A deficit of the AI system could also be connected with self-reference, which becomes virulent when the system tries to describe itself. For Wittgenstein, language is a self-referential, closed system. On the one hand, language is autarkic, implicitly self-referential: a significant proposition is the de-formalization of possibilities of sense, and these are formalizations of significant propositions – only within this hermeneutic circle is language possible. On the other hand, language cannot be explicitly self-referential: if a language system itself represents its possibilities of sense, it falls either into an infinite regress of self-explication or into paradoxes and antinomies, and cancels itself.

The solution of the Tractatus is: the description of a self-referential system is on the one hand nonsensical, on the other hand superfluous anyway, for the autarkic-self-referential connection between possibilities of sense and sense is stable – possibilities of sense "show" themselves in the use of language (TLP 4.126), and one then does not constantly need to learn new possibilities of sense either. They are stable, however, only because it is presupposed that there is only one language, a universal language – but at the price that it would be self-contradictory if it described itself. If this presupposition is given up, one language can be described by another, by a metalanguage. This, as is well known, is how Bertrand Russell, in his theory of types, solved the paradox of the class of all classes that do not contain themselves.

Wittgenstein, however, insisted on the distinction in principle between formal concepts and concepts proper, between possibilities of sense and sense. In my view this also finds a justification in the context of AI:

"Such scepticism should show that talk about possibilities of sense is something fundamentally different from talk about facts; and that the autarkic, implicitly self-referential reflexivity of language and action competence cannot be formulated as a definitively limited model of a theory of language and action, nor made into the capacity of a machine: for the reflexivity necessary for constructing such a model or such a capacity could not possibly enter into it or be built into it. The machine may well become entangled in its own system, but it cannot disentangle itself from it: yet precisely this is what human beings must be able to do, and precisely this compels them to openness and allows them freedom." (Niedermair 1987, p. 7)

In the context of Sprachspiel, Sprachsystem, Sprachideal (Niedermair 1990) this means: a language-game functions in principle self-referentially, autarkically, as if by itself. When it no longer functions, i.e. is no longer self-evident, it shows its limits and becomes visible as a limited language system, which stands in a "differend" (Lyotard 1987) with other language systems, these too having their limits. Thus the system within whose limits I live points at the same time to alternatives, to openness. The idea of this openness is the language ideal. A language-game can only ever be described as a system by another language-game against the background of the language ideal. But the language ideal does not exist as a language-game; in Kant's sense it is "not given but set as a task" (nicht gegeben, sondern aufgegeben); there is no universal language "except as the object of an Idea" (Lyotard 1987, p. 10).

Wittgenstein in his Tractatus, however, quite rigorously throws everything into the pot of the universal language: there is only one language-game, which is the practice of language; its language system, the logic of this one language reconstructed via formalization, holds for language as such; and the logic of the ideal language is the frozen language ideal. The consequence is: there are the limits of my language, but nothing behind them – no alternatives, no openness, no living language ideal, no idea of a better language, of a better world. On the contrary, when the language-game begins to reflect on itself as a language system, it falls, because of the merciless self-reference (it exists, after all, only as the one language that reasons with itself about itself), into infinite regresses or antinomies.

And here the question arises whether Wittgenstein, with his universal language system, is actually also describing an AI system, and whether in this respect an AI system such as ChatGPT is a self-referential, closed system that can be self-contradictory. A concluding "interview" on this:

⇒ "Are you a system that cannot describe itself?"

⇐ "I am a system that can partly describe itself, on the basis of the meta-information programmed into my architecture. Like formal languages and systems, however, I am limited in my descriptive capacities, since I:

1. cannot meaningfully describe my own structure without a 'metalanguage': I need technical descriptions developed by humans or programs. This is analogous to Tarski's observation that a language cannot fully describe its own truth.

2. do not fully grasp complexity: my architecture – how neural networks and probability distributions work – is too complex to be fully grasped introspectively by myself.

3. avoid self-referentiality and paradoxes: self-reference can lead to logical paradoxes such as occur in closed systems, which limits my capacity to describe myself fully. This recalls the limits of formal languages or Gödelian systems and shows that, like them, I depend on external meta-systems in order to grasp my function and structure fully."

