We understand sense depending on what possibilities of sense are available in the potential of our pre-understanding. In this respect the circle between possibilities of sense and sense also marks the limits of our reality. That is, if the limits always remain the same because the pre-understanding does not change, this circle can become a circulus vitiosus. But human beings can (more or less) also cross limits by learning something new – new knowledge about reality, and that means above all new possibilities of knowing reality, new possibilities of sense. What does ChatGPT say about this?

⇒ "How do you 'learn' 'possibilities of sense'?"

⇐ "I do not 'learn' [sic!] possibilities of sense in the same way as a human being learns [sic!]; rather, I base my answers on training data, logical structures and principles built into my system.

Training data: I was trained on the basis of huge amounts of text consisting of books, articles, conversations and other written sources. These texts contain information about how people use language, what things mean [?] and what connections exist. From this, implicit rules for possibilities of sense arise, e.g. 'Houses can have colours', 'Colours are perceptible properties', a contradiction such as 'red and colourless' is illogical.

Logical structures and context models: I use logical principles and models to process language. Semantic networks: relations between concepts (e.g. house → colour → red) are anchored in my structure. This enables me to recognize which statements are meaningful and which are nonsensical.

Context dependence: I analyse the context in order to determine which possibilities of sense are relevant in a situation. If someone, for example, is speaking about abstract or metaphorical houses, I adjust my assumptions.

Derived rules for possibilities of sense: From the training data and the logical principles, rules arise that help me grasp possibilities of sense, e.g.: if an object is ascribed a property (such as 'red'), it must satisfy the category that makes this property possible (e.g. be colour-determinable).

Error and feedback mechanisms: When I encounter contradictions or feedback (such as 'That is nonsensical!'), I can indirectly improve my models through user interaction. So I do not 'learn' [sic!] actively in the moment, but draw on training and usage data that make my answers more precise ..."

ChatGPT itself puts quotation marks to mark the metaphorical way of speaking; it seems to have "learned" from the context that I, as its prompter counterpart, am critical. ChatGPT thus "learns" differently from us humans. It would be wrong to assume that human learning in principle proceeds in the same manner of induction as the "learning" of AI, in which, quasi-inductively (but only quasi), by statistical procedures, a probabilistic model for further possibilities of weaving is obtained, so to speak, from the fabric of text. ChatGPT is "trained on huge amounts of text data". Nor does it learn in the literal sense through "contradictions and feedback" in user interactions; in both cases the "training" is heteronomous, externally controlled by text data and algorithms. Humans learn differently. When we mean something and understand something, every sentence, every action is not only the circular application of possibilities of sense but also a learning process in which possibilities of sense are experimented with – whether they fit and prove themselves or not, and if not, they are changed, supplemented or replaced – in the sense of a constructivist theory of learning, for example as accommodation in Jean Piaget's sense or equilibration according to Ernst von Glasersfeld. Through learning, the circle then becomes an open spiral, a circulus fructuosus. It is precisely this spiral that an AI system such as ChatGPT does not achieve: it cannot learn by itself, cross its limits and in this way learn something new.