We have been told that Artificial Intelligences “reason,” that next-generation algorithms flash moments of ingenuity, that we are now on the brink of a digital oracle. And yet, if we are to be truly honest, this science-fiction imagery we have built around them needs correcting. The truth is that much of the technological development of recent years—so magical it almost seems animated—rests on a single, enormous pillar: statistics. AI does not produce intelligent answers; it produces probable ones. Highly probable ones. Which inevitably raises the question: probable… based on what?
Think of a supermarket. Of how products are arranged. Of that odd, yet oddly reassuring, proximity between tuna and pickles. It is the result of a probabilistic calculation derived from thousands of silently scanned behaviors: an algorithm that learns, memorizes, and associates. Its strength is not intuition, but data gluttony. The same thing happens when you type “Leonardo da Vanci” into Google: the interface corrects you (and so does the text editor I am using to write this piece), with a digital raised eyebrow, because millions before you typed “da Vinci.” Correctness, in the end, is simply what is statistically most likely.
This mechanism—almost comical in its simplicity—becomes even clearer in the world of computational gaming. We like to imagine a chess-playing algorithm as something that reflects, evaluates strategies, and anticipates moves with glacial intelligence. None of this is true. Today, it is far easier to feed it millions of games and let it extract the statistically most winning combination of moves. That’s it. No strategic epiphany—just an immense leaderboard of accumulated probabilities.
It is at this point that we encounter the greatest paradox: AI models are, by their own creators’ admission, black boxes. Machines that devour data and spit out statistical verdicts, without even their programmers claiming to fully understand their inner workings. Does it work? Fine. The exceptions? Road accidents. Completely wrong answers? Social-media folklore. The illusion of intelligence remains intact, despite everything.
Meanwhile, this mountain of data—produced, refined, multiplied—must be stored somewhere. Where? In vast data centers that now spread across the planet like new cathedrals of our digital cult: aseptic environments kept at nineteen degrees, powered and cooled day and night, as energy-hungry as small cities. We have embarked on a path that is barely sustainable, scarcely virtuous, and difficult to reverse.
Our era seems to have embraced a cult of redundancy. A cult perfectly reflected in everyday life: from digital photography, where we take a thousand shots to keep just one, to the logic of “disposable data.” A society that does not choose, but accumulates; that does not select, but produces waste. Waste as a method. Redundancy as a philosophy.
Imagine, for instance, wanting to calculate the area of a triangle. Three seconds and a simple formula would suffice—base times height divided by two. Yet the statistical approach of AI is equivalent to examining twelve million school assignments to extract a reasonable convergence. An apparently functional madness. This is how even large LLMs (Large Language Models) operate: composing texts like probabilistic mosaics, selecting word after word the one most frequently found across the infinite plains of past data.
But every redundancy comes at a cost. Not only an environmental one—this ocean of silicon devouring energy—but also a cognitive one. We are building a self-referential model that risks collapsing in on itself: the much-feared Model Collapse. AIs beginning to learn from data generated by other AIs, in a degenerative cycle akin to biological inbreeding: generations of information increasingly sick, increasingly distorted, increasingly less human.
Technological redundancy thus risks producing cultural redundancy, turning knowledge into a smooth surface—without depth, without deviation, without intuition. Our actions—today shaped by experience, imagination, courage, and even error—risk becoming the billionth tile in a table of statistically likely choices. A meadow of blades of grass bending no longer to the will of the wind, but to the inertia of calculation.
And here, perhaps, lies the most delicate knot of all: not whether AIs “reason,” but whether we, by growing accustomed to their redundancy, will slowly stop doing so ourselves.
