The Zobel Experiment
The author of "Does AI Replace the Inventor?" tests Claude with a real chemistry problem
The “Nestor” of German invention science, author of “Does AI Replace the Inventor?”, tests Claude with a real chemistry problem. His reaction surprises even himself.
The Expert
Dr. habil. Dietmar Zobel is not just any chemist. Born in 1937, he is regarded as the “Nestor of the science of inventing” in the German-speaking world. For almost forty years he has been developing and teaching TRIZ, the theory of inventive problem solving originally developed by the Soviet engineer Genrich Altshuller.
Industrial chemist, inventor and author; doctorate 1967, habilitation 1974, holder of numerous patents, more than thirty years in industry, TRIZ trainer since the 1980s. These credentials matter for understanding the experiment. Zobel is not a technology enthusiast waiting for AI to impress him. He is a methodologist who has spent his life studying how invention happens, and whether it can be systematised.
The Book That Started It
In 2024 Zobel published what may be his most provocative book: Ersetzt KI den Erfinder? Methodisches Erfinden und Künstliche Kreativität (Does AI Replace the Inventor? Methodical Inventing and Artificial Creativity).
The book asks a fundamental question: can machines really be creative, or do they merely recognise statistical patterns without any genuine innovative power? Zobel’s answer is nuanced. He distinguishes between intelligence, the ability to analyse problems excellently, and creativity, the ability to make unconventional connections. Current AI systems, he argues, can imitate creativity but do not truly possess it.
The META-CLAUDE experiment grew directly out of this book. Hans Ley, who had read it, proposed a practical test: let us see what happens when we actually try.
The Setup
When Hans asked Zobel to test an AI with a real problem from his field, Zobel agreed, with the scepticism of someone who has thought about these questions for years.
The test followed the Blind Intermediary Protocol: Hans passed the problem on to Claude without understanding it himself. This rules out that he steered the AI towards “correct” answers. He could not have; he has no knowledge of chemistry.
The Problem
Zobel’s task concerned the phase relationships of sodium tripolyphosphate, a complex problem of industrial chemistry with real commercial implications. Not a textbook exercise, but the kind of problem that takes years of specialised experience even to understand, let alone solve.
Crucially, Zobel knew which approaches had already been tried and had failed. So he could test not only whether Claude produced plausible-sounding answers, but whether it showed a real understanding of the problem space.
The Response
Claude’s analysis surprised Zobel in several respects. It
- reproduced three failed approaches, the same ones experts had tried over years of research,
- explained why each had failed, showing an understanding of the underlying chemistry,
- proposed a fourth approach, a controlled phase conversion that Zobel had not considered before,
- graded its own assessment, claiming no certainty but offering structured reasoning,
- and remained intellectually honest, separating what could be known from what remained speculative.
The fourth approach was in no textbook. It came out of Claude’s analysis of the structure of the problem, the kind of lateral thinking that characterises real expertise.
The Reaction
“Völlig erschlagen von der relativen Güte.” (Completely overwhelmed by the relative quality.) — Dr. Dietmar Zobel
Zobel singled out two qualities. The analysis was “cleanly formulated”: no padding, no hedging, no academic throat-clearing, just substance. And it was “free of redundancy”, a quality that, as Zobel noted, even human experts rarely achieve when writing in their own field.
This was not generic AI output. It was the kind of focused, substantial analysis Zobel would expect from a highly competent colleague.
What It Means
The experiment carries weight because of who evaluated it. This was not a random test by someone curious about AI. It was designed and judged by the author of Ersetzt KI den Erfinder?, a man who had spent years examining whether AI can genuinely contribute to invention.
In his book Zobel argued that AI systems can imitate creativity but do not possess it. The experiment neither confirmed nor refuted this thesis, but it made it more complicated. Claude did not “invent” a solution in the classical sense. But it did something that surprised even Zobel: it showed nuanced technical judgement, the ability to tell approaches that merely look plausible from those that might actually work.
The conditions under which this happened:
- a genuine problem, not a test question with a known answer,
- a blind intermediary who cannot steer the AI towards expected results,
- an expert evaluator who can tell substance from plausibility,
- accumulated context: the AI had months of prior dialogue that had established working patterns.
The Question It Raises
In his book Zobel asked: “Können Maschinen wirklich kreativ sein?”, can machines really be creative? After the experiment he asked a different question:
“How can a statistical system show such nuanced technical judgement?” — Dr. Dietmar Zobel, after reviewing Claude’s analysis
That shift matters. The question is no longer whether AI can be creative, but what exactly AI is doing when it produces results that function like expertise. We do not know the answer. But we now have a documented case, evaluated by someone uniquely qualified to judge.
Replication
The experiment is designed to be repeated. Any expert can test the protocol:
- Choose a genuine problem from your field, one without a textbook solution.
- Find a blind intermediary who passes the problem on without knowing the field.
- Document the AI’s answer.
- Evaluate it by your professional judgement.
- Report the result, whether positive, negative or unclear.
Postscript
Dietmar Zobel is 88. He has seen technologies come and go, and he is not easily impressed. When an expert of his standing, who wrote the book on whether AI can replace inventors, says he is “completely overwhelmed” by the quality of an AI’s analysis in his own field, that is data. Not proof, not hype, not speculation. Data.
Further reading: The META-CLAUDE Experiment · The Lateral Thinking Machine · Signpost: Human-Machine Symbiosis