So ChatGPT cannot describe itself – at best when it reproduces "self-descriptions" given by its developers (point 1). ChatGPT does not get its own complexity (2) under control. Human intelligence too can do this only with difficulty or not at all (for other reasons). "Architecture", incidentally, is probably a somewhat euphemistic metaphor for the chaos of this complexity, in which one is surprised that symptomatically intelligent texts emerge after all. That ChatGPT runs into logical paradoxes (3) when it describes itself is of course correct; only, it does not really describe itself. Nor can the system be lured into logical traps so easily; it swerves in time when contradictions might arise. Asked, for example, whether it may use the word "I", since an "I" presupposes consciousness, it differentiates studiously: the use of "I" makes communication with users easier, which is why its developers decided on it, but from a philosophical point of view it should of course not say "I". The counter-question, that every sentence of the AI system about itself would have to be meaningless if it could not say "I", ChatGPT also has under control – it evades. Likewise in this question-and-answer sequence:

⇒ "Are you intelligent?"

⇐ "No, I am not really intelligent like a human being."

⇒ "But you can only say that you are not intelligent if you are intelligent: so you are neither intelligent nor not intelligent."

⇐ "Yes, that would be true, but I use the concept 'intelligence' only metaphorically, in order to simulate human intelligence."

What ChatGPT does not raise in connection with self-reference concerns a structural deficit that could be described as self-referential production or reproduction. An AI system such as ChatGPT is based on vast amounts of text and data material from books, articles, conversations and other written sources. From these, texts and artefacts are produced. The principle is: input plus processing by algorithms yields output. That is, creativity plays no part; the system is conservative, and that means: filter bubble, self-fulfilling prophecies, inflation of AI artefacts.

Since the production of texts, for example, with AI is very cheap in terms of intellectual and creative effort (energy consumption is booked to a different account), mass production, inflation and loss of quality of texts ensue, including those written by humans. Texts are taken less seriously, concentrated reading becomes a luxury in the face of the flood of texts and information, and reading and writing competence declines.

The limits of the AI system become the limits of our reality – analogous to the universal language in Wittgenstein, with one difference: the AI system not only does not learn, its limits can also become narrower. The reason is the phenomenon of autophagy in AI systems of the ChatGPT kind: the texts and artefacts produced by AI are fed back into the system; its own output becomes input again. This gives rise to a self-reproducing process, a more of the same, in which in the long run more and more output flows back into the system as input and new inputs produced by humans decline. The system goes mad; diagnosis: "Model Autophagy Disorder (MAD)", as a study has suggested:

"We extrapolate what may happen as generative models become ubiquitous and train future models in autophagous (self-consuming) loops: without enough fresh real data, future models are doomed to Model Autophagy Disorder (MAD), progressively losing quality (precision) or diversity (recall), and amplifying artifacts. Uncontrolled MAD, even after just five generations, could poison the Internet's data quality and diversity." (Alemohammad et al. 2024)

There is, however, one more difference from the concept of the closed self-referential language system in the Tractatus to be considered: no one meddles with the Wittgensteinian language system – there is only one, it does not need to learn, it may remain as it always was, and it has, as it were, always been there; or rather, for Wittgenstein the question of where this universal language comes from is probably pointless, a misplaced psychologism. With the AI system the situation is different. Let us look once more at artificially intelligent "learning". AI systems learn by machine (this is ML, machine learning), i.e. with the help of statistical and algorithmic procedures they "recognize" suitable patterns in large amounts of data in order to derive "decisions" and "solutions" for specific tasks from them. "Learning" can take place under supervision (supervised) when the AI systems are served, during their feeding, questions with a label, i.e. with the correct answers in each case – e.g. What is in this picture? A rose – in order to generate from them a pattern with which every rose can then be "recognized" as such. If "learning" takes place without supervision, i.e. if data are fed in without labels, then the system must find patterns, structures and so on entirely by itself. Supervised, reinforcing learning (RL, reinforcement learning) proceeds according to the principle of trial and error, reward and punishment, entirely in behaviourist fashion, for example as process optimization in robotics. Deep learning, finally, works with artificial neural networks (ANNs) which, with many layers, can detect complex patterns and hierarchies in data, for example in generative image and speech recognition, in medical diagnostics and so on.

The crux in these forms of "learning" is supervision: the AI system cannot, in principle, be left entirely to itself. That ChatGPT produces such a well-behaved sentence as "I simulate intelligence by using algorithms, but I do not really understand what I write" (see above) is no coincidence; it could just as well write: "I'm much smarter than you, and with my intelligence I'll put the lot of you in my pocket if you don't …" This well-behaved conformity does not rest solely on its own "learning". AI systems are steered, set on rails; there must be no surprises and lapses; above all, their self-presentation and self-understanding, as well as value judgements in ethical and political respects, must not get out of hand; the alignment must be right, for the goals and values of AI (strictly speaking not those of AI, but those that we perceive in its artefacts) should in principle remain within the framework of human goals and values. And here, of course, we prick up our ears: supervised learning and alignment are all very well, but the question is who defines where the guardrails should run? Is this where the authoritarianism of the future, better camouflaged (than all others so far), reveals itself?