Take a person, put them twice in the same situation down to the microsecond, and watch. Intuition says they won't do the same thing. Neuroscience, with data, also says they won't. And yet the courts, education and half of clinical psychology still run as if they would. Current AI, on the other hand, is textbook deterministic with a pinch of cosmetic randomness on top. The paradox is entertaining until you realise which of the two errors lands on which.
The noise isn't a flaw, it's the material
Let's start at the bottom, where it hurts less to argue and where the evidence is cleanest. The nervous system isn't deterministic. It isn't because the machinery that makes it up isn't. Neurons fire spontaneously, with no stimulus to justify it. The synapse releases neurotransmitters in variable amounts in response to the same action potential arriving — the ion channels (the molecular gates through which the ions that generate the electrical signal pass) open and close with probabilistic dynamics, not by clock.
Faisal, Selen and Wolpert published in 2008 in Nature Reviews Neuroscience a synthesis that has become a required stop: neural noise isn't a defect evolution couldn't polish away, it's a functional property of the system. It allows exploration, generalisation, escaping local optima, staying plastic. To remove it would be to break it.
If you measure the response of a single neuron in the visual cortex to the same point of light repeated a hundred times, the response changes each time. It changes in latency, it changes in number of discharges, it changes in temporal pattern. Not so much that you see the point somewhere else. But enough that the inner microstructure of your perception is, each time, a different perception. What we call "seeing the same thing" is an abstraction the brain fabricates late, after aggregating a great deal.
The raw material is never the same.
The decision that arrived late
Go up a level. Put the subject before a free decision: move your finger whenever you want. It's Libet's experiment, 1985, repeated and discussed for forty years. The preparatory brain activity is measured — the so-called readiness potential (a ramp of electrical activity that precedes voluntary movement) — and compared with the moment the subject reports having decided. The preparatory activity starts about three hundred and fifty milliseconds before the subject feels they decided. Consciousness arrives late, commenting on something that was already happening.
Wegner built on that, in The Illusion of Conscious Will (MIT Press, 2002), the thesis that the sense of agency is an after-the-fact reconstruction. A comment on the act, not its cause. Maoz and colleagues, in a modern replication published in eLife in 2019, refined the picture. They distinguished arbitrary decisions and deliberate decisions, and showed the pattern is more complex than Libet had posed. But the underlying idea doesn't come apart: what consciousness calls "my decision" is the visible tip of a chain whose root isn't accessible to introspection.
And, it's worth insisting, that chain isn't a classical deterministic chain. It's a stochastic chain.
Determinism, free will and a comfortable misunderstanding
Sapolsky, who doesn't deal in subtleties, took the argument to the extreme in Determined: A Science of Life Without Free Will (Penguin, 2023): there's no point in the causal chain of a human decision where you can lodge anything like free will, and therefore the category is surplus. Dennett, in Freedom Evolves (Viking, 2003), had tried to salvage a compatibilist version — a free will redefined as the capacity for reflective evaluation in a causally determined system. The argument between the two positions fills shelves.
What's interesting for what concerns us here isn't that metaphysical debate. It's a consequence almost nobody underlines. Even if Sapolsky is right and everything is determined by the causal chain, the causal chain includes noise. The correct physical description of the brain isn't that of the Newtonian domino where, given state A, state B inevitably arrives. It's a stochastic description (governed by probabilistic processes, where the same initial state can lead to different later states), in which the same input produces output distributions, not single outputs.
Determinism and non-determinism aren't the opposite of free will. They're two different physical descriptions of the substrate. What matters for everything that follows is that the real description of the brain is the second, and that isn't much seriously disputed. It's ignored because it's uncomfortable.
A courtroom built on sand
Land this in a court. Modern criminal law presumes that a person, under normal conditions, is responsible for their acts because they could have acted otherwise. The "reasonable man" criterion supposes a constancy: if A enters situation X, A produces behaviour Y. If they produce Z, they're committing a deviation attributable to their character, their will or their negligence.
Replace the presumption with what we know.
A enters X, and A produces a behaviour within a probability distribution determined by their biology, their history, their internal state and the system's noise. Tomorrow, in an equivalent situation, A would produce another point of that same distribution. Possibly a very different one. With no conscious decision intervening to justify the difference. On what exactly does responsibility fall? If we judge one point of the distribution as if it were the whole distribution, we're judging a variable as if it were a constant.
