A critique of Jobst Landgrebe’s claim that artificial intelligence does not exist

In a recent essay for Achgut, Jobst Landgrebe attacks what he sees as a new form of technological superstition. According to him, Western media and elites dramatically exaggerate the capabilities of artificial intelligence, ignore its mathematical limitations and mistake sophisticated statistical software for genuine intelligence.

Landgrebe argues that intelligence means responding appropriately to entirely new situations without prior training. Since today’s AI systems are created through extensive training and configuration, he concludes that they cannot be intelligent. Large language models, he writes, merely calculate probable sequences of symbols. They have no intentions, no understanding, no consciousness and no genuine creativity. Their stochastic nature inevitably produces errors and hallucinations.

He further argues that the human mind cannot be reproduced by a computer because consciousness emerges from an immensely complex biological system that cannot be fully represented mathematically. AI will therefore never think, develop intentions or undergo autonomous evolution. Model collapse caused by training systems on AI-generated material supposedly reinforces this limitation. Despite acknowledging that AI can automate repetitive tasks and detect patterns in large datasets, Landgrebe predicts that it will replace no more than five per cent of human labour within the next decade. He ultimately interprets belief in AI as a symptom of a broader spiritual and epistemic crisis in Western society.

There is a reasonable argument hidden inside this essay. AI reporting is often breathless, anthropomorphic and economically naive. Language models remain unreliable, hallucinations are real, and predictions of imminent superintelligence or the automation of almost all human work deserve scepticism.

But Landgrebe’s counterargument is considerably weaker than the hype he criticises.

An intelligence definition designed to exclude machines

The central problem is his definition of intelligence. Landgrebe describes it as the ability to respond appropriately to a novel situation “without prior practice” or training.

This is not a neutral definition. It establishes in advance that trained systems cannot qualify as intelligent and then uses that assumption to prove that trained systems are not intelligent.

More importantly, the criterion would exclude much of human intelligence as well. Human beings do not enter the world with language, mathematics, social competence or professional judgement fully developed. The human brain is shaped through continuous exposure, imitation, feedback, practice and interaction with its environment.

Human learning is certainly not identical to the training of an LLM. It is embodied, social, emotional and connected to real-world consequences. But that difference does not rescue Landgrebe’s definition. If prior learning disqualifies a system from being intelligent, most intelligent human behaviour would also have to be disqualified.

There is no universally accepted definition of intelligence. One influential attempt defines it more broadly as an agent’s ability to achieve goals across a wide range of environments. Learning and adaptation are then indicators of intelligence rather than reasons to deny it.

Intelligence is not the same as consciousness

The essay also repeatedly moves between intelligence, consciousness, understanding, intention, emotion and autonomy as though they were interchangeable.

They are not.

A system may display intelligent behaviour without possessing subjective experience. It may solve problems without having emotions, and it may pursue goals assigned by humans without having independent desires. Whether machines can ever become conscious is an important philosophical question, but it is not necessary to settle that question before evaluating what they can do.

For most practical purposes, the decisive issue is not whether an AI secretly “experiences” its answer. The issue is whether it can produce reliable, useful and contextually appropriate results.

This was already the strength of Alan Turing’s approach. Rather than trying to solve the metaphysical question of what thinking truly is, he proposed evaluating observable performance.

A sufficiently accurate simulation of intelligence may be economically, politically and culturally indistinguishable from intelligence itself. A simulated lawyer that analyses contracts better than most lawyers will affect the legal profession regardless of whether it feels anything while doing so.

“Only probability” is not a refutation

Landgrebe’s description of LLMs as systems that calculate probable symbol sequences is technically relevant but philosophically insufficient.

Describing the mechanism of a system does not determine the full range of abilities emerging from it. Saying that an LLM “only predicts tokens” resembles saying that the brain “only transmits electrochemical signals”. Both descriptions may be correct at one level while telling us little about the capabilities of the complete system.

Language prediction turns out to require the acquisition of extensive representations of grammar, concepts, relationships, styles and patterns of reasoning. This does not prove that a language model understands the world exactly as a human does. But neither can its demonstrated capabilities simply be dismissed by repeating the mechanism through which they arise.

Research on GPT-3 already showed that a pretrained language model could adapt to previously unspecified tasks from instructions or a small number of examples, without additional parameter updates. Its performance remained highly uneven, but such in-context adaptation sits uneasily with the claim that these systems merely reproduce fixed responses encoded during training.

