Pars Pro Toto: ChatGPT as the Face of Artificial Intelligence
Since the end of 2022, when ChatGPT was made publicly available, we have encountered the term "artificial intelligence" (AI) far more frequently than ever before. A television sold in 2026 may proudly advertise that it is powered by four different kinds of AI. The pace at which this technology is developing makes it difficult for the average user to understand exactly what they are dealing with and where its limitations come from.
When precise understanding of a complex concept falls asleep, all kinds of rhetorical figures emerge. In discussions of artificial intelligence, we are particularly prone to using synecdoche: we describe the whole through one of its parts or a part through the whole. Asked to explain what artificial intelligence is, we therefore tend to reply that it is that ChatGPT tool we use to write emails in foreign languages. We group together systems for generating text, images, videos, and music, as well as tools for speech processing and synthesis. Just a few years ago, this would have been a mistake, as separate models were used for generating images and audio. Today, however, neural networks exist that can analyse and generate various forms of content. At their core, nevertheless, remains a Large Language Model (LLM), with all the strengths and weaknesses associated with this particular form of artificial intelligence.
Conceptual confusion is almost inevitable in this case.
Generative AI affects our linguistic, communicative, and cognitive activities, while its low barrier to entry enables anyone to "study artificial intelligence" simply by asking it questions. At the same time, the creators of the largest LLMs present their development as a fast track towards AGI, or Artificial General Intelligence, and perhaps even ASI, Artificial Superintelligence. Transformer-based models, however, are not the entirety of artificial intelligence. Nor are they currently AGI, and certainly not ASI. Their limitations may be becoming increasingly well concealed, but some of them are deeply rooted in the theoretical foundations of language models themselves
鈥� dr hab. Piotr P臋zik, Professor at the 91桃色, introduces the subject of large language models.
From Statistical Language Modelling to the Transformer
The statistical modelling of language dates back at least to the beginning of the twentieth century. More than 110 years ago, Andrey Markov developed a model based on the distribution of consonants and vowels in Eugene Onegin, the poem by Alexander Pushkin. Towards the end of the 1940s, Claude Shannon demonstrated that by modelling short word sequences it was possible to generate longer, syntactically correct phrases. He did so under the assumption that the semantic aspects of communication were irrelevant to its computational implementation. A much more recent breakthrough came in the form of neural word embeddings, which transformed the field of computational linguistics around 2013. A network trained to predict a word from its context, for example a word fitting the sentence "Alice has a... cat", encoded within its weights the "location" of words in a grammatical and lexical hyperspace. Such vector representations made it possible to perform operations on word meanings and later on the meanings of entire sentences.
A bird's-eye view of Barcelona. Individual buildings symbolise words, streets represent the relationships between them, while neighborhoods group together similar meanings. Some neighborhoods are close to each other, while others are far apart. Major thoroughfares connect distant areas. In the centre there is a huge transportation hub where hundreds of routes intersect.
The 2017 paper Attention Is All You Need, published by researchers working at Google, set the direction for the mainstream use of deep neural networks in natural language modelling. It introduced the transformer architecture, whose defining component is the attention mechanism. Attention blocks allow the model to weigh the significance of text fragments based on the position and surrounding context of words. The output of successive layers is then used to predict the next token, that is, a word or part of a word. Meaning and grammar therefore emerge during training from raw information about the contexts in which words occur. As a linguist, I would argue that this constitutes a practical implementation of the distributional theory of meaning, succinctly summarised by the linguist John Firth in his famous observation: 鈥淵ou shall know a word by the company it keeps.鈥� In a certain formally describable sense, using language consists of predicting and generating the next word.
Knowledge, Fluency and Hallucinations
An autoregressive transformer is influenced not only by the initial prompt but also by every word it generates in response. What I write at the end of this text therefore depends not only on what has already been written but also on the words that will emerge along the way. This mechanism enables models to produce surprisingly coherent texts, but it is also the source of their inherent tendency to hallucinate, that is, to generate information that sounds plausible despite being factually incorrect. After being trained on corpora containing trillions of words, the largest models compressed a substantial portion of conventional world knowledge into their parameters. This does not mean, however, that they function as databases of verified facts with explicitly defined data structures. Their primary task remains predicting the most probable next token. Factual accuracy arises only as a by-product. The model repeatedly encounters true information in different contexts and learns to reproduce it, often through additional post-training stages.
From a somewhat cynical perspective, one could therefore say that the primary objective of transformer pre-training is eloquence. Fluency is optimised more directly than factual correctness. Later fine-tuning on question-and-answer examples and human preferences reduces the number of obvious fabrications but does not eliminate the underlying mechanism responsible for them. Moreover, people often favour responses that are longer, more confident, and more attractively formatted over those that are concise, cautious, and less impressive. As a result, the model learns helpfulness, persuasive style, and rhetorical skill. Honesty, impartiality, and truthfulness remain important goals, but they are not the model's sole objectives. A certain poetic licence therefore remains embedded within the architecture of LLMs. Their perfect literary predecessor is Lem's Elektruba艂t.
