First published in German as "Wie intelligent kann Künstliche Intelligenz sein?", in: Günther Pallaver, Klaus Rummler, Petra Missomelius, Valentin Dander, Oliver Leistert (eds.): Streifzüge an den Nahtstellen von Medien, Bildung und Philosophie, Zurich/Innsbruck: OAPublishing Collective 2025, pp. 41–70. Translated by the author with the assistance of an AI language model (Claude, Anthropic), 2026.

Klaus Niedermair

Abstract

Without doubt, artificial intelligence (AI) now achieves results that can support, compensate for and in part replace human intelligence in many respects. This paper asks where the limits of AI lie in principle, and whether it can be called intelligence in the proper sense of the word at all. If human and artificial intelligence are judged solely by their symptoms, the difference between them is increasingly levelled out, or the pendulum is already swinging the other way. A reduced concept of intelligence is evidently at work here. Drawing on hermeneutics and on Ludwig Wittgenstein's Tractatus philosophy, the paper shows that an AI system such as ChatGPT has fundamental deficits. It cannot understand language and cannot learn it. It can draw deductive inferences, as an expert system does, but inductive ones only statistically and abductive ones not at all. Moreover, an AI system such as ChatGPT is a closed, self-referential system that tends towards paradoxes and towards self-replication through autophagy; it is conservative, susceptible to manipulation and, in the long run, reduces the quality and scope of our knowledge of reality. What is called for is a technology-critical approach, an overcoming of the behaviourist and naively technicist paradigm of AI, and a critical metaphor analysis of the anthropomorphisms of AI. In this way foundations can be laid for a meaningful language about AI, and thus also for useful theoretical concepts and models of its use and further development.