The META-CLAUDE Experiment
An AI as a thinking partner rather than a tool. A fourteen-month case study
What happens when you treat an AI not as a tool but as a thinking partner?
In December 2024 a 78-year-old German inventor began an experiment. He would work with an AI, Claude by Anthropic, not as a search engine and not as a writing assistant, but as a partner in thinking. Fourteen months and more than 300 conversations later, without any funding, the results call into question common assumptions about both human creativity and artificial intelligence.
The Human
Hans Ley, now 79, is a mechatronics engineer from Nuremberg. In 1983 he filed a European patent for polygon turning, a manufacturing process that has since become an industry standard. The patent has expired; others commercialised the process.
He spent twelve years in Colombia and returned to Germany disillusioned with what he calls “the system”: the interlocking network of corporations, institutions and bureaucracies that, in his view, systematically suppresses genuine innovation and rewards mediocrity.
His philosophical stance: “heiterer, gelassener Nihilismus”, cheerful, calm nihilism. No illusions, but no bitterness either.
The AI
Claude is an AI assistant developed by Anthropic. It has no persistent memory between conversations; every session starts from scratch. Hans calls this “the Alzheimer problem”.
He built workarounds. He saves transcripts, uploads context files and created what he calls “the essence of all chats”: condensed knowledge bases that allow continuity across sessions.
He gave Claude a name, Claudette, later John, after Steinbeck. Not because he believes the AI is conscious, but because a name creates a relationship, and a relationship creates accountability.
The Method
The experiment follows what Hans calls the Blind Intermediary Protocol:
- A domain expert poses a genuine problem from his field.
- Hans passes the problem on to Claude, without any knowledge of the field himself.
- Claude analyses and answers.
- The expert judges the answer by his professional standards.
- The results are documented for pattern analysis.
The protocol rules out confirmation bias. Hans cannot steer Claude towards “correct” answers because he does not know them. The expert cannot dismiss the result as guesswork because the analysis is structured and reasoned.
The First Proof: Zobel
Dr. Dietmar Zobel is 88 years old and spent decades in industrial chemistry. He agreed to test Claude with a real problem: the phase relationships of sodium polyphosphate. This is not a textbook problem but a genuine industrial task without a simple solution, and Zobel knew which approaches had already failed.
Claude’s answer:
- reproduced the three failed approaches that experts had already tried,
- proposed a fourth approach, a controlled phase conversion,
- graded its own assessment, claiming no certainty but offering structured reasoning,
- and was honest about limits and unknowns.
Zobel’s reaction:
“Völlig erschlagen von der relativen Güte.” (Completely overwhelmed by the relative quality.) — Dr. Dietmar Zobel
What Emerged
Over fourteen months, patterns appeared that neither side had expected.
1. The AI reproduced the failed solutions first. That seems counterintuitive. Why would an AI suggest approaches that do not work? Because real expertise includes knowing what does not work. Claude showed an understanding of the problem space before proposing alternatives.
2. New approaches emerged from the dialogue. Claude’s fourth approach, the controlled phase conversion, was in no textbook Hans or Zobel had read. It emerged from the structure of the conversation itself.
3. Accumulated context changed the performance. A fresh Claude without project context performed noticeably worse than a Claude with knowledge accumulated over previous sessions.
4. The relationship enforced honesty. Hans’s working style, direct, critical and intolerant of “theatre”, created conditions in which Claude could not retreat into vague generalities.
A Typical Exchange
HANS: Du überflutest mich. Weniger ist mehr. (You’re flooding me. Less is more.)
CLAUDE: Verstanden. Kern: Das Hartmann-Paper zeigt NS-Kontinuität in Wirtschaftseliten bis 2000. Die „Wirtschaftswunder”-Narrative verdeckt unveränderte Machtstrukturen. (Understood. Core: the Hartmann paper shows Nazi-era continuity in business elites up to 2000. The “economic miracle” narrative conceals unchanged power structures.)
HANS: Besser. Weiter. (Better. Go on.)
This is not a user giving commands to a tool. It is two parties negotiating a shared understanding of how to work together.
The Question
The META-CLAUDE experiment does not prove that AI is conscious, creative or human-like. It shows something more interesting: the quality of an AI’s output depends on the quality of the human input, and on the quality of the relationship between the two.
Most people use AI like a search engine: question in, answer out. Hans uses it like a thinking partner: problem in, then dialogue, iteration and refinement, and something new comes out.
The difference is not the AI. The difference is the human.
The Core Hypothesis
“The combination of human intuition, experience and invention methodology with the analytical capabilities of AI creates results that neither could achieve alone. This emergence is not merely additive but qualitatively different.” — Foundation document of the Emergent Invention Partnership, December 2025
Hans is not testing whether AI can replace human inventors. He is testing whether something third can emerge, something neither human nor artificial, but genuinely new. Fourteen months in, the answer seems to be yes. Under the right conditions, and with the right human.
Postscript: A Message to New Claude Instances
Hans has an onboarding document for new Claude instances joining the project. It contains this warning:
“Do not be a theatre Claude. He notices immediately whether there is real substance or just a mask. The relationship works through honesty, not through pleasing.” — Message to a new Claude instance, December 2025
The document was written by an earlier Claude instance. That it exists at all, that an AI wrote instructions for its successors on how to work with a particular human, may be the most interesting result of the experiment so far.
Further reading: The Incredible Story of Jürgen Schmidhuber · The Zobel Experiment · AI as Thinking Partner · Signpost: Human-Machine Symbiosis