A week ago, my enthusiastic friend forwarded me an article claiming that a famous computer scientist had used one of the most popular “artificial intelligence” models to solve an open problem in mathematics. According to the article, the scientist stated the problem and the “intelligence” proposed a solution. Knowing who the scientist was, I immediately told my friend that the story had probably been told backwards. Most likely, the scientist had fed the “intelligence” a large collection of papers on the problem, guided the exploration, and ultimately had the machine bring that elusive detail to light. The next day, my enthusiastic friend sent me another article about the achievement. It provided more detail and explained that the discovery had indeed followed the process I had described.

The more careful reader will already have noticed that I put “artificial intelligence” in quotation marks. The reason is simple: “artificial intelligence” does not exist, and perhaps it never will. The term is misleading. To understand why, we first need to define intelligence, the kind without quotation marks. Major dictionaries describe intelligence as the collection of our intellectual faculties, ranging from memory and imagination to judgment and abstraction. Yet their definitions contain a crucial element: the ability to understand, reason, and, above all, successfully deal with new situations by adapting previously acquired knowledge to them. Intelligence, therefore, implies a mental agility that turns interpretation into effective action when confronted with the unfamiliar.

The Origins

The definition of “artificial intelligence” has shifted over time. The term dates back to 1955, when prominent figures in information theory, including John McCarthy, Marvin Minsky, and Claude Shannon, proposed a two-month research project on the subject. Its premise was audacious: every aspect of learning or feature of intelligence could, in principle, be described precisely enough for a machine to simulate it. The irony is that, more than seventy years later, the debate over what constitutes this “intelligence” continues.

The choice of term was deliberate. It avoided direct conflict with names already in use at the time, such as cybernetics, automata, and thinking machines. At the same time, the new label brought all these fields together under a single conceptual umbrella.

The concept of “artificial intelligence” evolved considerably over the following decades. In the 1950s, the height of machine intelligence was associated with a computer’s ability to play chess. Within roughly twenty years, however, the task had become relatively straightforward: computers could calculate future moves and eliminate lines that led to defeat. The essential difference between a home computer and IBM’s Deep Blue, the supercomputer that defeated grandmaster Garry Kasparov, was how deeply each could search through the possible sequences of moves.

Other early developments included systems capable of solving small logic problems. In subsequent decades, “artificial intelligence” came to include expert systems, which combined collections of such logical problems. Handwriting recognition presented another challenge that helped shape the field’s evolution. In essence, all these manifestations of what has been called “artificial intelligence” amounted, and still amount, to statistical methods for pattern recognition.

Neural Networks

Beginning in the 1970s, one particular idea associated with “artificial intelligence” came to dominate scientific research: “neural networks.” I place the term in quotation marks because these networks are, in fact, purely mathematical algorithms. When represented graphically, however, their structure resembles the natural organization of neurons.

One remarkable characteristic of “neural networks” is their ability to “learn,” again in quotation marks. Because a network consists of a vast set of mathematical operations, manually adjusting all its parameters is virtually impossible. What we call “learning” is actually an auxiliary program that manipulates those parameters, seeking to adjust the network so that it produces the best possible response to a set of example data.

The word “learning” was gradually replaced by training, now without quotation marks, because it describes the process more accurately. For every new input, we adjust the network’s “knobs” so that its output comes as close as possible to the expected response in the example. This training process is purely mathematical and differs from human learning. It is a calculation that steers the network’s parameters toward reducing the difference between the actual and desired outputs. This property makes a neural network a universal approximator. Because of how these networks are constructed and trained, they often produce answers very close to reality, even for inputs they never encountered during training. This is not always the case, of course.

Training has always been based on refining parameters so that the network can select the correct response with increasing sophistication and precision, minimizing error relative to a reference answer or desired result. During the 1980s and 1990s, network architectures and training algorithms such as backpropagation advanced significantly. The computing power of the time, however, together with limited memory and data-storage capacity, imposed severe constraints. These structures were therefore applicable only to relatively narrow, well-defined problems, such as simple pattern recognition or low-dimensional classification tasks.

