected to one, and so when it reaches the edge of what it learned, it keeps writing anyway.

A research paper from OpenAI and Georgia Tech, published in September 2025, adds a second cause that matters to anyone choosing or using these tools. The authors argue that language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty. Most benchmarks score a right answer as a point and both a wrong answer and “I do not know” as nothing. Under that scoring a model that always guesses beats one that admits doubt, so models are shaped into good test-takers rather than honest ones. The authors propose changing how benchmarks are scored so that confident errors cost more than admitting uncertainty.

For a business user, the practical lesson is simple. The default behaviour of most tools is to guess. Much of the work of reducing hallucinations is changing that default, through your instructions, your source material and your process.

Can you stop AI Hallucinations completely?

No, and any vendor who says otherwise should be asked for evidence. Anthropic’s guidance on reducing hallucinations, which lists the main techniques, closes with a plain warning: although they significantly reduce hallucinations, they do not eliminate them entirely, and critical information should always be validated, especially for high-stakes decisions.

The best independent evidence comes from law, where errors are costly and easy to measure. Stanford researchers tested AI legal research tools that were sold on the claim that they were built on real legal sources. They found the tools made fewer errors than general-purpose chatbots, but two of the products still gave incorrect information more than 17 percent of the time, and a third more than 34 percent. The same team’s earlier work found general-purpose chatbots wrong between 58 and 82 percent of the time on legal queries. Their conclusion was that even retrieval-based systems are not hallucination-free.

So the realistic goal is a lower rate of error, caught before it matters. Think of it as risk reduction, in the same way you would treat an unreliable but fast junior colleague: useful, supervised, and never trusted blindly with the final word.

How do you reduce AI Hallucinations with better Prompts?

Good prompting will not make a model accurate, but it changes what the model is trying to do. The first four checks are habits worth building into every prompt your team uses for factual work.

Check 1: give it permission to say it does not know. Anthropic’s guidance puts this first: explicitly allow the model to admit uncertainty, with a line such as “if the document does not contain the answer, say you do not have enough information”. Because models are shaped to guess, an explicit alternative removes the pressure to fill the gap. It is the single cheapest improvement available.

Check 2: tell it to use only what you provide. The same guidance describes restricting the model to the documents you give it and not its general knowledge. Write it as a rule: “answer only from the text below; if the answer is not there, say so”. The rule is far more effective than a vague request to be accurate.

Check 3: ask for quotes before conclusions. For long documents, the guidance recommends asking the model to extract word-for-word quotes first and then do the task using those quotes only. This forces the answer to rest on text that exists. You can also run the check afterwards: ask the model to find a supporting quote for each claim in its draft and to remove any claim it cannot support.

Check 4: compare more than one run. Anthropic’s guidance also suggests running the same prompt several times and comparing the results, since inconsistent answers can signal a hallucination. If three runs give three different figures, none of them should be trusted.

None of this needs technical skill. It needs a clear brief, which is the same discipline as when you brief an AI tool like a new hire. Once you have wording that works for a recurring task, save it in a prompt library so the whole team uses it rather than improvising. A common question is whether simply telling ChatGPT not to hallucinate works. It helps a little, but only when it is paired with the three concrete moves above.

How do you make AI use your own Documents?

The strongest protection against invention is to give the model the facts and tell it to stay inside them. This is called grounding. Instead of asking the AI what your returns policy says, you paste in the policy, or connect it, and ask the AI to answer from that text.

There are three ways to do this, from simplest to most involved.

Paste the text into the chat. For a one-off task this is enough. Paste the document, state the rule about using only that text, and ask your question. The limit is the context window, the amount of text the model can hold in view at once. If your document is longer than that, the model works from part of it and may fill the rest from memory.

Use a project or a saved set of files. Most business tools now let you attach reference files and standing instructions to a workspace, so every conversation starts from the same approved material. Pair this with custom instructions that repeat the core rules: use only the supplied files, quote the source, and say when the answer is not there.

Connect the AI to a document store. Larger set-ups retrieve the relevant passages from your files automatically before each answer. That is the subject of the next section.

Whichever you use, check 5 is to make sure your sources are right first. Grounding an AI in an out-of-date price list produces confident, well-sourced, wrong answers. The tool is only as reliable as the documents behind it.

Does Retrieval-Augmented Generation stop AI Hallucinations?

The approach called retrieval-augmented generation, usually shortened to RAG, is where the system first searches a set of documents for relevant passages and then asks the model to answer from them. It is the standard way to build an assistant that knows your own material, and it does lower the rate of invented answers compared with asking a model cold.

But it does not stop them. The Stanford study cited earlier found legal tools built on retrieval still produced wrong answers at a high rate. The authors attributed this to three problems. The search step can fail to find the right passage, which is hard when the answer depends on judgement rather than a single sentence. The passages it finds can be real but inapplicable, for instance from the wrong jurisdiction or the wrong period. And the model can be sycophantic, going along with a false assumption built into the user’s question.

Each of these has a business equivalent. A retrieval tool pointed at a messy shared drive may pull an old version of a policy. It may find the right page for the wrong product line. A customer who asks “since you offer free returns for a year, how do I return this?” may get an answer that accepts a premise the policy never made.

If you are weighing this against training a model on your own data, the trade-offs are set out in RAG vs fine-tuning. For most small businesses, retrieval over a tidy, current set of documents is the more practical and more controllable route. Treat it as a way to reduce errors, and plan for the review step anyway.

Which AI tasks are most likely to hallucinate?

