AI hallucinations: why AI gets things wrong and how to verify its answers

AI hallucinations are answers that sound confident and plausible but are not true: invented facts, non-existent laws, wrong figures, links to sources that do not exist. They happen because a language model predicts the most likely continuation of a text rather than checking against a base of verified knowledge. The defence is simple but requires discipline: check everything checkable against the primary source, and use the model for drafts and structure, not as a source of truth.

The better AI writes, the harder its mistakes are to spot: the text is fluent, logical and confidently phrased. That is exactly why hallucinations are dangerous in business: a wrong figure in a report or a non-existent rule in a letter can cost you your reputation. Here is why it happens and how to set up verification.

Why AI gets things wrong

• It predicts rather than knows. A model is trained to continue text the way it is usually continued. When the plausible answer and the correct answer differ, it may choose the plausible one.
• Data goes out of date. A model has a cut-off date after which it knows nothing unless search is connected. Laws, prices, heads of organisations and addresses change.
• Rare topics. The less text about a subject in the training data, the higher the risk of error. Local specifics such as regional rules, local organisations and Uzbek terminology are exactly that case.
• Ambiguous questions. If a request can be read in several ways, the model picks one interpretation and answers it confidently.
• The urge to answer. Models tend to give an answer even when "I don't know" would be more honest. Unless explicitly allowed to say so, they fill the gaps.
• Long context. In large documents the model can mix up details from different parts or "lose" a caveat.

It is important to understand that hallucinations are not a bug that will soon be fixed but a property of the technology itself. Newer models err less often, but the problem does not disappear, so the habit of checking stays essential whatever tool you use.

Where errors are most dangerous

• Numbers of laws, articles and resolutions. A model may name a document with a plausible number and date that does not exist, or attribute the wrong content to a real one.
• Statistics and figures. Percentages, market sizes and user counts are classic territory for invention.
• Quotes. Words attributed to real people need particularly careful checking.
• Links and sources. A model can generate a web address or the title of a study that does not exist.
• Names, positions, dates. Who heads an organisation and when an event happened are frequent errors.
• Medical, legal and financial advice. Here an error can harm people directly.
• Calculations. A model can make arithmetic mistakes, especially in multi-step calculations.

Signs of a possible hallucination

Not every error is visible, but some signals should put you on alert:

• very precise figures with no source;
• references to documents you cannot find by searching;
• confident answers about recent events;
• suspiciously convenient facts that perfectly support your point;
• different answers to the same question when you ask again;
• vague wording where you expected specifics, since the model may be masking a gap.

A confident tone says nothing about whether an answer is correct. AI writes truth and fiction equally persuasively.

How to check answers: a step-by-step method

1. Break the answer into claims. List every fact, figure, name, date, link and conclusion. You can ask the model itself to produce this list.
2. Decide what is critical. Not everything needs the same scrutiny. A figure in a report for management matters more than a nice phrase in an internal email.
3. Check against the primary source. Laws against official publications, statistics against the originating organisation's data, quotes against the original speech or interview. A retelling in another article is not a primary source.
4. Open every link. Make sure the page exists and actually contains what the answer claims.
5. Recalculate the numbers. Repeat every calculation in a spreadsheet or calculator.
6. Compare with the document if one was the input. If the model summarised your document, check that every point of the summary really appears in the original.
7. Flag what you could not verify. An unverified claim is better removed or rephrased without specifics.

A worked example

Suppose the AI drafted a paragraph for a report: it mentions a regulation with a number and date, gives a percentage growth figure for an industry over a year, and quotes a representative of a relevant organisation. The review goes like this. Look up the regulation on the official legal information portal; if nothing matches the number and date, remove the reference. Look for the percentage in a publication by the organisation that collects that statistic; if you find only media retellings without the original source, do not use the figure. Check the quote against a video or transcript of the speech. In the end the paragraph contains only confirmed information and the wording becomes more cautious, and that is fine.

Checking with another AI model can help reveal contradictions, but it does not replace the primary source: two models can be wrong in the same way.

How to reduce the risk at the prompt stage

• Supply the sources yourself. Paste in the document text or data and ask the model to answer only on that basis.
• Allow it to say "I don't know". Add to the request: "If there is not enough information, say so; do not guess."
• Ask it to flag uncertainty. "Mark the statements you are unsure of and those that need checking."
• Require quotes from the source. When working with a document, ask for the fragment each conclusion is based on.
• Narrow the question. The more specific the request, the less room there is for invention.
• Use tools that search your own documents and return answers with links to sources. We described this approach in our article on AI in document workflows.

More on structuring good requests in our piece on how to write prompts for business.

A verification process for teams

When a whole team uses AI, verification should be part of the process rather than a personal habit:

• A four-eyes rule for anything going to clients, partners or the public.
• Checklists by content type: for a press release, a report, a commercial proposal, a social media post.
• A fact owner for each piece: the person who confirms that figures and links have been checked.
• An error log. Record the AI errors you encounter; it is the best training material for colleagues.
• A clear publishing rule: no figure, quote or link goes out without being checked against the primary source.

Common mistakes when working with AI answers

• Checking "by eye" only. Plausible text looks right until you verify it.
• Asking the model whether it is sure. It may confidently confirm its own mistake.
• Trusting links without opening them. A generated link may lead nowhere or to a different topic.
• Using AI as a news search engine without connected search and without checking dates.
• Keeping "nice" figures because they look good in a presentation. One invented number devalues the whole document.

Checklist before using an AI answer

• All figures checked against the primary source or recalculated.
• All links opened and containing the stated information.
• Names of laws, organisations and positions checked against official sources.
• Quotes confirmed against the original.
• Any document summary checked against the document itself.
• Unverified claims removed or rephrased.
• A second person has reviewed the material if it is going outside the organisation.

If you need templates for reports, presentations and checklists with room for sources and verification built in, the Dasta Creative team can help through our design support service.