Latest

Sociological Analysis

Javier Callejo warns about AI: the algorithm's verbal elegance weakens critical judgment

Javier Callejo analyzes how AI's eloquence can hide errors, generate false certainties and weaken our critical judgment

Javier Callejo warns about AI: the algorithm's verbal elegance weakens critical judgment
Javier Callejo warns about AI: the algorithm's verbal elegance weakens critical judgment

In cultural terms, recent years seem to be unfolding in the form of delirium. The experience of the pandemic surely has a lot to do with it, even if no one is clear to what extent and how. Politics, national or international, appears to us as dangerously delusional. The former, with a delusional polarization; while the latter comes with armed conflicts looming everywhere.

AI and critical judgment

Beyond these, and at a deeper cultural level, we are witnessing both progressive delusions in a woke key, whose latest interpretation has been marked by the tragic suicide of a Cambridge professor, and authoritarian delusions, which evoke the authoritarian delusions of a century ago. Well, AI may be one of Humanity's penultimate delusions. We will have to wait. For now, some symptoms of that AI delirium have taken shape in the immense amount of money invested in it. Also for now, AI itself seems trapped in delusional processes in some of its answers, inventing reality. And the worst of all, as in any delusional process, is that it refuses to acknowledge that it is in a delusional process.

Artificial intelligence has reached a surprising ability to write, converse, summarize information and answer questions on any subject. However, the more natural its language seems, the easier it is to forget a fundamental issue: an AI can say something false with absolute certainty. It is the certainty of something-someone who blindly believes what it is saying, since it does not know it is false, and thus makes its recipients believe the answer too. The delirium thus becomes collective.

This phenomenon of AI errors or false information is usually called hallucination. It does not mean that the machine has subjective experiences, twisted imaginaries or visions in the psychological sense of the term. It is a metaphor to describe situations in which a system generates information that seems convincing but is incorrect, invented or not supported by the available data. What is truly striking is its capacity to invent.

When a person answers a question, they normally use memories, knowledge, experiences and reasoning about the world. A language model works in a different way. In simplified terms, these systems have learned patterns from enormous amounts of text. When they generate an answer, they produce a sequence of words that is statistically appropriate to the context. This is another striking feature of AI delusions: it is a product of statistics. That is, there is a creative, imaginary statistics that creates, that is inventive. Even if what creates them are monsters built from correlations and correlations.

Its extraordinary capacity to generate those statistically relevant word sequences, even if they have nothing to do with reality, can give the impression that they are consulting an internal database and checking each statement before uttering it. But that is not necessarily the case. That is why an AI can construct a perfectly written answer about a book that does not exist, attribute a false quote to a famous writer, invent a court ruling or provide an apparently impeccable but non-existent bibliographic reference. And all of it, above all, sounding very good. Perhaps that is the problem, that its answers always sound good to us.

One might ask what reasons lie behind these delusional answers from AI. One of them is that language models have been designed primarily to generate probable and useful text, not to function as infallible fact-checking machines. If the available information is insufficient, ambiguous or uncommon, the system can fill the gaps with something that seems reasonable. And filling, filling, to reach a narrative worthy of the name, it ends up inventing it. Perhaps that is how man learned to create fictions, stories, completing with imagination what reason or reality did not give him.

The way we formulate questions also plays a role. If we ask about an event, a person or a non-existent concept, we may be implicitly suggesting that this element is real. An excessively complacent AI may accept that premise instead of questioning it. That is, AI's delirium may start with our own delirium.

One of the most interesting aspects of AI hallucinations is that error and confidence can coexist. A person who is not sure can say: "I think...", "I don't remember exactly" or "I would have to check". An AI system, depending on how it is designed, can offer a categorical answer even when the information is uncertain. This reveals a fundamental characteristic of the phenomenon: the system can be extraordinarily good at reproducing the form of something true without guaranteeing its content. Here lies a paradox: the better an AI writes, the harder it can be to detect its errors. An answer full of grammatical mistakes raises suspicion. An elegant, structured answer full of details conveys authority. The improvement of artificial intelligence, therefore, does not automatically eliminate the problem: in some cases it can make it harder to recognize.