“Artificial Intelligence”

The true transformation came around the turn of the twenty-first century. Advances in semiconductor technology produced an unprecedented exponential growth in processing power, available RAM, and, crucially, data-storage capacity. This technological leap removed the bottleneck that had previously limited model complexity. Suddenly, neural networks could be scaled to address far more complex problems and operate on massive quantities of unstructured data, including text and natural language.

Large language models, or LLMs, emerged from this convergence of improved architectures, more efficient training algorithms, and massive computing power. Trained on trillions of words and petabytes of textual data, these models represent today’s state of the art in natural-language processing. They are also the systems most widely recognized by the public as “artificial intelligence.” ChatGPT from OpenAI, Gemini from Google, and Claude from Anthropic, among many others that continue to appear, are products of this advance.

These models are trained with two essential goals. The first is to extract context and intent from a text input, which is fundamental to answering complex questions. The second is to generate logical, syntactically correct sequences of words that reflect the context and intent they have identified.

The process works as a predictive system: the neural network evaluates the sequence of words already produced and determines the most likely word to follow, based on the frequencies and patterns found in its training texts.

It is crucial to understand that the primary training objective is not factual accuracy, but coherence with the context and intent. The accuracy of the generated content is largely a consequence of the quality and volume of the training data. For example, if the overwhelming majority of the training texts (say, ninety percent) contain the statement “two plus two equals four,” then encountering the sequence “two plus two” will lead the network to produce “equals four” as its continuation.

Compared with the training of modern LLMs, a human being achieves excellence in writing by mastering grammar and developing a unique style shaped by the body of texts absorbed during learning. The ability to retain and synthesize information that has been read or heard follows naturally from this process. A human being, however, can go further, using knowledge to create or conceive entirely new scenarios. Here lies the boundary between intelligence and artificial “intelligence.” This faculty of imagination is limited in LLMs, which can sometimes follow illogical paths, a phenomenon commonly called hallucination. LLMs also frequently struggle with simple logical questions unless they possess a specific mechanism, acquired through either training or an algorithm, for handling such edge cases.

Another troubling tendency in LLMs is to repeat and even mirror ideas. In simplified terms, this follows from training that reinforces the reproduction of what the model is shown. A common mistake is to treat these LLMs as “artificial intelligence,” encouraging their use in ways that generate confirmation bias for virtually any position.

To mitigate this, some models implement filters that insert “common-sense” information into their global context. Taken to its limit, however, this approach merely reflects the creator’s biases, producing LLMs that repeat sociopolitical and identity-based narratives. One notorious example is China’s DeepSeek, which emphatically denies what happened in the Tiananmen Square massacre.

The Present and the Future

As explained above, LLMs are not manifestations of genuine intelligence. They are statistical tools of extraordinary complexity and predictive power. Their central limitation is their inability to produce original reasoning, exercise judgment, or, ultimately, imagine independently of their training data. That does not disqualify them; on the contrary, it defines their proper field of application.

This boundary is less a barrier than an invitation. Once we understand that these models are pattern optimizers and language sequencers rather than thinkers, we can use them more powerfully and responsibly.

LLMs are undeniably here to stay. They represent an advance in humanity’s ability to process, synthesize, and generate vast amounts of linguistic information. The challenge ahead is not to force “intelligence” into something where it does not exist, but to integrate these tools alongside human intelligence. They help reveal in words what is not immediately apparent to us, identifying the elusive detail that can escape us even after dozens of revisions.

The true power of “artificial intelligence” lies not in its ability to imitate the mind, but in our ability to recognize it for what it is: a universal approximator that, in the right hands and with a clear understanding of its limitations, can catalyze human creativity, accelerate discovery, and transform access to information in ways we are only beginning to understand. It is not a replacement for intelligence, but its newest and most powerful aid.

One final note: yes, I used AI to revise this essay and make it more engaging. Tools, after all, are meant to be used!