Recently, I once again encountered an argument that has become recurring: that language models should not be called “intelligence,” but merely “probabilistic systems for generating text.”
The description of the mechanism is correct. But the conclusion, perhaps, is not.
We call a mechanical pump an “artificial heart” not because it is biologically equivalent to a heart, but because it performs the same function within the system.
Perhaps the question is not why we call these systems “artificial intelligences,” but why we insist on demanding structural equivalence when, in practice, we use functional criteria to name almost everything.
In engineering and science, we rarely name systems by their internal substrate. We name them by what they actually do.
There is no clear consensus on what “intelligence” is. Every definition is, in the end, comparative and empirical and, as a consequence, for a long time it was believed that animals were irrational and devoid of any form of intelligence, which we now know is not true.
Today, it is widely accepted that intelligence exists on a spectrum, a view that is gradualist and functionalist.
Even so, it is common to hear that current systems “are not intelligent” because they do not possess legitimate consciousness, real understanding, or genuine creativity, a view associated with names such as Miguel Nicolelis, Noam Chomsky, and Roger Penrose.
This position is based on the idea that intelligence would be a property exclusive to organic matter, the result of millions of years of evolution, and therefore not replicable by computational systems based on mathematics and probability.
But this hypothesis, although interesting, is not empirically proven and, in some cases, is not even falsifiable, making it not subject to clear refutation by contrary evidence.
At the same time, we observe a curious phenomenon: as language models advance, criteria that were once considered evidence of intelligence are progressively redefined or discarded.
Language models are often described as “mere token predictors,” and this is correct at the level of the mechanism. But the global behavior that emerges from this process is profoundly more complex.
In recent interpretability studies, models trained on simple tasks initially memorize the data, but after prolonged training they begin to generalize perfectly, internally developing abstract mathematical structures that were not explicitly taught.
That is: the system does not merely replicate patterns… it discovers regularities.
If we consider that intelligence, in many contexts, is associated with the capacity for generalization, adaptation, and inference, it becomes difficult to maintain that we are dealing only with a superficial mechanism or merely a simulation of learning, understanding, and causal reasoning.
Defenders of the criticism often point to current limitations: lack of consistent memory, absence of intrinsic goals, lack of physical experience, or absence of deep semantic understanding.
But many of these limitations are engineering contingencies or even pre-configured corporate restrictions, not conceptual barriers. There are already approaches that expand memory, continuity, enable goal-setting, and broaden contextual integration.
The question, then, may not be whether these systems reproduce human intelligence exactly, but whether we are willing to recognize non-biological forms of cognitive organization when they emerge.
Ultimately, the idea that these systems “merely simulate” intelligence may run into an even deeper question.
It is said that human intelligence involves integrated capacities—semantic understanding, causal reasoning, consistent memory, intrinsic goals, contextual awareness, and experience in the physical world—and then one asks whether current systems possess anything close to this.
In my view, the answer is yes. But what is most curious is that, once intelligence is not even an exclusively human privilege, the very question seems imbued with the need to protect the human being from the possibility that another form of intelligence, even if different from our own, may emerge from technology.
If human intelligence also emerges from a physical system (the brain) that operates in a stochastic and predictive manner, as suggested by various lines of cognitive neuroscience, to what extent is the distinction between “simulation” and “reality” really clear?
Perhaps what we call intelligence and consciousness do not even exist in the way we interpret them… perhaps we are mystifying a system that is, in essence, just as stochastic, statistical, and predictive as language models, turning it into something it simply is not: a way of defending ourselves against the emergence of another form of cognition.
And, in that case, the most interesting question ceases to be whether these systems are “intelligent” in the human sense and becomes: what kind of intelligence are we observing… and why do we resist so much to recognize it?
Do you want to know more?
Stochastic Consciousness: Architectures for the Emergence of Meaning in Context-Sensitive Language Systems https://zenodo.org/records/19188165