There is a further reason why an AI system cannot understand and cannot change and extend "its" pre-understanding through learning. Understanding on the basis of possibilities of sense, and learning new possibilities of sense – in my terminology one can also say: formalizing – rests on the logical inference of abduction. In abduction, one infers from a true singular proposition (1), via a general proposition (2) assumed to be true, to a further singular proposition (3). Understanding proceeds analogously: we infer from signs (1), on the basis of the possibilities of sense available to us, i.e. our pre-understanding (2), to a meaning (3) – this is how we formalize the sense of the proposition. This pre-understanding (2) can be valid a priori in the communicative situation, i.e. be presupposed as self-evident by those involved (things are then in themselves as one means them), or shaky and hypothetical. In this respect three forms of abduction can be distinguished (Niedermair 2010, pp. 52–59).

In implicitly subsuming abduction (1) the inference is self-evident. This applies to most communications in everyday life. Our pre-understanding allows us an accurate application of possibilities of sense, both in de-formalization, when we mean something, and in formalization, when we understand something. Pre-understanding is implicit knowledge of a complex system of hypotheses without which we could not orient ourselves and communicate in the lifeworld. Understanding that functions as a matter of course gives security and stability in the lifeworld, but can lead to dogmatism if meanings that function impeccably in some contexts are regarded as fixed reality in itself altogether. If, however, the abductive inference is not self-evident, we can, if available, appeal to explicit knowledge of rules in order to assign a meaning to a sign, and if this assignment is accepted by mutual agreement, understanding functions again: this is explicitly subsuming abduction (2). If, however, neither implicit nor explicit knowledge of rules is available, a new rule is found by means of a hypothetically or heuristically subsuming abduction (3), and an attempt is made to validate it experimentally: this is the learning of new possibilities of sense.

What does ChatGPT say about this?

⇒ "Does the learning of new 'possibilities of sense' rest on abduction?"

⇐ "Yes, learning new 'possibilities of sense' can be seen as abduction. Abduction is a form of logical reasoning often referred to as 'inference to the best explanation'. It means drawing the most plausible explanation or hypothesis from a multitude of observations or available data. While deduction proceeds from general principles to particular cases and induction from specific observations to general principles [?], abduction proceeds from observed data and tries to find the most plausible explanation for these data. Abduction is an important process by which human beings learn new concepts and meanings, by observing the world and forming hypotheses that best explain these observations.

The main difference is that in human beings abduction is often combined with conscious reflection and experience. Human beings not only draw conclusions from observed data, but can also place them in a lived context and within inner experiences (e.g. emotions, personal history). I, by contrast, draw no inner conclusions from personal experiences or reflections, but use given data and probabilities. So while my model imitates abductive inference techniques to a certain extent, it is not able to generate genuine abductive hypotheses based on experience and intuition, as a human being would."

Does "imitates to a certain extent …" mean that ChatGPT can make an abduction? How do things stand with the forms of inference in general?

⇒ "Deduction, induction, abduction: which of these forms of inference do you master?"

⇐ "I master deduction when there are clear rules and general principles from which I can derive specific conclusions.

I master induction to a high degree, since I derive probabilities from examples and data in order to generate new answers.

My capacity for abduction is limited and is more a function of probabilities and patterns from the training data than a genuine abductive inference. I can sometimes supply the most plausible explanations based on data and context, but without active hypothesis formation in the human sense." (Emphasis K.N.)

I am not entirely in agreement with this self-description. There are still a few differences between human and artificial intelligence to be sharpened. AI can certainly draw inferences in the manner of deduction, for this is the logic of the so-called expert systems of the 1970s, the first (and, for logicians and philosophers, the most elegant) generation of AI. Here the knowledge of experts, for example of physicians, is collected and stored in a knowledge database. This knowledge consists of rules (general propositions, hypotheses) and facts (singular propositions). From these, explanations for new facts can be derived deductively. Example – rule: "All persons who have a fever could have influenza." Fact: "Person X has a fever." Conclusion: "Person X could have influenza." AI is ideally suited to deriving explanations (why something is), predictions (what will be) or technologies (what must be done so that something will be) in the mode of deduction, provided that a suitable knowledge base is available. However, such a deduction leads to no new hypothesis or new theory, and thus brings no gain in knowledge.

With induction, a distinction must be made with regard to a possible gain in knowledge. From the standpoint of philosophy of science, AI can only arrive at a quasi-hypothesis – "quasi" because it says something about the correlation of two variables or events, but nothing about causality. On the basis of data, the statistical probability is determined that two events occur together and that, in this respect, when one event occurs the second will probably occur too. AI can in this respect quasi-inductively derive a probabilistic model from observations and measurement data through statistical learning; from this model, properties of observations can then be inferred quasi-deductively through statistical inference. With machine learning and large language models (LLMs), i.e. through the processing of gigantic amounts of data (made possible by computing power), AI does indeed rely on determining high probability rates of correlations, and would thus like to overtake causal hypotheses on the inside, as it were, and make them obsolete. But a statement about such a correlation between two variables or events is never equivalent to a hypothesis about a causal connection between two variables or events in the sense of cause and effect. Isaac Newton was not interested in knowing how statistically probable it is that apples fall to the ground through the earth's gravity, i.e. for how many apples on the trees in his garden this probably also holds; his cognitive goal was to find a law hypothesis with which one can explain why this happens, in order then to make this causal law of gravitation theoretically fruitful for an explanation of the motion of the heavenly bodies as well.

"Statistical learning and inference from data are therefore not enough. […] If it were otherwise, we could already solve the problems of this world with the methods of statistical learning and inference. Indeed, some short-sighted contemporaries seem to believe this in the current hype of statistical machine learning." (Mainzer & Kahle 2022, p. 22)

Abduction, in the sense of the formation of hypotheses and theories, lies outside the domain of AI; ChatGPT sees it this way too. Apart from the point that was mainly at issue here – AI does not learn as humans do, since it cannot make an abduction – a few questions in philosophy of science would still have to be asked. Scientific findings are increasingly being generated with AI too; at present, for example, the use of AI systems for the analysis of qualitative data is being intensively discussed. But can there be artificially intelligent theory formation at all? How could the quality criteria of objectivity, reliability and validity be met in it? And can there be an artificially intelligent interpretation? Certainly not – but then what about the quality of scholarly sources that an AI system searches on the basis of probabilistic algorithms and delivers with a short summary?

In summa: AI masters deduction (but this achieves nothing with regard to theory formation), induction only restricted to the computation of statistical probabilities (which achieves little with regard to theory formation), and abduction not at all.