It's Time to Talk About AI as a Science
AI works better than we understand it. Like the steam engine before thermodynamics, it needs to become a science: laws, measurements and verifiable predictions.
The Background Noise
Every week a new model comes out, a benchmark is beaten, an announcement promises to change an entire industry. On the other side, just as regularly, come the alarms: AI will replace us, deceive us, slip out of our hands. The two narratives seem to be in conflict, but they have something in common: they talk about artificial intelligence the way we talk about a product or a weather event. Something that arrives, that you buy, that you endure.
Almost always, a third way of talking about it is missing, and it is the oldest and most useful one: artificial intelligence as an object of study. Not as a technology to sell or a threat to fear, but as a phenomenon to understand, with the same method we used to understand falling bodies, combustion or heredity.
The thesis of this article is simple to state and harder to defend. AI has reached the point where it needs to become a science in the full sense of the word: a body of laws, measurements and verifiable predictions. Not because it sounds nobler, but because it is the only way to stop chasing it.
A Technology That Works Better Than We Understand It
There is a fact that people in the field know well and that is rarely discussed outside it: we know how to build systems that work far better than we know how to explain why they work. We know how to train a network with hundreds of billions of parameters, we know that it generalizes to data it has never seen, we know that at a certain scale capabilities appear that were not there at smaller scales. As for why, the answers are partial.
Classical statistical learning theory had taught us a prudent rule: a model with too many parameters relative to the data memorizes the noise and fails on new cases. Modern neural networks violate this rule systematically. In 2017 a group of researchers showed that the same networks that generalize well can also memorize labels assigned completely at random. The title of the paper was already a program: understanding deep learning requires rethinking generalization.
That same year, at an award ceremony at NeurIPS, Ali Rahimi told the community that machine learning had become a form of alchemy. It wasn’t an insult. Alchemy produced real results: metallurgy, medicines, glass, dyes. What was missing was a theory that said why one recipe worked and another didn’t. Rahimi’s point was this: we can keep going like this, but sooner or later the bill comes due.
Almost ten years later, the bill has arrived. These systems have entered the processes of companies, hospitals, public offices, schools. And the question “why does it behave this way?” is no longer an academic curiosity: it is a matter of reliability, of responsibility and, in many cases, of law.
What We Mean When We Say “Science”
The word must be used with care, because it is easy to abuse. Calling any activity that uses formulas and computers a science is a way of emptying it of meaning. It is better to say precisely what it would take for AI to truly become one.
A science, in the sense that matters here, has at least three characteristics. First: it identifies stable regularities, which hold beyond a single experiment, a single model, a single laboratory. Second: these regularities allow quantitative predictions, that is, they let us say in advance what will happen and by how much. Third, the one Popper taught us to consider decisive: the predictions can be proven wrong. A theory that explains everything after the fact and never risks anything beforehand is not a theory, it is a story.
This helps distinguish three activities that are often confused in public debate. Technique knows how to do: it knows the recipe. Engineering knows how to do it reliably and repeatably, with margins and tolerances. Science knows why, and therefore can say what will happen even in conditions nobody has tried yet. Today AI is a very advanced form of engineering resting on a still young science. The point is not to downgrade engineering, which has produced extraordinary results. The point is that engineering without science, at some stage, can no longer extrapolate.
Feynman put it in his own way: what I cannot create, I do not understand. With AI we find ourselves in the opposite situation, and it is a historically rare one: we know how to create something we do not fully understand.
First the Machine, Then the Law
If the situation seems anomalous, the history of science suggests the opposite: it is the norm. Very often technology arrives before the theory that explains it.
The clearest example is the steam engine. Newcomen’s engine had been pumping water out of English mines since 1712. Watt made it far more efficient in the 1760s and 1770s. For decades the Industrial Revolution ran on machines designed through experience, trial and error, and the skill of extraordinary craftsmen. Only in 1824 did Sadi Carnot publish his Reflections on the Motive Power of Fire, asking what the maximum possible efficiency of a heat engine was, whatever its design. From that question, and from the work of Clausius, Kelvin and others in the following decades, thermodynamics was born.
It is worth noting what changed. Before Carnot, people knew how to improve a machine. Afterward, they knew how far it could be improved at most, and why. The second law does not build engines, but it tells us which engines can never be built. This is exactly the kind of knowledge AI lacks today: not how to build the next model, but what no model will be able to do, at what cost, within what limits.
The same pattern shows up in the passage from alchemy to chemistry. Alchemists had accumulated centuries of effective procedures. What Lavoisier brought at the end of the eighteenth century was not new recipes, but the balance, the conservation of mass, a rigorous nomenclature. In a word: measurement. Chemistry was born when people began to weigh what they had previously only observed.
AI today is at a similar point. We have the machines. What we largely lack is our Carnot.
The signs that a science is being born
Saying that a complete theory is missing does not mean saying that there is nothing. In recent years, results have emerged that have exactly the shape you would expect from the beginnings of a science: robust empirical regularities, still waiting for a deep explanation. I will mention four, because each one illuminates a different aspect.
Scaling laws. In 2020 a research group showed that the error of a language model decreases according to fairly regular power laws as parameters, data and compute grow. Two years later, the work known as “Chinchilla” corrected the optimal proportions between parameters and data, showing that many models had been trained on too little data for their size. The interesting detail is the correction itself: a law that can be corrected is a law that makes predictions, and therefore a law that can be wrong. That is the typical behavior of a science, not of a fad.