Training does not mean copying

The claim that AI “never creates anything new” is similarly underdefined.

Almost all human creativity recombines earlier perceptions, experiences, techniques and cultural material. A novelist uses an existing language. A composer works with inherited musical structures. A scientist develops ideas from existing theories and observations.

Novelty does not require creation from nothing. A more useful question is whether a system can produce an output that was not explicitly stored, that differs meaningfully from its examples and that solves a new problem.

Current AI systems clearly do this in at least a limited functional sense. Whether that should be called creativity depends largely on how creativity is defined—not on the mere fact that the system was trained on previous material.

Errors do not disprove intelligence

Landgrebe is right that stochastic systems cannot be made perfectly reliable. But intelligence has never meant infallibility.

Humans misremember, misunderstand, hallucinate, rationalise and confidently produce false answers. We nevertheless regard them as intelligent because intelligence is a matter of degree and capability, not error-free operation.

The relevant questions are therefore comparative: How often does a system fail? Under which conditions? Can its results be verified? Does it outperform humans on a specific task? Can safeguards reduce the remaining risk?

Hallucinations impose serious limits on AI deployment. They do not constitute proof that no intelligence is present.

The missing mathematical proof

The essay repeatedly refers to mathematical proofs showing that genuine AI is impossible. Yet it does not present a specific theorem, its assumptions or the logical steps connecting it to the claimed conclusion.

The fact that the human brain is a complex biological system that cannot be completely modelled does not establish that no artificial system can reproduce some or even many of its cognitive functions.

An aircraft does not need to reproduce the biology of a bird in order to fly. Artificial intelligence may likewise achieve functional abilities through mechanisms radically different from those of the human brain.

The argument quietly assumes that human cognition must first be completely modelled before a machine can display intelligence. That assumption is neither demonstrated nor self-evident.

Model collapse is real—but not universal

Landgrebe is on firmer ground when discussing model collapse. Research has shown that indiscriminately training successive models on recursively generated data can cause them to lose parts of the original data distribution and progressively deteriorate.

But the crucial word is indiscriminately. The research does not demonstrate that any use of synthetic data necessarily causes collapse, nor that AI development must inevitably degenerate. It identifies a data-management problem involving selection, mixing, provenance and the preservation of high-quality original information.

A genuine limitation is turned into a universal law.

A prediction without a foundation

The claim that AI will automate no more than five per cent of human work over the next five to ten years is presented without a model, dataset or transparent calculation.

It also fails to distinguish between jobs, tasks, working hours and productivity. AI does not need to eliminate an occupation entirely to transform it. Automating 20 per cent of the activities performed by millions of office workers could have enormous economic consequences without replacing any profession in full.

The precise scale of future automation remains uncertain. That uncertainty is a reason for cautious analysis—not for replacing grandiose industry forecasts with an equally unsupported five-per-cent figure.

From scepticism to cultural diagnosis

The weakest part of the essay is its final move from technological criticism to cultural polemic. Belief in AI is attributed to secularisation, postmodern insecurity and the decline of Christian metaphysics, before being associated with Covid policy and wind turbines.

None of this strengthens the technical argument. It merely sorts AI enthusiasm into a broader catalogue of positions the author already rejects.

The result is a mirror image of the discourse he criticises. AI evangelists treat every new benchmark as evidence that machine superintelligence is imminent. Landgrebe treats every limitation as evidence that machine intelligence is impossible.

Both sides mistake certainty for analysis.

The more reasonable conclusion

Today’s AI systems are not human minds. They have no demonstrated consciousness, no biological embodiment, no stable independent intentions and no guaranteed understanding of the world. They remain dependent on training data, human objectives, technical infrastructure and systems of verification.

But it does not follow that they are unintelligent, incapable of novelty or economically marginal.

The important question is not whether AI conforms to a definition deliberately modelled on human consciousness. It is what kinds of cognitive performance artificial systems can achieve, how reliably they can achieve them and what happens when these capabilities are deployed at scale.

AI hype deserves criticism. So does the dogmatic claim that artificial intelligence cannot exist.

Between technological salvation and technological impossibility lies the far more interesting reality: machines that may not think like us, may not feel anything at all—and may nevertheless become extraordinarily consequential.