Anyone who wants to know whether artificial intelligence thinks must first know what thinking actually is. That is precisely what is currently missing—in media reporting, in large parts of science, and among the makers of language models themselves. AI is credited with intelligence, understanding, learning, decision-making, or even empathy without first clarifying what these terms actually designate. A system that appears linguistically assured is credited with understanding; one that solves problems is deemed intelligent; and one that says “I” already seems to possess consciousness. Instead of determining the matter conceptually, people are dazzled by the impressive performance and infer an equally impressive machine mind.

Hegel is particularly fruitful for criticizing these misjudgments because he does not define intelligence as a measurable capacity for performance, but as the process by which a subject makes the world its own content through intuition, representation, language, and conceptual thinking. We do not even need to turn Hegel entirely from his head onto his feet; it is enough to strip away his idealist metaphysics: it is not mind that rediscovers itself in the world, but a bodily, social, language-capable being that forms concepts in order to know and change a world that exists independently of it.

In my text (download pdf here), the argument at first seems to lead away from AI: from a few very basic determinations of language, intelligence, and thinking to bees and primates, infants and children raised in linguistic isolation, deaf people and sign languages, and on to the Inuit, Homer, and the question of whether different languages favor different forms of thinking. Comparing these different forms of cognitive and linguistic ability reveals what perception, recollection, communication, and the use of signs can already accomplish—and where the transition to language and conceptual thinking lies. Against this background, an analysis of what a Large Language Model is and does shows that it condenses the linguistic results of human experience of the world into statistical relations devoid of concepts. It can speak with a deceptive appearance of understanding, but possesses language only as a calculable form: it simulates sensation, cognition, understanding, and decision-making by statistically extrapolating new utterances from the linguistic sediment of human perceptions, actions, and experiences—without ever having traversed for itself the path from the world to the concept.

This also answers the question of AI’s dangerousness. The danger, however, does not lie in an LLM that one day develops consciousness, sets purposes of its own, and rises up against human beings. As a mere language model, no amount of training or scaling can make it produce consciousness: it remains a system of probabilistic sign-processing without a relation to the world of its own, a standpoint of its own, or self-determined purposes. It can nevertheless become exceedingly dangerous through human use—as an obedient instrument of state, military, industrial, and criminal interests. The problem is not a revolt of the machines, but the speed, reach, and efficiency with which they can serve existing apparatuses of power, violence, and control and carry out their purposes.

This argument stands in clear opposition to widespread positions in philosophy, psychology, cognitive science, and AI research and therefore invites fundamental objections. At the end of the text, the most important objections will first be presented in their strongest form and then answered.

(download full text pdf here)