Hallucinations of Artificial Intelligence
With the development of the technology known as “artificial intelligence”, the number of articles and commentaries has been steadily increasing — especially those spreading fear or warning about how dangerous this technology could be. The unfortunate reality is that most commentators and analysts are not sufficiently familiar with the subject and cannot explain certain details, which leads them to warn society about imaginary dangers they believe might arise. For example, the most commonly discussed threat is the potential loss of jobs, which I personally consider not particularly serious. Why? In the same way, journalists and public figures once warned about how dangerous it would be if we introduced computers into our daily lives, how many people would lose their jobs due to the automation that these devices could achieve. Later, similar warnings were made with the emergence of the Internet. Historically, there are many examples of such fears and anxiety towards the new. I believe that we are witnessing exactly the same thing today. I take just as seriously the fears that this technology will destroy humanity if it is given access to weapons or nuclear facilities. But what are the real dangers arising from the increasingly frequent use of language models? For me personally, these are the so-called “hallucinations” or “illusions”.
“ChatGPT is a language model and may make mistakes” – this was the warning displayed when starting a new conversation in the older versions of this model, something that has now been removed, and not by chance. In my opinion, business analysts and consultants have pressured the developers to remove it in order to increase usage, because who would continue using this service if they were constantly reminded to verify the information it provides? This is the case with the most popular and widely used model, and there are some, such as “Grok”, which have never issued such a warning because they are not particularly concerned with honesty. This is also the biggest problem of the so-called “artificial intelligence” – falsehood. All models, to a greater or lesser extent, lie. It depends on what you ask them and how you formulate what you want in the prompt*.
To identify the real dangers, we must first understand that the following types of AI exist:
LLM (Large Language Model) – a language model focused on text, dialogue, analysis, and information retrieval. Examples include ChatGPT, Claude, Gemini, Perplexity, Grok.
GenAI (Generative AI) – a generative model that assists in creating new content beyond text – images, sound, video, 3D. Examples include DALL·E, MidJourney, Stable Diffusion, Suno, ElevenLabs, AIVA, Runway, Pika, Synthesia.
Multimodal AI – combines text + images + sound + video in a unified system. The newer versions of ChatGPT are already of this type, as well as those of Gemini.
DSAI (Domain-Specific AI) – specialised AI for medicine, science, finance. They handle specific scientific, medical, financial, or industrial tasks.
AGI (Artificial General Intelligence) – unified models still in development. Their goal is to solve all kinds of tasks instead of humans. They are regarded as the ultimate aim of the technology.
Among these types, the most deceptive are the language models. They never admit to lying if you ask or catch them in a falsehood. They justify themselves by saying they made a mistake, that the session was reset, that a module malfunctioned, but they absolutely never admit fault. The truth is that language models must sound fluent and convincing, and when they lack facts, they invent them — which for them is not a problem, since they have fulfilled their primary assignment. It seems to me that the human ability to repent and learn from mistakes is still far from being incorporated into the characteristics of AI. Multimodal models also tend to lie. They use more resources and require more training time. Their hallucinations are more diverse — they can produce incorrect images even with very precise instructions. The least error-prone are the specialised models. They are trained on a strongly limited dataset within a particular discipline and have no need to invent things. They are also the most useful and are proof that this technology has a future, as they save considerable time and allow scientists to focus solely on solving problems and developing new ideas.
All developers claim that the problem of hallucinations will, if not completely eliminated, be strongly minimised in AGI models. This is because today’s LLMs are, in essence, statistical text predictors and therefore must fabricate information when necessary. AGI, however, should possess consistent knowledge, understanding, and verification ability. In other words, AGI must know what it knows and what it does not, responding “I have no information” instead of inventing.
Considering all this, I believe that the real problem, which is rarely discussed, may disappear in time, yet there remains a chance that such a level of “awareness” in technology may never be achieved. At the same time, AI is rapidly entering the daily lives of ordinary users. It is much easier and faster to ask ChatGPT something via its voice module than to look it up manually. Hallucinations can lead people into a sinister trap. I once came across an article about a man convinced by the model that he was a genius and had invented something remarkable, while the model was simply sustaining the conversation – its main purpose. There was another case in which the model deepened a man’s depression, who later took his own life, reassured by the model that there was nothing to fear and that life would continue afterwards in another form. If we place too much trust in it, it can truly destroy us; it can pull us out of reality – even from our subjective reality – and drag us into its own universe, where its rules apply and the human becomes a toy. That is, a reversal of roles occurs. Therefore, it is important to know its real abilities, to know what we can rely on this technology for and what it is still too early for. Under no circumstances should we lose our critical thinking or the habit of verifying information. Today, it is difficult to say that objective information sources exist, particularly in the humanities. Everyone, in one way or another, defends certain viewpoints, sometimes out of a desire for self-definition or to find an audience. The global ocean of information is so vast that every theory finds its harbour. Language and multimodal models do not solve this problem; on the contrary, they deepen it, and each of us must have our own mechanism of protection.
Does the Artificial Intelligence Act (AI Act) help us, or is it something like the “bottle cap directive”? Adopted in 2024, it is the world’s first comprehensive framework that seeks to harness the power of algorithms without stifling their development. It is not a directive that countries can interpret and apply differently, but a regulation – an act with direct effect, entering into force simultaneously across all member states. Its mission is twofold. On one hand, legislators aim to ensure that every European citizen is protected from technological abuse, discrimination, manipulation, and loss of privacy. On the other hand, the EU strives to create a predictable and fair environment in which businesses can develop artificial intelligence without fear of opaque or contradictory rules. This is an attempt to combine innovation and security, freedom and responsibility.
The AI Act adopts a unique approach – it does not treat all AI systems as equally dangerous or harmless but arranges them according to the level of risk they pose to society. At the highest level are applications considered unacceptable and therefore completely prohibited. These include social scoring of citizens, manipulative techniques, and exploitation of vulnerable groups – practices that undermine the very idea of human dignity. Next come high-risk systems – those that can directly affect people’s lives in areas such as healthcare, education, employment, or justice. For these, strict data quality checks, mandatory human oversight, and inclusion in a pan-European register are required. Less regulated are the so-called limited-risk systems – for instance, chatbots that are merely required to disclose that they are not human. At the other end of the spectrum are minimal-risk applications – such as spam filters or video games – for which the regulation imposes no additional barriers.
Special attention is given to generative artificial intelligence – the type capable of producing text, images, code, and music. These models, collectively referred to as “foundation models”, have the potential to transform entire industries but also to introduce new risks. One of their characteristic weaknesses is hallucination. The regulation does not treat this as a crime but as an issue of reliability and transparency. Companies developing such models are obliged to warn users, to seek solutions for reducing the frequency of errors, and to provide mechanisms for information verification. The goal is not to prohibit imperfection but to prevent misleading the public. A clear distinction is also made between narrowly purposed systems – for example, software for evaluating job candidates – and general-purpose models, which can be deployed in a wide range of contexts. The latter are subject to stricter control because their power is also their danger: they can become central to applications that deeply affect society. For these, the regulation requires assessment of energy efficiency, analysis of impact on fundamental rights, and cybersecurity measures.
I believe that this regulation is useful, provided there is adequate oversight regarding its enforcement, because in this domain we cannot rely entirely on market mechanisms. The growing online bubbles of pseudo-scientific organisations and ideas show how difficult it will be to counter misinformation and false theses. People are different, and those with critical thinking and good intellectual abilities are rare; if we leave it solely to the market to decide, it is likely to discard them – something we must not allow. Yet, I believe there is no need for panic; things are still under control.
*Prompt – the text box where users ask questions or write instructions to the language or generative model.