Double descent. The classical curve said that as you increase a model’s complexity, the error on new data first goes down and then goes back up. Between 2019 and 2020, several papers showed that if you keep increasing complexity beyond the point where the model can interpolate the data perfectly, the error goes down again. It is a measurable, reproducible phenomenon that contradicts the initial intuition: the kind of anomaly from which new theories usually arise.
Grokking. In 2022 some researchers observed networks that, trained on small arithmetic problems, first memorize the examples without generalizing and then, much later in training, switch almost suddenly to the general solution. Learning, in other words, is not always gradual: it can have transitions similar to phase changes. The vocabulary of statistical physics is not a affectation here, it is a working tool.
Mechanistic interpretability. This is perhaps the line of work closest to the spirit of Lavoisier: open up the model and measure what happens inside. Specific circuits have been identified, such as the so-called induction heads that allow a model to copy patterns from its context. Researchers have studied superposition, the tendency of networks to represent more concepts than they have neurons, and developed tools such as sparse autoencoders to separate them. For someone like me, who works on attention architectures, this is the point where the question “how does it do that?” stops being rhetorical and becomes experimental.
None of these results is a theory yet. Taken together, they are what Boyle’s measurements on gases were before kinetic theory: data that demand an explanation and that constrain the possible explanations.
Why this is not only a matter for laboratories
One might think this is a discussion for specialists. It is not, and you can see it clearly in at least three areas.
Regulation. The European AI Act asks for risk management, accuracy, robustness and transparency in high-risk systems. These are reasonable requests. But accuracy with respect to what, robustness up to what point, transparency in what sense? Without measurable quantities and shared criteria, every rule tends to become a paperwork exercise: you produce paper, not guarantees. Serious regulation needs serious science underneath it, just as bridge safety codes need the science of structural engineering.
Companies. Anyone who introduces AI into a real company knows the problem well. A prototype works in the demo and behaves differently on real data. An updated model improves on one task and gets worse on another. The question a technical lead should be able to ask, “what happens if the data changes by ten percent?”, today gets an answer almost only empirically, case by case. Every project redoes from scratch experiments that a mature theory would make unnecessary. The cost of this gap is enormous, and largely invisible.
Education and public debate. AI is mostly taught as a set of tools: this library, that framework, that prompting technique. This is useful knowledge with a short shelf life. What lasts are the principles: why a model generalizes, what it means for a system to learn, which limits are structural and which are contingent. The same goes for public debate, which swings between enthusiasm and panic precisely because it lacks a vocabulary to say what is possible, what is probable and what is ruled out.
In all three cases the problem is the same. Without laws, every claim about AI is worth as much as its opposite, and whoever repeats it with more conviction wins.
Two objections to take seriously
A thesis like this has legitimate objections, and it would be dishonest to ignore them.
The first: AI is not a natural phenomenon but an artifact. It makes no sense to look for its laws the way we look for the laws of the planets, because we wrote the rules ourselves. The answer is that we wrote the code, not the behavior. We know every line of the training algorithm, and yet what emerges from billions of adjustments on billions of examples is written nowhere. It is a complex system built from simple rules, like a gas made of molecules or a market made of exchanges. Herbert Simon spoke of “sciences of the artificial” for exactly this reason: even what we build can have properties that must be discovered, not just designed.
The second: it is too early. The field changes every six months, and any law formulated today will be superseded by tomorrow’s architecture. That is a real risk. But it is also the argument that could have been made in Carnot’s day, when machines changed with every generation. The strength of a good law lies precisely in the fact that it does not depend on a particular design: Carnot’s maximum efficiency holds for every heat engine, including those that nobody imagined in 1824. If a regularity in AI holds only for one architecture, it is not yet a law. The task is to look for the ones that hold for all of them.
Finally, there is an implicit objection, rarely spoken aloud: that talking about science cools the enthusiasm and takes away some of the magic. It is true, and it is a good thing. Chemistry took the magic out of alchemy and gave us synthetic drugs. The price was acceptable.
Changing the question
If there is one thing I would like to remain from these lines, it is a change of question. Instead of asking “what can the latest model do?”, a question whose answer is valid for a few months, we could ask “what holds for any system that learns from data?”. It is a slower question, less spectacular, and far more productive.
It is the question I started from when writing Le leggi fondamentali dell’Intelligenza Artificiale, the first volume of a broader project, which not by chance carries the subtitle Verso una nuova scienza. The book does not claim to settle the matter: it tries to gather what can be said today with reasonable solidity and to organize it in the form of principles, with their limits clearly stated. I also use it in the classroom, because I believe that those entering this field today need principles more than they need instructions for use.
But the point is not a book. The point is an attitude to spread: treating artificial intelligence with the same seriousness, the same patience and the same willingness to be wrong with which we have treated every other great phenomenon we wanted to understand. Measure before proclaiming. Predict before explaining. Accept being proven wrong.
The steam engine changed the world before thermodynamics did. But it was thermodynamics that made us understand what we had in our hands, and how far we could push it. With AI we are more or less there. It is time to start talking about it as a science.
