Page language: Italiano English Français Deutsch

Is artificial intelligence reliable? When it can get things wrong

By the FAIND newsroom · Updated on 17 October 2026

A woman holding an apple while her mirror reflection holds a pear

“Is artificial intelligence reliable?” and “can it make mistakes?” are two of the most searched questions. The honest answer: it is reliable for some things and not for others, and it does not warn you when it is wrong.

In brief

AI can make up facts, dates and quotations: these are called hallucinations.
It is more reliable on texts you give it than on facts it has to remember.
Checking always remains the user's job.

Why artificial intelligence can get things wrong

A language model does not consult an archive of facts: it generates the most plausible answer. When it does not know, it does not say “I don't know”: it builds a believable answer. This is what is called a hallucination. The most cited case is that of some lawyers sanctioned in the United States for filing documents with non-existent precedents, invented by a chatbot.

When it is reliable

It works well when it works on material you give it: summarising a document, rewriting a text, translating, tidying up notes, explaining a well-known concept. It also works better when it searches the web and cites its sources, because you can check them.

When it is not

It is less reliable on numbers, dates, quotations, names, legal provisions, recent news and complex calculations. And in everything to do with health, money and legal matters, where a mistake has consequences: there it can help you understand, but not decide.

Is artificial intelligence safe?

Safety also means data. What you write to a chatbot is processed on the company's servers and, depending on the settings, may be used to improve the models. It is better not to enter other people's personal data, confidential documents or passwords, and to check the privacy settings.

How often it gets things wrong: the numbers

Two studies help measure it. In October 2025, 22 public service broadcasters from 18 countries, coordinated by the EBU and the BBC, checked more than 3,000 answers about the news from ChatGPT, Copilot, Gemini and Perplexity: 45% contained at least one significant error, often in the sources cited. In April 2025 OpenAI stated that its o3 and o4-mini models made up information in 33% and 48% of answers to a test of questions about real people. The tests differ and models change fast, but the message is the same: on facts and news, mistakes are not rare.

How to check an answer

Ask for the sources and open them. Check numbers and quotations on an official site. Ask the question again in a different way, or ask another AI: if the answers change, there is reason to doubt. And remember that a well-written answer is not necessarily a right one.

Frequently asked questions

Can artificial intelligence make mistakes?

Yes. It can invent false information and present it with confidence, especially on facts, numbers and quotations. That is why answers need to be checked.

Can you trust ChatGPT?

For summarising, rewriting and explaining, yes, with a final check. For decisions about health, money and legal matters you always need a professional.

How reliable is artificial intelligence?

Very reliable when it works on texts you give it, much less on facts and news: in a 2025 European study, 45% of answers about the news contained at least one significant error. Use it with a final check.

Does artificial intelligence always tell the truth?

No. It does not lie on purpose, but it generates the most plausible answer, and when it does not know it can make things up with confidence. That is why it is worth asking for sources and opening them.