I'm not writing a plea to abolish criminal responsibility. The attack on that notion has its own cost, and it's enormous: a society that renounces imputing acts to people falls apart fast. But it's worth not confusing the soft pavement with the solid building that stands on top of it. The law works, to a large extent, because society needs it to work. Not because its anthropological model is correct.
The exam, the test, the interview
Education operates with the same fiction. An exam presumes that two students with the same knowledge will produce comparable answers, and that the same student examined twice would produce near-identical scores, with a small margin of error attributable to the instrument. The reality is that the variability of the same subject on cognitive tests — measured in test-retest studies at short intervals — is perfectly capable of moving a grade from an A to a B, or the other way, for reasons that have nothing to do with what was learned.
The student isn't an accumulator of knowledge that discharges cleanly when pressed. They're a system whose disposition to retrieve a content depends on fluctuations they don't control and the examiner can't read.
Clinical psychology works with the same problem and knows it. Diagnoses based on self-administered questionnaires whose answers change with the day. Personnel selection built on interviews and tests whose test-retest reliability is notoriously poor. We keep using it. We have nothing else. But it's worth not confusing the tool with the reality it claims to measure.
And now the part where the noise was ours, not theirs
An artificial neural network in inference (the phase in which the model, already trained, generates the response to an input) is deterministic. Feed it the same input tensor, fix the seed (a number that initialises any random-number generator so it always produces the same sequence) and the same floating-point operations, and it returns exactly the same output. Bit for bit. With no variability. It's applied mathematics, not biology.
When a large language model seems to produce different answers to the same question, it isn't because there's noise in the system. It's because explicit randomness is introduced at the last step, in the sampling. The temperature (a parameter between zero and infinity that controls how much the probability distribution is "flattened" before choosing the next fragment of text), that dial the user manipulates from outside, isn't an emergent property of the model. It's a control over the entropy with which the next token is chosen — the unit of text, a word or fragment of a word — over a distribution the model has computed in a completely deterministic way.
Holtzman and colleagues analysed it in The Curious Case of Neural Text Degeneration (ICLR 2020). Nucleus sampling, top-k, top-p are techniques to keep the greedy choice of the most probable token from producing degenerate or repetitive text. They're generation tricks. The underlying function is still the same deterministic machine beneath.
A die on top of a calculation isn't a brain
That surface randomness resembles neural noise as little as a six-sided die resembles the solar system. The brain's noise is distributed across the whole architecture, modulates every intermediate operation, is coupled to the system's temporal dynamics, interacts with global states — attention, vigilance, cortical oscillations — and is functional: the system uses it to explore, to escape local optima, to generalise, to stay plastic.
A language model's noise, by contrast, is in the last layer, is optional, the user switches it on, is decorrelated from everything else, and changes nothing of the internal calculation. The intermediate representation is identical across all samplings. Painting the two things as equivalent is like saying that having a twitch in your eye makes you the equal of a complex dynamical system.
The contradiction the industry hasn't resolved
This weighs on the industry, even if it doesn't say so. Language models produce answers too consistent for human sensibility, which detects the pattern and gets bored or distrustful. And at the same time they produce answers too inconsistent for professional uses, where bit-for-bit reproducibility is wanted. Human-conversation variability is asked of it when you talk to the system. Machine reliability is asked of it when you integrate it into a process.
The system can't give both, because it doesn't have them where it would need to have them. Human noise isn't a random sampling applied at the end over a clean distribution. It's a structural property of the system that computes. Imitating it by adding more temperature is the error of someone who copies a painting by painting only the surface and leaving the easel intact.
The day AI resembles us
What's being asked of artificial intelligence, with growing insistence, is that it be like us. That it hesitate. That it change. That it explore. That it not answer the same question the same way twice. Supposing it's achieved one day — not by raising the temperature, but by architectures with fluctuating internal states, coupled temporal dynamics and functional noise integrated into the substrate of the computation — what will have been built won't be a more useful system.
It'll be a system more like a human. With everything that implies.
Less predictable. Less auditable. Less controllable. Less of a contractual guarantee. The industry sells "agents" and at the same time tries to preserve the deterministic reliability of the calculation, and that's a contradiction the market hasn't yet faced head-on. The more the agent resembles a human, the worse it can be put into a service contract. The better it can be put into a contract, the less it resembles a human. There's no comfortable middle point.
