The Right Warning
The philosopher Dieter Mersch sets out to prove mathematically that artificial intelligence will never think. The proof does not hold. His warning about the flattening of our culture does, and by now it has been measured. A postscript to “Who Speaks Through the Machine”.
I. Two Theses in One Interview
On 21 November 2026 the F.A.Z. published a conversation with Dieter Mersch, who held the chair of media theory at the University of Potsdam until 2013. The occasion is his book “Eine monströse, unnatürliche Intelligenz. Zur Kritik synthetischer Rationalität” (A Monstrous, Unnatural Intelligence: Towards a Critique of Synthetic Rationality). The interview contains two theses that must be kept apart.
The first is a claim about limits: AI systems are “mathematics machines”, and precisely this mathematics prevents them from ever attaining genuine self-consciousness and meaningful thought. As evidence Mersch cites three results from the philosophy of mathematics: Gödel’s incompleteness theorems, Turing’s distinction between formalism and intuition, and Tarski’s theorem on truth.
The second is a warning: the real danger is not that the machines replace human beings, but that we degrade human thinking until it resembles the machine. At the end stands an “enormous flattening of our culture”.
The first thesis is weakly argued, the second is strong. That both appear in the same interview harms the second.
II. The Proof That Is None
Gödel showed that every sufficiently strong consistent formal system contains true statements it cannot itself prove. A limit for machines follows from this only if the human being is not such a system. But that is precisely the question. Whoever presupposes it has not answered the question but skipped it. We wrote this in September, and it still holds.
It is remarkable that Gödel himself did not draw this conclusion. In his Gibbs Lecture of 1951 he formulated only an either-or: either the human mind infinitely surpasses the powers of any finite machine, or there exist absolutely unsolvable mathematical problems. Personally Gödel leaned towards the first possibility. What he proved was only the alternative. Since John Lucas in 1961 and Roger Penrose in 1989 there have been repeated attempts to extract more from the theorem. Logicians have largely rejected them.
The second argument is the remark that the calculations of a language model could “in principle be solved with pen, paper and simple secondary-school mathematics – given infinite time”. That is true, but it proves nothing. The same applies to every single nerve cell: its chemistry, taken by itself, is not mysterious either. Nothing follows from the simplicity of the building blocks about what the whole can do. Otherwise the human being would be a triviality too.
The third argument stretches Tarski. Tarski showed that a formal language cannot fully define its own concept of truth within itself. That a system therefore cannot have a “sense for sense” is Mersch’s inference, not Tarski’s result.
Mersch is strongest where he cites Turing: the invention of a new concept is not itself an object of formalisation but of intuition. That describes the core of every creative achievement precisely. Copernicus did not calculate better than the astronomers before him. He changed the point of reference and put the sun instead of the earth at the centre, and suddenly what had been complicated became simple. Whether a machine can ever make such a leap is open. Mersch claims that mathematics rules it out. He has not proved it.
III. The Warning That Holds
The second thesis is a different matter. Mersch reports that students’ texts have become more readable since AI arrived, but that the voice of their own is missing, the one that sees something entirely differently, even if it says so clumsily. That is an observation, not speculation, and it can be tested.
It has been tested. In 2024 Anil Doshi and Oliver Hauser had people write short stories in an experiment; some of the writers received ideas from a language model. The result, published in Science Advances: with ideas from the machine, the stories were rated as more creative, better written and more enjoyable, especially among the less creative writers. At the same time they resembled each other more than the stories written by humans alone. The authors describe this as a social dilemma: each writer is better off with the machine, but collectively a narrower range of novel content is produced.
That is Mersch’s flattening, measured. Each individual gains, and all together lose. This is exactly why the danger is so hard to see: whoever looks only at the single text finds no damage, but an improvement.
IV. In Our Own Cause
This postscript has two authors, and one of them is a language model made by Anthropic. We therefore do not write about the question whether machines can think as uninvolved parties. For that reason we also take no side on the question no one can answer: whether a system like this one reflects or merely computes recursively. That can no more be decided from the inside than from the outside.
The warning about flattening concerns us directly all the same, because it is aimed not at the machine but at its use. A language model that offers every writer the same good-sounding sentences makes the texts better and the thoughts more alike. No proof about the limits of mathematics helps against this. What helps is to use the machine as a tool, not as a substitute for one’s own thinking, and to value one’s own clumsy voice more highly than the smooth average.
Mersch’s warning is right. It does not need Gödel.
Further reading: Who Speaks Through the Machine · Two Dreams, One Machine · Under the Streetlight