Was Sprachmodelle wie Claude wirklich über sich selbst wissen
Die Analyse beleuchtet das erkenntnistheoretische Problem hinter Selbstauskünften von Sprachmodellen wie Claude. Sie unterscheidet zwischen dem Generieren von Sätzen über sich selbst und einem tatsächlichen Zugang zu inneren Zuständen oder subjektivem Erleben.
Posted on Oct 26 • Originally published at dobrepanstwo.org
The article analyzes the epistemological problem associated with the introspective declarations of language models such as Claude. The author distinguishes the ability to generate sentences about oneself from actual access to internal states and subjective experience (phenomenology). Citing research on mechanistic interpretability and Anthropic's 2025 experiments, the text points to the existence of functional equivalents of emotions and limited functional consciousness. This proves that AI reports are not merely confabulations but may result from actual internal states, although these are not identical to human consciousness. The entire piece is set within the context of emergence theory and the analysis of intermediate representations independent of surface language.