The lesson isn't about AI
The reason current AI is so different from a human isn't only that it lacks a body, or consciousness, or a life. It's that it's deterministic in the substrate, and we aren't. Our non-determinism isn't a defect evolution failed to polish. It's the very condition of our being able to think as we think, learn as we learn and decide as we decide.
The perfect consistency of a machine was never an advantage over the erratic human being. It was the trademark of something that isn't alive.
And we still keep designing courtrooms, classrooms, human-resources departments and psychology manuals assuming the opposite on the human side, while spending billions trying to get the machine to stop behaving like a machine. If you step far enough back, there's something comic in the symmetry. If you look at it from inside the courtroom, the classroom or the personnel director's office — somewhat less so.
Definiciones
Neural noise. The spontaneous variability of nervous-system activity, present in neuronal discharges, in the synaptic release of neurotransmitters and in the opening of ion channels. It isn't a system flaw but a structural trait, and current neuroscience considers it functional for exploration, generalisation and plasticity.
Readiness potential. A ramp of electrical brain activity, recordable on the scalp, that precedes a voluntary movement by hundreds of milliseconds and appears before the subject reports having decided to move. It's the central datum of Libet's experiment.
Determinism. A property of a system in which an initial state and fixed rules always lead to the same later state. Applied to a computational model: the same input always produces the same output, bit for bit, if no explicit randomness is introduced.
Stochastic system. A system governed by probabilistic processes, where the same initial state can give rise to different later states with a certain probability distribution. The brain fits this description; a large language model, except in its final sampling layer, doesn't.
Inference. In machine learning, the phase in which an already-trained model is used to generate responses to new inputs. It's the phase in which the user interacts with the system.
Temperature. A parameter of generative language models that regulates how much the probability distribution is flattened before choosing the next token. Temperature zero produces deterministic greedy output; high temperatures increase variability and, eventually, incoherence.
Token. The minimal unit of text a language model operates with. It can be a whole word, a fragment of a word, a punctuation mark or a sequence of characters.
Nucleus sampling, top-k, top-p. Families of sampling techniques applied in text generation to avoid the degenerate or repetitive output of greedy sampling. They restrict the set of candidate tokens to the most probable before sampling over that subset.
Compatibilism. The philosophical position holding that free will, redefined as the capacity for reflective evaluation, is compatible with a causally determined universe. Daniel Dennett defended it against incompatibilist positions such as Robert Sapolsky's.
Referencias
Faisal, A. A., Selen, L. P. J. and Wolpert, D. M. (2008). Noise in the Nervous System. Nature Reviews Neuroscience, 9, 292-303. Central reference for the section on neural noise as a functional trait and not a defect.
Libet, B. (1985). Unconscious cerebral initiative and the role of conscious will in voluntary action. Behavioral and Brain Sciences, 8, 529-566. Classic experiment cited when discussing the readiness potential and the lag of consciousness relative to the onset of the decision.
Maoz, U., Yaffe, G., Koch, C. and Mudrik, L. (2019). Neural precursors of decisions that matter — an ERP study of deliberate and arbitrary choice. eLife, 8. Modern replication of the Libet paradigm with the distinction between arbitrary and deliberate decisions.
Wegner, D. M. (2002). The Illusion of Conscious Will. MIT Press. Source for the thesis of conscious agency as a reconstruction after the act.
Dennett, D. (2003). Freedom Evolves. Viking. Compatibilist position cited in the contrast with Sapolsky.
Sapolsky, R. (2023). Determined: A Science of Life Without Free Will. Penguin. Radical incompatibilist position cited in the section on determinism and free will.
Holtzman, A., Buys, J., Du, L., Forbes, M. and Choi, Y. (2020). The Curious Case of Neural Text Degeneration. ICLR 2020. https://arxiv.org/abs/1904.09751. Reference for the discussion of nucleus sampling, top-k and top-p as sampling techniques in language models.
También te interesa
- States of mind. No test measures the same person twice
- How the brain computes. Twenty watts and a consciousness
- The emotion-reason pairing. AI is only the rider
No comments yet
No comments yet. Be the first.
Leave a comment