Check 6 is to match the task to the risk. Risk depends less on the tool than on the task, and a rough sorting helps a team decide where to be relaxed and where to be strict.

Lower risk: reshaping text you provide. Summarising a meeting transcript, turning notes into a draft, changing tone, translating a paragraph, reformatting a table, or extracting named items from a document. The facts are in front of the model, and a reader can compare the output with the input. Errors still happen, such as a dropped caveat, but they are easier to spot.

Higher risk: recalling facts from memory. Asking for a statistic, a legal requirement, the terms of a competitor’s product, a historical date, or a list of sources. Here the model has nothing to compare against and will often produce something plausible. Treat any figure or citation that came from the model’s memory as unverified until a person has found the original.

Highest risk: anything that commits the business. Quotes, refunds, contract wording, compliance statements and replies that promise something to a customer. These are risky because of what a wrong answer costs, not because the model is worse at them. This is also where the line between automating a task and automating a decision falls, and where a person should stand behind the result.

A related point concerns tools that act on their own. An agent that takes several steps in a row can carry an early invented fact into later steps, so one error spreads. Keep such set-ups to narrow jobs with a clear stopping point, and see a plain-English explanation of what an AI agent is before you give one any authority.

How do you check AI Answers for Hallucinations?

Prevention narrows the problem and checking catches what is left. Check 7 is a person, and the habits that make that check work in practice are few.

Check claims, not style. Read for every statement that can be verified, such as a number, a name, a date, a policy or a quotation, and trace each back to the source document. Do not use the AI’s own summary of the source as the source. The longer guide to how to check AI output before it goes to a customer sets out a six-question order that takes a few minutes per item.

Open the citations. If the tool gives sources, click them and read the passage. A reference that looks complete can be wrong in the details or not exist at all. A well-known example came in 2025, when reports said a consultancy had agreed to partially refund the Australian government for a report that contained apparent AI-generated errors. Reporting on the case described a fabricated quotation from a court judgement and references to research papers that could not be found. A report that size had still gone out with the errors in it.

Match the depth of checking to the risk. Read every word of anything touching money, contracts, personal data, health or safety. Spot-check low-stakes internal drafts. A fixed sample of lower-risk items each day is enough to show whether the error rate is changing.

Keep a human in the loop at the point of action. The check has to happen before the message is sent or the record is changed, not after. Name the person responsible. A review step that belongs to nobody is the first thing to be skipped when the week gets busy.

Record what you find. A line noting what the reviewer changed and why shows you which kinds of error recur. If the same mistake keeps appearing, the fix is usually in the prompt or the source material, not in more checking.

What should a Small Business do about AI Hallucinations?

A small business does not need a programme. It needs a short set of decisions, written down, that the whole team follows.

Decide which tasks AI may do alone, which need a check, and which are off limits. Put it in a one-page AI use policy so nobody has to guess. The three risk groups above are a sound starting point.

Standardise the instructions. Write the “say you do not know” rule, the “use only the supplied documents” rule and the “quote your source” rule into your shared prompts and standing instructions, so they are applied every time rather than when someone remembers.

Keep your source documents current and in one place. An AI that is grounded in a tidy folder of approved material is far more reliable than one drawing on whatever it can find. Name an owner for that folder.

Be most careful where customers talk to the AI directly. The clearest warning is Moffatt v. Air Canada, a 2024 decision of a Canadian tribunal. A customer asked the airline’s chatbot about bereavement fares and was told he could buy a full-price ticket and claim the reduction afterwards, which was not the airline’s policy. The airline argued the chatbot was responsible for its own statements. The tribunal rejected that, found the airline had not taken reasonable care to make sure its chatbot was accurate, and ordered a partial refund of CAD 650.88 plus damages for interest and fees. The amount was small, but the principle was not: what your AI tells a customer is treated as what your business told them. If you are considering AI customer support, keep the AI to questions it can answer from approved text, and hand anything involving money or exceptions to a person.

Test before you trust. Before relying on any tool for a task, run 10 to 20 real examples from your own work through it and record every wrong answer. That gives you your own error rate, which matters more than any benchmark.

What to do next

Pick one task your team already does with AI and sort it into one of the three risk groups above (check 6). Add the four prompt checks to the instructions for that task (checks 1 to 4), point the tool at a current set of source documents (check 5), and name the person who checks the output before it goes anywhere (check 7). Run it for a month and note what the reviewer changes.

If you would like a second opinion on where checks belong in a specific process, a discovery call is a quick way to find out.

Sources

• Why Language Models Hallucinate: research paper by Kalai, Nachum, Vempala and Zhang, submitted 4 September 2025, arguing that models hallucinate because training and evaluation reward guessing over acknowledging uncertainty, and proposing changes to how benchmarks are scored.
• Reduce hallucinations (Anthropic documentation): the guidance on allowing “I don’t know”, extracting word-for-word quotes, citing and retracting unsupported claims, comparing multiple runs and restricting the model to provided documents, with the warning that the techniques do not eliminate hallucinations.
• AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries (Stanford HAI): the study of retrieval-based legal research tools, with error rates of more than 17 percent and more than 34 percent, and the finding that retrieval systems are not hallucination-free.
• AI Gone Wild: Airline Has to Honor a Refund Policy Its Chatbot Fabricated (Manatt): summary of Moffatt v. Air Canada, covering the chatbot’s statement, the airline’s defence, the tribunal’s reasoning and the award.
• Deloitte to partially refund Australian government for report with apparent AI-generated errors (Associated Press via CTV News): report of 7 October 2025 on the partial refund after apparent AI-generated errors in a government report.