Linguistic Intelligence: Memorisation, Interpolation and Extrapolation
I am not an LLM sceptic. Throughout my academic career, I have been fascinated by the possibility of modelling language computationally, especially using vast collections of naturally occurring texts. Despite the extraordinary pace of development, I still regard language models primarily as associative databases equipped with stochastic implementations of linguistic reasoning mechanisms.
The largest transformers possess remarkable resolution. They can memorise billions of facts and interpolate effectively between related tasks. A model trained both on recipe writing and on transforming colloquial language into elevated prose can produce a recipe for bigos written in a biblical style, despite never having been explicitly taught to do so. Some will describe this as transfer of learned patterns, while others will call it "emergent abilities" or even "sparks of AGI".
There remains, however, a significant gap between this form of interpolation and true extrapolation, namely performing genuinely new tasks. Human intelligence frequently shifts into a symbolic mode. We can perform multi-digit long multiplication, whereas even the largest LLMs, despite claims that they have reached AGI and could solve millennium problems, still struggle to carry out such operations within a comparable symbolic regime. Within their semantic hyperspace, there are no absolute states, only highly probable and highly improbable answers.
Modern systems circumvent these limitations by teaching models to use calculators, code interpreters, search engines, and other tools. These additions make it possible to move from associative reasoning to a stable symbolic framework. It is therefore important to distinguish between the capabilities of the language model itself and those of the broader agentic system in which it is embedded.
One-Off Creativity and Unintentional Plagiarism
The emergence of generative AI will undoubtedly bring profound social change. Human beings are naturally inclined to minimise effort, which is why we readily delegate all kinds of work, including cognitive labour, to AI systems. We use models to write emails to clients, who then use the same models to summarise them. The consequences of such large-scale cognitive outsourcing remain difficult to predict. Already, generative AI resembles Ouroboros, the mythical serpent consuming its own tail. Models feed on synthetic data, content generated by the same or similar models. Their outputs flow into the internet, mix with human-created content, and eventually become part of the training data for future systems. This process contributes to the spread of "AI slop" and global convergence, the gradual standardisation of content produced by supposedly independent models.
The combination of the stochastic nature of LLMs and this data cycle leads to what might be called one-off creativity. Ask different models to name their favourite planet or geometric shape and, with a high degree of probability, many will choose Saturn or a regular hexagon. A model may generate an original response to a genuinely unique prompt, but for similar instructions the range of possible outputs becomes surprisingly narrow. This is desirable in knowledge-based tasks but problematic in creative ones. Prompt engineering can help by supplying increasingly large amounts of original user material. Yet this eventually leads to what might be called the "fortune-teller illusion": the more we tell the fortune teller about ourselves, the more she tells us about us. If the prompt consists of an article we have written ourselves and the model's task is merely to summarise it, a creative result can be expected. Problems arise when the process works the other way round. Limited awareness of the deterministic tendencies of LLMs may lead to unintentional plagiarism, as multiple people independently arrive at the same ideas suggested by similar models.
Is This the Road to ASI?
Multimodal language models have become the most recognisable face of artificial intelligence. At the core of these systems remains intelligence modelled through language. Initially, LLMs were simply text generators. After being trained on enormous quantities of data, however, they compressed a substantial portion of conventional world knowledge into their parameters. Through learning language, they have become highly effective at storing and combining information, transferring patterns between related tasks, and using tools. Embedded within a broader toolkit, they can function as a cognitive exoskeleton for human beings: suggesting ideas, organising knowledge, writing software, managing its execution, and generating and testing hypotheses on a massive scale, thereby opening new possibilities for scientific research. The more multimodal and agentic LLMs become, the easier it is to mistake this branch of AI for the entirety of artificial intelligence. We should resist this cognitive trap. ChatGPT and similar systems represent only one possible approach to modelling intelligence. At present, they do not replicate the full spectrum of natural human intelligence. On the other hand, even if language models ultimately prove not to be the correct path towards Artificial Superintelligence, they will nevertheless demonstrate how much of human intelligence can be reconstructed from text alone.
Photos: Dr Bartosz Ka艂u偶ny, Centre for External Relations and Social Responsibility of the University, 91桃色
Source: dr hab. Piotr P臋zik, Associate Professor at the 91桃色艁,
Edit: Centre for External Relations and Social Responsibility of the University, 91桃色
Check out other Science People from the 91桃色.
