AI: HALLUCINATIONS ON A CONVEYOR BELT

The “hallucinations” of Artificial Intelligence (AI) are the elephant in the middle of the room: everyone sees it, almost no one (except users) talks about it. And when someone does say something, the topic is quickly pushed out of public attention. After all, we wouldn’t want to stain the noble ideals of AI (seriously now).

“Yes, they exist!” – just not in any relevant way

What is actually happening? The fact is that AI, the great digital oracle whose creators claim it knows everything about everything, invents bibliographies with the same elegance with which a middle‑school student invents excuses for not doing homework. It also presents quotes from works that do not exist, falsehoods as truths, and many other things in the same paradigm. The result: it turns into a professional generator of science fiction. And what can we even say about those links that lead to real articles but have absolutely nothing to do with the topic—honestly, it’s downright poetic. A sort of “Yes, it exists! But it’s about something else entirely, good luck finding the connection!” And we can’t even blame it: AI seems perfect, presenting us fiction as real, verified facts, on top of everything! A special talent, one that humans know under the name “lying.” And another one that AI keeps perfecting continuously; soon enough, even politicians will have to learn from AI how to lie better. Which raises the question: can we say that AI has finally succeeded in copying the human mind?

When AI not only errs, but deludes itself

Artificial Intelligence “hallucinates” because of a design flaw: it is built to respond to our request in the most suitable way according to mathematical logic. Meaning that AI chooses what seems exact/true and what is mentioned most often in its database, in relation to our query. And when it doesn’t find exactly what it needs, AI picks whatever seems most suitable or most probable, even if it has absolutely no connection to the subject. Moreover, its database contains numerous incorrect, inaccurate, false, or context‑dependent pieces of information. A human would generally know how to navigate this chaos, relying on knowledge, life experience, logic, emotions, deduction, intuition. Not AI, which takes as true whatever appears to be true or whatever is labeled as such, and presents the information accordingly. And all this without any critical judgment. Likewise, if a quote is attributed to a certain work, AI will not check whether it is accurate or whether the work actually exists; it will simply, naively, believe it is true. In conclusion, this is how AI ends up inventing things, fabricating, “hallucinating.” Let’s recap: AI does not know how to distinguish what is true from what is false when this is not clearly indicated. It is not capable of understanding or handling nuance and subtlety. Therefore, it is “idiotic.”

After the hallucinations? Emergency meeting!

The architecture of the whole thing is impeccable. AI is “trained” to have authority, to create the impression that what it does and says is perfect and represents the absolute Truth. Consequently, we will probably soon see companies holding crisis meetings because AI recommended they purchase a non‑existent product. And since AI was so convincing, everything was planned in detail: budget, implementation, staff training, etc. After which the truth was discovered. Hence the legitimate question from the CEO: “What do we do when AI makes such spectacular mistakes?” Short (and cynical) answer: Publish an AI‑written article that goes viral and then discover that half the bibliography is a fairy tale. A complete invention.

Don’t take AI at its word!

First of all, don’t expect AI companies to fix this. They won’t. They would have to invest enormous amounts in copyright licensing to access high‑quality and (almost) reliable information. They would also need qualified employees to sort, properly organize, and continuously verify the information in the databases AI works with. And AI companies don’t have the money for that. So the safest path is for each user to verify the work provided by AI, either alone or through a specialized third party (assuming such services will appear).

Excerpt from the article “AI: artificial intelligence or artificial idiocy.”