What are AI hallucinations?

Walter Damián Cabrera, founder of LIA Engine · November 1, 2026

What does it mean when an AI «hallucinates»?

The word sounds odd, but the idea is simple. A large language model (LLM), like the ones behind ChatGPT, Gemini, Claude or DeepSeek, doesn't look the answer up in a database: it writes it, word by word, choosing what seems most likely. Most of the time, likely matches true. When it doesn't, the AI fills the gap with something that sounds right. That's called hallucinating, or confabulating.

Examples of AI hallucinations

These are the ones that show up most when AI is used for work:

• Calculations and totals: a tax calculated wrong, or a total that doesn't add up to its parts.
• Impossible dates: a due date on September 31, which doesn't exist.
• Data from documents: an amount misread from a photo or a PDF.
• Made-up sources: quotes, laws or links that don't exist.
• Ignored requirements: you asked for ten fields and got nine; you asked for five invoices and got four.
• Diagrams: overlapping labels, or an arrow coming out of the wrong box.

Every line is right except the last one. It looks tidy, it sounds sure, and it's wrong.

Why does AI make things up?

There are three reasons, and none of them is bad intent:

1. It writes what's likely, not what's verified. The model doesn't do the math like a calculator: it predicts which text comes next. A number «close to the right one» is just as likely to it as the right one.
2. It has no way to know whether it got it right. It can't check its answer against your data on its own.
3. It's trained to always answer. Saying «I don't know» is almost never the most likely response, so it fills in instead.

That's why errors show up mostly where the answer has to be exact: long calculations, data that has to be copied as is, and rules that aren't spelled out in the request.

Are hallucinations errors or lies?

They're errors. To lie you have to know the truth, and the AI doesn't: there's no intent. The danger is something else: it doesn't warn you. A wrong total looks exactly like a right one, with the same format and the same confident tone.

How often do AI models hallucinate?

It depends on the model and, above all, on the task. For writing, summarizing or brainstorming, they make few mistakes. For calculations, dates and exact data, many more. And it changes with every version. One example of our own: in our electronics test bench (in Spanish), an AI on its own got none of 24 tasks right; with LIA checking and correcting, it got 23, with no wrong answer certified. The honest way to know how much your AI hallucinates is to measure it with your own tasks.

Can AI hallucinations be fixed?

The honest answer has two parts:

• Eliminated completely? No. No AI is infallible, not even the newest ones. Be wary of anyone who promises otherwise.
• Kept out of your work? Yes. They go down with good prompts and by giving the AI the data instead of asking it to remember. And anything that can be checked with code (calculations, totals, data against your spreadsheet, requested fields, diagrams) can be verified and corrected before you use it.

What can't be checked with code, like an opinion or a fact about the world that isn't in your documents, still needs a person to review it. To compare options, see Is there a tool that fixes AI hallucinations?

How LIA Engine solves it

LIA Engine sits between your systems and the AI you already use. Your AI answers. LIA checks. It verifies every answer against your data and what you asked for, corrects whatever can be calculated, and delivers a signed certificate that anyone can verify. What it couldn't verify, it doesn't certify, and it tells you why.

It doesn't replace your AI or change it: it makes it trustworthy where it matters most.