What is an AI hallucination, and why do answers err?

An AI hallucination is an engine stating something incorrect in a confident tone: instead of saying it does not know, it gives you a definition, a date, a figure or a reference with nothing behind it.

The companies that own the engines say so on their help pages. OpenAI says ChatGPT can produce incorrect or misleading outputs and may sound confident even when it is wrong, calling this a hallucination; its examples include incorrect definitions, dates or facts, and fabricated quotes or references. Google says Gemini can hallucinate and present inaccurate information as factual. Anthropic says Claude can write things that might look correct but are very mistaken.

It is one of the terms in the Tasadr glossary. The part that matters to you is that the error can land on your business's name, price or location, in front of a customer asking before they buy.

In a hand check of twenty answers to buying questions in the UAE, asked on 11 October 2026, we checked every answer against its sources and found a note on seven of them. We do not attribute these errors to any particular engine:

• A stale figure, in two answers: a Sharjah hospital-stay price taken from a news article published in 2016, and a café-breakfast price built from a 2019 hotel offer and fast-food menus.
• A vendor's claim presented as customers' praise, in two answers: a "95% satisfaction" figure taken from the software firm's own page, and agencies described as "trusted by Gulf brands" on the strength of the agencies' own sites.
• A name mix-up, in one answer, which attributed one rental firm's name to another firm's site.
• A price in another country's currency, in one answer, which quoted a Sharjah resort in another country's currency, taken from a booking site's edition for that country.
• One source making the whole answer, in one answer about the best medical centres in Abu Dhabi, whose five sources all came from a single directory, and whose first pick was an ambulance company.

Because it writes from one of two places: what the large language model learnt from earlier text, or what it finds on the web when asked. The first may not know what changed at your business recently; the second takes what it finds, current or old.

Google says in its developer documentation that grounding Gemini models' answers in Google Search reduces hallucinations by basing responses on real-world information. That is what grounding means, and it rests on retrieval-augmented generation.

But grounding does not correct the source itself. Anthropic says the quality of Claude's responses depends on the sources it references, so if the page it found is old, or the whole answer rests on one source, the error passes into the answer as it is.

Nobody edits an engine's answer directly, but you own the sources it draws on. Write the facts about your business on a current, clear page with its update date, and in the same form everywhere your business is written about:

• Figures with dates: prices, hours and offers, with the date of the last update, so the engine does not take last year's figure.
• Claims with their source: if you cite a satisfaction rate or an award, say who measured it and when, because an answer may pass your claim off as your customers' opinion.
• One name everywhere, in Arabic and Latin letters, so the engine does not confuse you with someone else; the idea is in what is an entity in search?
• Currency and country beside every price, and a page per country if your prices differ, as in how do you lead AI answers in more than one Arab country?

Four terms that explain where an answer comes from, and where an error enters it:

• Grounding: an engine tying its answer to content it finds on the web at the time of the question.
• Retrieval-augmented generation: a system retrieving pages first, then writing the answer from them.
• Citation: the link that tells you where a piece of information came from, so you can check it yourself.
• Knowledge Graph: Google's database of facts about people, places and things.