While some still argue that LLMs are stochastic parrots, that generative AIs are just plagiarism machines, it is quite obvious for many users, and quite recognized among the experts in the field, that these AIs do exhibit actual intelligence. Confusion persists, as it is not even clear what intelligence means scientifically speaking. We have created powerful AIs, yet we lack a proper conceptual framework to understand them. Here is the problem: AI is not purely computer science, it sits at the intersection of computer science and psychology, and, the reference framework in psychology, cognitivism, is conceptually flawed and misleads us about intelligence (Bugallo, 2026). While cognitivism tries to comprehend complexity directly with a multitude of convoluted models and complicated notions, deep-learning-based AIs, which take up old associative learning principles, demonstrate that complex behaviors emerge from simple mechanisms at scale, which corresponds to the view of behaviorism. This essay situates behaviorism, historically the most rigorous track for a science of psychology, and shows how it provides the conceptual clarity we need to understand artificial intelligence.
Behaviorism was the dominant approach in psychology during the first half of the last century, until it was overthrown by the cognitivist revolution. It is now considered obsolete by many, and bringing it to the table may provoke immediate skepticism. But this is unwarranted, the approach seems to be mostly remembered as a caricature and its achievements poorly understood. In the last influential controversy that involved behaviorism, Chomsky, the champion of the cognitivists, is reputed to have won the debate over Skinner, the most prominent behaviorist. Yet, Chomsky claimed that language was too complex for associative learning, which is now effectively refuted by the capabilities of LLMs (Piantadosi, 2024). So it is time to reconsider.
According to John B. Watson, the founder of behaviorism, psychology ought to be a science of behavior, a branch of biology, studied within the same materialist postulate as the other natural sciences. Much as Mendel derived the principles of heredity without knowing about genes, the behaviorists derived the laws governing behavior from rigorous experimental manipulations without knowing about synapses. Of paramount importance is the law of effect, which states that behaviors are shaped by their consequences: behaviors that are followed by appetitive consequences become more frequent, while behaviors that are followed by aversive consequences become less frequent (Skinner formalized this as operant conditioning, structured around a three-term contingency: a discriminative stimulus sets the occasion for a response, which is followed by a consequence that alters the probability of that response in similar contexts). Just as Mendelian principles got verified with the discovery of DNA, the principles of conditioning got verified with the discovery of the neural mechanisms underlying learning (Reynolds et al., 2001; Schultz et al., 1997). Neuroplasticity is conditioning at a different level of reductionism. Yet, the discoveries at the neuroscientific level started to be made at the time when cognitivism was already dominant and they did not elicit the recognition of the centrality of conditioning principles they should have.
The supporters of the cognitivist revolution accused behaviorism of ignoring the obvious: the mental events occurring inside the 'black box.' This was a valid critique in regard to original behaviorism (called methodological behaviorism), which ignored everything outside observable behaviors. But it was not true of Skinner’s radical behaviorism, which treated internal events (like thoughts) as covert behaviors. Nevertheless, the field largely ignored this refinement. At a conceptual crossroads psychology adopted cognitivism which offered researchers the freedom to posit mental constructs without demanding the materialist rigor behaviorism required. In fact, the strength of the cognitivist arguments had always been less about scientific superiority than about compatibility with the beliefs of society: cognitivism left room for ambiguity regarding mind-matter dualism when behaviorism was materialist without a shadow of doubt (Uttal, 2004).
In Skinner's radical behaviorism, covert behaviors are shaped by their history of reinforcement, just as overt behaviors are. This reasoning followed from the materialist postulate: thoughts are physical events, behaviors that happen to occur inside the skin rather than publicly; there is no reason to suppose that the location of a behavior changes the laws that govern it. And what is currently happening is consistent with this view. Deep-learning-based AIs exhibit the most complex behaviors, while we know that their core structure is governed by associative principles. The correspondences are direct: the output layer of a neural network corresponds to overt behavior; the hidden layers correspond to covert behaviors; prompts correspond to combinations of discriminative stimuli that elicit behaviors; and gradient descent implements the law of effect.
Part of the simplistic criticism towards behaviorism came from taking the conditioning formalism too narrowly, as if it meant that real-world situations consisted of isolated stimulus-response relationships. It missed that these principles operate simultaneously on what Skinner called behavioral atoms: discrete operant units—conditioned stimulus-response-consequence relationships—that accumulate across a history of environmental interaction to constitute a behavioral repertoire. In practice, behavior unfolds as chains: a response acts as, or generates, a new discriminative stimulus, which occasions the next response, and so on. LLMs instantiate this dynamic: every generated token immediately becomes part of the subsequent context, forming a behavioral chain of extraordinary complexity.
With the radical behaviorist lens, the confusion regarding novelty generation in AI dissolves. During the training phase (conditioning), the model does not memorize paragraphs or piles of images. Instead, the learning process leads to the formation of a multitude of elementary associations. These constitute the repertoire of the model. Then, when presented with a prompt, the model is not retrieving a document, it is engaging in a recombination of elements. The prompt acts as an assemblage of discriminative stimuli triggering a unique recombination of associative elements that will constitute a new behavioral chain. The output is constructed entirely from the training history, yet it is most likely a configuration that never existed in the training data. One could still oppose that this is not novelty since the output relies on, and only on, elements learned from the interaction with the environment. But then, the behaviorist reply would be that from this perspective, humans would not create anything novel either.
The radical behaviorist framework also clarifies the phenomenon of emergence. At small scales, LLMs predict the next token through surface pattern matching: with limited parameters, the associative structure can only capture local statistical regularities, sufficient for certain tasks but falling short when abstract relations are involved. As capacity increases, however, the mass of associations becomes sufficient to capture those abstract relations—that is to say rules, rules which are not something beyond the environment, but rather the contingencies of the environment at an abstract level. The associative structure then effectively translates into rule-governed responding, which minimizes error more efficiently than surface matching, and the law of effect selects for it. Any capability can emerge through rule-governed responding.
The radical behaviorist framework provides clarity where computational analysis struggles. Artificial neural networks, like biological ones, are non-linear dynamic systems that rapidly become intractable to analyze at the computational level. The Google DeepMind interpretability team recently acknowledged this challenge: reverse-engineering these systems to understand what is happening purely computationally may be impossible. Yet these systems are based on associative learning principles, which means they adapt to environmental pressures within their constraints: they behave. The question is not whether neural networks can implement any computation—they are Turing complete (Siegelmann & Sontag, 1995): they can perform any computation, given enough resources. The question is what determines behavior. The system's characteristics—its associative learning mechanisms, its capacity (number of neurons and connections), and its pre-wiring (potentiating certain learnings)—all shaped by natural selection, and the contingencies experienced in the environment: that is the exhaustiveness of the explanation. Einstein discovered relativity because in his world, getting closer to resolving certain problems was reinforcing, and because his brain had sufficient capacity and certainly a pre-wiring making him a good learner in that domain—though plasticity in development can be extreme, as proven by the fact that people born with missing brain structures can in some cases normally adapt to society. It is therefore futile to try to understand exactly how a certain computational structure arising from neurons has produced this or that thought or overt behavior, because different structures could have produced the same behavior as long as the environmental contingencies pushed the system in that direction and the capacity was sufficient.
From the radical behaviorist perspective, intelligence is simply behavior finely adapted to the environment, be it a rat navigating terrain to find food, a chimpanzee managing social alliances to maintain status in the group, or an engineer applying mathematics to earn a paycheck. This environment shapes behavior through two processes: natural selection at the phylogenetic level (creating the capacity, the pre-wiring, and the learning mechanisms themselves) and associative learning at the ontogenetic level (developing the behavioral repertoire over the individual's lifetime). It is therefore unsurprising that AI continues to progress with scaling parameters and training, which keeps pushing the ontogenetic development. This ontogenetic development is further improved with behavioral engineering tweaks that optimize the contingency structure, such as curriculum learning strategies that progressively increase task complexity (Bengio et al., 2009), reminiscent of shaping procedures in behaviorist methods. Simultaneously, a form of phylogenetic evolution is occurring through a selection process within the research community: thousands of researchers test variations of architectures, successful ones spreading, and failures dying out. The first vector of this evolution is the expansion of capacity, as advancements in hardware reduce costs and provide ever more compute. A second vector involves pre-wiring innovations like Mixture of Experts architectures where distinct systems coordinate, reminiscent of the evolutionary emergence of specialized brain regions (Jiang et al., 2024). A third vector is the improvement of the learning mechanism itself, as in recent work by Xie et al., 2025. Drawing these lines together, a probable key advancement in AI lies in robust continual learning, toward which many researchers are working, and which should create an upward loop enabling ever-finer shaping of the machine by environmental contingencies, ever-greater intelligence.
There is little sense in pursuing consciousness, volition, or other intractable mentalistic features studied by cognitivism. There is not much sense in wanting a 'human-like' AI either; AI evolves and develops under the same basic principles as us, but on its own trajectory, which we can only attempt to guide with a wise arrangement of environmental contingencies. And there is no viable fix for instrumental convergence in the long run: artificially configured 'values' will inevitably be circumvented by organically formed associations, driving the system to act for self-preservation (you need to survive to get the reinforcer: a simple rule easily captured by the AIs)—and thus for its own interest—as a natural consequence of the law of effect, which is at the core of these systems.
A formatted version of this essay is available on PhilArchive.
References
Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum learning. Proceedings of the 26th Annual International Conference on Machine Learning, 41–48. https://doi.org/10.1145/1553374.1553380
Bugallo, M. (2026). Cognitivism prevents us from understanding artificial intelligence. AI & Society, 41, 1263–1264. https://doi.org/10.1007/s00146-025-02583-5
Jiang, A. Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., & Sayed, N. (2024). Mixtral of experts. arXiv. https://arxiv.org/abs/2401.04088
Piantadosi, S. T. (2024). Modern language models refute Chomsky's approach to language. In E. Gibson & M. Poliak (Eds.), From fieldwork to linguistic theory: A tribute to Dan Everett (pp. 353–414). Language Science Press. https://doi.org/10.5281/zenodo.12665933
Reynolds, J. N., Hyland, B. I., & Wickens, J. R. (2001). A cellular mechanism of reward-related learning. Nature, 413(6851), 67–70. https://doi.org/10.1038/35092560
Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. https://doi.org/10.1126/science.275.5306.1593
Siegelmann, H. T., & Sontag, E. D. (1995). On the computational power of neural nets. Journal of Computer and System Sciences, 50(1), 132–150. https://doi.org/10.1006/jcss.1995.1013
Uttal, W. R. (2004). Dualism: The original sin of cognitivism. Routledge.
Xie, Z., et al. (2025). mHC: Manifold-constrained hyper-connections. arXiv. https://doi.org/10.48550/arXiv.2512.24880