Reasoning vs. Language - what do we know from humans
One of the most striking developments in Generative AI over the past two years has been how far language-based systems have progressed â and how quickly. At first, this progress felt intuitive. Large language models work by predicting the next word or token in a sequence, drawing on vast amounts of prior text. This maps surprisingly well to what cognitive psychology calls fast, automatic thinking: pattern recognition, linguistic intuition, and reflexive responses. Daniel Kahneman famously referred to this as System 1 thinking â quick, effortless, and largely unconscious.
What became more surprising was the apparent emergence of deliberate reasoning in newer models. By adding mechanisms like self-reflection, intermediate steps, or iterative âthinking loops,â these systems began to show behaviors that resemble System 2 thinking: slower, more effortful processes associated with planning, reasoning, and problem solving. The idea that something as simple as structured self-dialogue could unlock this kind of capability feels both elegant and unsettling. If reasoning can emerge from language alone, how far can this paradigm really take us?
This question â whether language is merely a communication tool or a foundational mechanism of thought itself â has a long history in cognitive science. A recent Nature article titled âLanguage is primarily a tool for communication rather than thoughtâ offers a sharp and surprisingly decisive perspective on this debate. Its conclusions challenge the intuition that increasingly powerful language models will naturally evolve into general reasoning systems.
Language is not necessary for any tested forms of thought
The classical approach for making inferences about brainâbehaviour associations and dissociations is to examine individuals with brain damage or disorders. If linguistic ability mediates our ability to engage in certain forms of thought, then linguistic impairments should be associated with concomitant difficulties in those aspects of thinking and reasoning. The evidence is unequivocalâthere are many cases of individuals with severe linguistic impairments, affecting both lexical and syntactic abilities, who nevertheless exhibit intact abilities to engage in many forms of thoughtâthey can solve mathematical problems, perform executive planning and follow non-verbal instructions, engage in diverse forms of reasoning, including formal logical reasoning, causal reasoning about the world and scientific reasoning, to understand what another person believes or thinks and perform pragmatic inference, to navigate in the world, and to make semantic judgements about objects and events.
[âŚ]
Neuroimaging evidence complements the evidence from individuals with brain damage. Using tools such as fMRI, we can identify the language areas in intact, healthy brains and then examine the response in those areas while individuals engage in tasks that require different forms of thought.
[âŚ]
Responses in this network to diverse non-linguistic inputs and tasks have been examined, and the evidence demonstrates that all regions of the language network are largely âsilentâ during all tested forms of thought, including mathematical reasoning, formal logical reasoning, performing demanding executive function tasks such as working memory or cognitive control tasks, understanding computer code, thinking about othersâ mental states, and making semantic judgments about objects or events. Instead, these tasks engage other brain areas that are non-overlapping with the language network (Fig. 1b), although they sometimes lie in close proximity to the language areas.
Iâm still on the fence with respect to the question how far the current LLM paradigm will take us, how much of a paradigm shift is needed to get us to something that can be called AGI. This here argues for a longer timeline with more fundamental changes needed.