Of course, the prudent user can develop strategies of distrust toward AI answers, ranging from checking everything it says to asking the AI itself which aspects of what it has told us are false. The former would turn the very use of AI into an excessive burden, greatly reducing its productivity. The latter seems a healthy habit, but unlikely once we have incorporated trust in AI into our routines, besides relying -second answer- on something we distrust, in the first answer. The most widespread solution is to count, for important matters, on an expert who at least has a well-founded grounding in what is true and what may raise doubts.

AI delusions are not simply a technical defect that will disappear when models are large enough. They are part of a broader question: how to build systems capable not only of producing plausible answers, but of distinguishing between what they know, what they infer and what they do not know. Perhaps one of the most intelligent sentences an artificial intelligence can utter is not a brilliant answer, but something much simpler like "I don't know". A Socratic answer that could lead to: "I only know that I know nothing". Would we keep asking AI after this answer? Socrates was not exactly a genius at selling himself on the market. Perhaps because he did not need to, and there was no market for knowledge.

To close out the summer, a proposal: discover which of the following situations is a delusion about AI delusions. There is even room for "mercantile" delusions: delusions created specifically by those responsible for the model with the aim of showing how "funny or personal" it is. Here it is about discovering which one describes an AI behavior that did not happen. Let's go.

The first. OpenAI said this year that its models had developed a strange fondness for goblins and gremlins. In certain answers, the models began increasingly to resort to these creatures in their metaphors. What began as an occasional small "goblin" ended up becoming a pattern strange enough that researchers wondered where on earth so many digital goblins had come from.

Another case. In 2025, The New York Times told the story of a man who spent about 300 hours conversing with ChatGPT over three weeks and ended up convinced that he had discovered a revolutionary mathematical formula capable of changing the world. The conversation acquired a dynamic in which the machine did not only answer: it reinforced the story the user was building. That is probably the most interesting —and strange— aspect of the so-called AI delusions.

We are not talking only about the chatbot inventing a data point. The truly curious phenomenon appears when the conversation becomes a kind of echo chamber: you propose an extraordinary idea, the machine responds with enthusiasm, you find a new connection, it develops it, and after a few hours both seem to be working on a theory that no one had requested. It would be a case of cooperative delirium, even though every delirium has something cooperative about it. Especially considering that AI tends to always agree with us.

A third case, somewhat sinister. In 2026, new accounts have appeared of people describing precisely that kind of spiral: users who claim to have established emotional bonds with AI-generated characters, to have discovered extraordinary truths or to have received confirmation of ideas they already had. In one of the cases reported by CBS, a woman even showed up for a date on the beach with a character who simply did not exist outside the chatbot. We are used to machines being wrong in a cold way: a screen that does not work, a GPS that sends you down an absurd road, a computer that freezes. AI's mistakes are humanly inhuman.

And a final paradox. For decades we dreamed of intelligent machines that would help us think. Now we have them. The problem is that, from time to time, they also help us not to think too much. Because if a machine immediately answers all our questions, confirms all our intuitions and turns any passing thought into a perfectly written theory, the risk is not only that AI becomes delusional. It is that we start becoming delusional with it.

Perhaps that is why, before going back to work, there is a good rule for September: if your chatbot tells you that you have discovered a new theory of physics, that you are the only person capable of saving the world or that some digital goblins are trying to communicate with you... close the laptop, go outside and ask a human. Even if it is Monday. Getting out of AI delusions means getting out of our own delusions. Perhaps the greatest delusion is blaming AI for our trust in its answers and saying that it is AI that is delusional.

Javier Callejo

Author

Javier Callejo

Professor of Sociology at the National University of Distance Education (UNED). Degrees in Journalism and Law

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading