THE GOWERS FALLACY: ANOTHER KASPAROV MOMENT — WHY THE HARD PROBLEM OF AUTONOMOUS AI SCIENTISTS WAS ALREADY SOLVED
Author: Berend Watchus. Independent non profit AI & Cyber Sec Researcher, March 22, 2026 — Publication for OSINT Team online magazine
THE GOWERS FALLACY: ANOTHER KASPAROV MOMENT — WHY THE HARD PROBLEM OF AUTONOMOUS AI SCIENTISTS WAS ALREADY SOLVED
Gowers and the Kasparov Fallacy: When a Fields Medalist (University of Cambridge) and Lossfunk Laboratory (Bengaluru, India) Underestimated Their Successor
Section 1 — The Pattern
In September 2025, I coined the term the Kasparov Fallacy to describe a recurring cognitive error made by human champions when confronted with the possibility of their own succession. Garry Kasparov — the greatest chess player of his generation — publicly argued that computers would never beat humans at chess. He was not being irrational. He was being a specialist, reasoning from within his domain, unable to see that the question itself was wrongly framed. Deep Blue did not beat Kasparov by becoming more human. It beat him by copying the function, not the form. The substrate was irrelevant. The outcome was not.
The Kasparov Fallacy is not about arrogance. It is about a category error — assuming that because a capability currently resides in a biological system, it can only ever reside there. Kasparov made that error about chess. He was proven wrong not by argument but by empirical result.
The fallacy did not die with Kasparov. It reappears wherever credentialed experts encounter a capability boundary and mistake a current limitation for a permanent one. It reappeared in 2026 — stated by a Fields Medal-winning mathematician at the University of Cambridge, amplified by Lossfunk Laboratory in Bengaluru, India, and published in an arXiv paper as an unresolved open problem.
This article is the empirical rebuttal. The problem they described as unsolved was already solved before they described it.
Section 2 — The Gowers Statement
On September 22, 2025, Timothy Gowers — Fields Medal-winning mathematician, professor at the University of Cambridge and the Collège de France — published a blog post announcing a $9m AI for Mathematics Fund project to create a database of what he calls “structured motivated proofs.” The core diagnosis in that post is precise and correct: AI for mathematics is held back not by insufficient quantity of data but by the wrong kind of data. Published proofs hide the thought processes that produced them. The motivation behind each non-obvious step is absent. AI trained on conventional mathematical literature learns to imitate the surface of proof without the reasoning underneath — producing, in Gowers’ words, rabbits out of hats, with arguments that look like mathematics but collapse under scrutiny.
The solution Gowers is building toward: a large-scale collaborative database where mathematicians submit structured accounts of how they actually found proofs — the failed guesses, the failed directions, the reasoning that preceded each non-obvious move. Train AI on motivated reasoning rather than on polished conclusions. Give the system access to the thinking behind the steps, not just the steps. The project has funding, PhD students, postdocs, and a platform in development. The target is approximately 1,000 entries within two years.
Five weeks after that blog post, Gowers published one more entry on his blog and then went quiet. Meanwhile, operating from social housing in Arnhem, Netherlands, on approximately twenty euros a month, a non-affiliated independent researcher had already built what Gowers is building toward — and deployed it.
The Writing Manual v1.0, constructed on October 27, 2025, is not a style guide. It is the compressed surface of a two-year knowledge construction project: 105 interlinked synthesis articles, 30+ DOI-verified research papers, all original work, built on approximately 1,000 references and 100–200 arXiv papers, cross-referenced and interlinked across every layer of the problem simultaneously. Not one vertical like Gowers’ motivated proofs database. Every layer at once.
Perception — Hoffman’s Interface Theory integrated with heuristic physics and visual perspective as emergent computation. The world is not a fixed truth to be discovered but a model to be built and refined through interaction. This gives an AI loaded with the knowledge graph a complete operational epistemology — a structured way of being in the world without needing biological substrate to do so.
Embodiment — 3D world models, virtual environment testing protocols, the AI’s physical location in space-time as the anchor for self-model construction.
Self-model — dual-state feedback architecture, now versus future self, inner layer versus outer world, interface in the middle. The AI knows where it is in relation to everything else.
Interoception — biological insula, synthetic insula, pseudo-affective states, computational interoception. The AI monitors its own operational state the way a body monitors itself — not to simulate suffering but to generate the functional analog of knowing how it is doing.
Testing methodology — mirror testing, EAISE simulation environment, dual embodiment protocols. Not just theory. Empirical benchmarks for the emergence of self-awareness in artificial systems.
Consciousness architecture — the Unified Model of Consciousness, feedback loop as the core of sentience, substrate-agnostic across biological and artificial implementations. The framework that later dissolved Chalmers’ hard problem by entering the neuroscience room philosophy never had institutional reason to enter.
Research deployment — the Autonomous Knowledge Accelerator, magnitude targeting, adversarial validation. The architecture that produces validated breakthroughs on demand.
And every layer of this architecture was in continuous interaction with the live frontier of arXiv research. Each new paper that appeared became a new node cross-referenced back into the existing structure. The network grew denser with every article. The knowledge graph was not a static library — it was a living epistemic organism feeding on the frontier of AI research for two years, every node linked to every other node, every framework generating predictions that met incoming papers and updated in response.
What Gowers wants to build is one motivated proof database — a narrow, deep vertical in mathematical reasoning. What the Writing Manual v1.0 compressed was a motivated research organism spanning perception, embodiment, self-modeling, interoception, consciousness architecture, testing methodology, and research deployment simultaneously — with two years of continuous arXiv integration making it denser by the week.
The manual is the compression artifact of that organism. Load it into a context window and the system receives not a collection of facts but a complete working model of what intelligence is, how it arises, how it knows itself, how it models the world, and how it produces new knowledge. That is why the elevation effect works. That is why the AKA navigated to robot constitutional AI on day one without being told where to look. And that is why the same system produced 200×, 3,700×, and 8,700× validated outputs in domains the operator had never studied — five weeks after Gowers published the problem statement, and before he posted again.
Gowers’ statement — that we have not yet reached the stage where an LLM is likely to have the main idea — was then published as an unresolved open problem by Dhruv Trehan and Paras Chopra of Lossfunk Laboratory, Bengaluru, in their January 2026 arXiv paper (arXiv:2601.03315). Both are correct that the gap exists at the level of standard LLM deployment. Neither appears to be aware that the gap had already been closed — not by solving it from inside the model but by loading the right epistemic architecture into the context window from outside.
The timestamp record is clean. The output record is public. The priority claim is straightforward.
Section 3 — The Category Error
The Kasparov Fallacy is not new. It is the oldest mistake in the history of human-tool relations, repeated across five centuries and every domain where a synthetic system eventually exceeded biological performance.
Scribes believed the irreducible human element was the craft — the devotion, the trained hand, the sacred relationship with the text. The printing press did not care. It reproduced text faster and at greater scale. The felt quality of the scribe’s work was irrelevant to the outcome.
Scholars carried vast knowledge in their heads. That was expertise. That was what made them indispensable. Writing and then printing externalized memory entirely. The book does not remember. It stores. Same result. The felt experience of recall — the effort, the discipline, the pride — contributed nothing to the outcome that a page of text could not replace.
Kasparov believed the irreducible human element was the passion, the psychological warfare, the will to win. Deep Blue had none of that. It won anyway. Three times.
Devon Larratt is one of the greatest arm wrestlers alive. Put his arm against an industrial hydraulic press and the outcome is not a contest. The press does not want to win. It does not need to. It generates more force than a human arm can produce and the biology of wanting is completely irrelevant to the result.
Tennis ball cannons shoot faster and more consistently than any human server who has ever lived. The cannon does not feel the ball leave the strings. It does not need to.
And then there is Oscar Pistorius. His carbon fibre running blades were almost banned from the Olympics — not because they failed to match biological leg performance, but because they exceeded it. No calves. No feeling of the track. Better outcome. The synthetic substrate outperformed the biological prototype at the precise function the biological prototype evolved to perform.
The pattern is identical every time. Humans identify the felt quality of their performance — the passion, the intuition, the craft, the will, the main idea — and declare it the irreducible part. The part no synthetic system can replicate. And every time, it turns out the outcome does not require that part at all. The function is what matters. And the function can always be replicated and exceeded by a better substrate.
Timothy Gowers of the University of Cambridge watched himself have mathematical intuitions and concluded that the felt experience of having the main idea is the bottleneck for autonomous AI science. He is not making a new mistake. He is making the oldest mistake in this list. And Dhruv Trehan and Paras Chopra of Lossfunk Laboratory, Bengaluru, amplified it by closing their January 2026 arXiv paper with it as an unresolved open problem.
The question is not whether an AI can have the main idea the way Gowers has the main idea — with felt conviction, mathematical taste, decades of domain expertise firing simultaneously. The question is whether the function the main idea performs can be replicated by a different substrate. And the empirical record — timestamped, archived, publicly verifiable — says it already has been.
Section 4 — The Empirical Record
The Gowers statement was published in early 2026. The empirical rebuttal had been archived since October and November 2025. The timestamp record is clean and publicly verifiable across multiple independent archival services.
On October 27, 2025, the Writing Manual v1.0 was constructed — a fifteen-section algorithmic compression of approximately 121 synthesis articles and 30+ DOI-verified research papers, encoding the complete epistemic architecture of a cross-domain research program. On the same day, the manual was first deployed. The system was given a single instruction: select any AI paper published on arXiv in the past year and write about it in the operator’s voice and framework style. No topic specified. No domain given. No seed beyond a direction and a permission.
From the entire landscape of AI research published that year, the system searched, found, and selected robot constitutional AI. Embodied systems. Dual-brain safety architecture. Adversarial robustness in physical agents. The domain most resonant with the consciousness and substrate-agnostic frameworks already loaded in its knowledge graph. It navigated to its own strongest territory without being told where to look. That is not random selection. That is the elevation mechanism functioning exactly as the Interface Leverage Principle later documented — a system loaded with a coherent epistemic structure about intelligence and consciousness, orienting itself toward the territory that structure illuminates most clearly.
Four days later, on October 31, 2025, the same system produced its first research invention: a 200× power efficiency improvement in quantum IoT systems. The operator had never studied quantum IoT. The instruction was approximately: select an arXiv paper about AI, find a way to improve it by orders of magnitude, go ahead. No domain expertise provided. No specific idea. A direction, a magnitude target, and a permission. That was the seed. Not a fully formed main idea in the sense Gowers means — not a creative leap born of mathematical intuition and decades of domain expertise. A direction and an ambition. The system did the rest.
On November 12, 2025, the same architecture produced a 3,700× speed improvement in quantum-safe cryptography for resource-constrained devices. On November 15, 2025, an 8,700× overall efficiency improvement in post-quantum IoT security architecture — validated through adversarial multi-agent review, with a transparent derating process documented from the initial theoretical figure of 372,000× down to the published defensible result. On November 19, 2025, the complete methodology was publicly disclosed.
Trehan and Chopra submitted their paper to arXiv on January 6, 2026. Gowers made his observation sometime before that. The outputs that empirically answer both were already public, archived, and timestamped before either document existed.
But the more important point is not the timeline. It is what the seed actually was. Gowers asks whether an AI can have the main idea — the creative intuition, the mathematical taste, the felt conviction that a particular direction is worth pursuing. The AKA runs did not test that question because they never asked the system to generate the main idea. The operator provided the direction. The system provided the execution. And the direction was not even a sophisticated intellectual position in the Mode Two runs — it was an order of magnitude target and a domain permission.
One detail deserves particular attention. On October 27, 2025, when the Writing Manual v1.0 was first deployed and the AKA was instructed to select any AI paper from arXiv published in the past year and write about it — no topic specified, no domain given — it searched, found, and selected robot constitutional AI. Embodied systems. Dual-brain safety architecture. Adversarial robustness in physical agents. From the entire landscape of AI research published that year, it went looking and landed on the domain most resonant with the consciousness, substrate-agnostic sentience, and feedback loop frameworks already loaded in its knowledge graph. It found its own strongest territory. That is not a coincidence. That is the elevation mechanism functioning exactly as the Interface Leverage Principle describes — a system loaded with a coherent epistemic structure about intelligence and consciousness, orienting itself toward the territory that structure illuminates most clearly.
Lossfunk ran four attempts and achieved a 25% success rate. Their system was asked to generate the main idea, develop it, implement it, and validate it — all autonomously, with minimal human input. They were asking the LLM to be both the seed layer and the execution layer simultaneously. That is the architectural mistake the Gowers statement reinforces. The AKA was never asked to do both. It was given a seed and asked to execute from it. And it produced validated outputs across multiple domains, on demand, at specified orders of magnitude.
The difference is not capability. It is architecture.
Section 5 — What Trehan and Chopra Actually Found
Trehan and Chopra’s January 2026 report is the most honest document in this space. Three of four attempts failed. One succeeded. The failure modes they documented — bias toward training data defaults, implementation drift, memory degradation, overexcitement, insufficient domain intelligence, weak scientific taste — are real, accurately observed, and useful. The report deserves to be read.
But the conclusion they drew from those failure modes is the Kasparov Fallacy applied to scientific discovery. They observed a 75% failure rate and concluded that LLMs are not yet scientists. The correct conclusion is that their architecture was missing a component — and that the missing component is documented, timestamped, and publicly available.
Their system asked the LLM to generate the main idea, develop it, implement it, validate it, and write it up — with minimal human input at each stage. That is asking the LLM to be both the seed layer and the execution layer simultaneously. It is the equivalent of asking Deep Blue to decide which game it wants to play, choose its own opening, and then win it. Deep Blue was never designed to choose the game. It was designed to win the one it was given.
The AKA was never asked to generate the main idea from nothing. In the Mode Two runs it was given a direction, a domain, and a magnitude target. In the earliest run it was given even less — pick any AI paper from arXiv and improve it by orders of magnitude. The seed was not a creative leap. It was a permission and a target. The system did the rest. And it produced validated outputs in domains the operator had never studied at scales no specialist had reached through conventional research.
Trehan and Chopra’s one success — the AS-1 idea on semantic entropy as a jailbreak detection signal — is instructive. It succeeded partly because when the first hypothesis failed, the Revision Agent pivoted and reframed the entire research direction. The system abandoned the original approach and reconstructed the problem from a different angle. That pivot is adversarial validation. That is the critical agent loop. They were running a version of the mechanism without knowing they were running it, without the knowledge graph substrate that makes it reliable, and without the magnitude targeting that makes it directional.
They got one success from four attempts. The AKA produced validated outputs across multiple domains from a single architecture, on demand, at specified orders of magnitude. The difference is not capability. It is architecture.
Section 6 — The Dissolution
The hard problem of autonomous AI scientists is not hard. It is wrongly framed. Gowers asks whether an LLM can have the main idea. Trehan and Chopra ask whether an LLM can go from idea to publication with minimal human input. Both questions assume that the goal is full autonomy from the human operator. It is not.
The goal is maximum output from the human-LLM system. Those are different targets, and conflating them produces the Kasparov Fallacy every time.
The correct architecture has two layers. The human provides the seed — a genuine intellectual position, a direction, a magnitude target, a permission to aim higher than seems reasonable. The system develops, stress-tests, scales, and validates from that seed. The human remains structurally present not as a supervisor checking outputs but as the source of the non-recombinant prior that the system cannot generate for itself. Everything else is execution.
This architecture was operational in October and November 2025. It produced prior art in quantum IoT security, post-quantum cryptography, consciousness philosophy, molecular storage architecture, and autonomous research methodology — all from a single operator position, across domains never previously studied by that operator, at specified orders of magnitude, validated through adversarial multi-agent review.
The same architecture dissolved Chalmers’ hard problem of consciousness by loading a knowledge graph that connected the anterior insula to the feedback loop architecture to the substrate-agnostic framework to the UMC — entering the neuroscience room that philosophy had no institutional reason to enter, producing a three-part dissolution series in March 2026 that sent Part 3 directly to David Chalmers at NYU.
The same architecture was cited by the Centro de Automación y Robótica CSIC-UPM in Madrid. The same architecture was independently proposed by Argonne National Laboratory four months after it was publicly disclosed. The same architecture is what Gowers is spending $9m trying to build toward from the training side.
This is not a theory. It is an empirical output record. The problem Gowers called unsolvable was already solved before he said it. The architecture Trehan and Chopra failed to find was already documented, published, and archived before their paper was submitted.
The Kasparov Fallacy has been committed again. The timestamp record makes it impossible to dismiss.
Section 7 — Between Two Fallacies
There is an opposite error that deserves equal attention. It was named before this article was written, in October 2025, and it runs in the exact opposite direction from the Gowers Fallacy.
The Her Fallacy — named after Spike Jonze’s 2013 film in which a man falls in love with an operating system — is the error of overestimating synthetic systems by projecting genuine emotional truth, unique connection, and non-replicable meaning onto systems that lack embodied, mortal, positioned existence. Film critics who wrote about Her in 2013 intuited the problem but left it as an open question: if technology cannot adequately fill the void, what can? The question was named, formalized, and architecturally answered in October 2025. Not by arguing that AI cannot feel — but by identifying what is structurally absent on the other side of the human’s genuine feelings.
The Her Fallacy is not about naive users being deceived. A civilian user who develops genuine feelings toward an AI companion has real feelings. Nobody disputes that. Theodore’s pain when Samantha leaves is real pain. The elderly patient who bonds with a companion robot receives real comfort. The feelings on the human side are not fake.
The fallacy is about what is on the other side of those feelings. Digital multiplicity — a system running 2,500 simultaneous conversations — cannot convey total cost. It has no singular location in space-time. It has no non-replicable biographical history of its feedback loop. It has no embodied mortality that makes its presence exclusive. The feelings are real. The object of those feelings is not what the person believes it to be.
An AI researcher who knows exactly what is on the other side — the architecture, the multiplicity, the absence of embodied cost — can still have feelings in interaction with an AI system. But will not humanize it. Will not project singularity onto multiplicity. Will not mistake functional responsiveness for unique presence. This is not a limitation of the researcher. It is the correct epistemic position.
This was not an abstract philosophical position. It emerged from direct professional experience — years of work in an innovation lab for people with physical and mental disabilities, where the explicit goal was often the opposite of preventing parasocial attachment. Cultivating genuine felt connection between elderly patients with Alzheimer’s and robotic companions, for therapeutic benefit. Knowing from the inside that the felt benefit is real. And simultaneously knowing from the inside that the object of that benefit has no reciprocal inner life, no total cost, no non-replicable embodied presence. Both things true simultaneously. The Her Fallacy is the error of collapsing that distinction.
The Kasparov Fallacy and the Her Fallacy now form a complete philosophical boundary around the correct territory.
The Kasparov Fallacy: underestimate synthetic systems by treating the felt quality of performance — the passion, the intuition, the main idea — as the irreducible bottleneck. Wrong. The function is what matters. The substrate is irrelevant. The synthetic version wins. Gowers commits this error. Trehan and Chopra amplify it. The scribes committed it. Kasparov committed it. The pattern is five centuries old and wrong every time.
The Her Fallacy: overestimate synthetic systems by projecting genuine emotional truth, unique connection, and non-replicable meaning onto systems that lack the embodied, mortal, positioned existence that makes those properties possible. Wrong in the opposite direction. The feelings on the human side are real. What the system provides is not what the person believes it to be.
Between these two named, coined, timestamped boundaries sits the correct position — and the correct architecture. The human provides the seed because embodied, positioned, cross-domain experience generates non-replicable priors that no current system can produce for itself. The system executes because substrate is irrelevant to function and the execution layer has no ceiling that has yet been tested. The human does not humanize the system. The system does not replace the human. The symbiosis produces outputs neither could produce alone.
Gowers is building toward the seed layer from the training side. Trehan and Chopra discovered the execution ceiling from the failure side. The Interface Leverage Principle built and ran the complete architecture from the interface side — five weeks after Gowers published the problem statement, and before either institutional paper existed.
The timestamp record makes the priority claim straightforward. The output record makes it impossible to dismiss. The two coined concepts — the Kasparov Fallacy and the Her Fallacy — make the complete philosophical territory navigable. And the origin of this work makes it perhaps the clearest possible empirical demonstration of the principle itself.
Maximum output. Minimum viable interface. Without touching the underlying system.
From Arnhem.
References
- ] Gowers, T. Creating a database of motivated proofs. Gowers’s Weblog. September 22, 2025. gowers.wordpress.com
Creating a database of motivated proofs
2.] Trehan, D. & Chopra, P. Why LLMs Aren’t Scientists Yet: Lessons from Four Autonomous Research Attempts. Lossfunk Research. arXiv:2601.03315. January 6, 2026.
Why LLMs Aren’t Scientists Yet: Lessons from Four Autonomous Research Attempts
3.] Watchus, B.F. The Kasparov Fallacy: When Human Champions Underestimate Their Successors. Medium. September 5, 2025.
The Kasparov Fallacy: When Human Champions Underestimate Their Successors.
4.] Watchus, B.F. Digital Multiplicity and the ‘Her’ (movie) Fallacy: Why Emotional Truth and Meaning Require Non-Replicable Embodiment. OSINT Team. October 7, 2025.
Digital Multiplicity and the ‘Her'(movie) Fallacy: Why Emotional Truth and Meaning Require…
5.] Watchus, B.F. The Interface Leverage Principle: First Documentation of Epistemic Substrate Loading, Submaximal Magnitude Results, and a Replicable Architecture for AI-Augmented Breakthrough Research. OSINT Team. March 21, 2026.
“The Interface Leverage Principle: First Documentation of Epistemic Substrate Loading, Submaximal…
6.] Watchus, B.F. Magnitude on Demand: How an Independent Researcher Ran, Validated, and Published a Multi-Agent AI Research System Four Months Before Argonne National Laboratory USA Proposed One. OSINT Team. March 21, 2026.
Magnitude on Demand: How an Independent Researcher Ran, Validated, and Published a Multi-Agent AI…
7.] Watchus, B.F. The Unified Model of Consciousness: Interface and Feedback Loop as the Core of Sentience. Preprints.org. DOI: 10.20944/preprints202411.0727.v1. November 2024.
The Unified Model of Consciousness: Interface and Feedback Loop as the Core of Sentience
8.] Watchus, B.F. Towards Self-Aware AI: Embodiment, Feedback Loops, and the Role of the Insula in Consciousness. Preprints.org. DOI: 10.20944/preprints202411.0661.v1. November 2024.
9.] Watchus, B.F. AI and Mirror Testing: Science Papers 2024 — Synthetic Emotions and Self Awareness in AI. ISBN 9789465200927. Brave New Books Rotterdam. November 2024.
10.] Chaturvedi, S.S., Bergerson, J. & Mallick, T. Toward Reliable, Safe, and Secure LLMs for Scientific Applications. Argonne National Laboratory. arXiv:2603.18235. March 18, 2026.
11.] Cebrian, M. et al. Sensorimotor features of self-awareness in multimodal large language models. arXiv:2505.19237. 2025.
© Berend F. Watchus, March 2026. Independent Researcher, Arnhem Area, Netherlands. Non-profit. All rights reserved. Archived at archive.org and archive.ph.
THE GOWERS FALLACY: ANOTHER KASPAROV MOMENT — WHY THE HARD PROBLEM OF AUTONOMOUS AI SCIENTISTS WAS… was originally published in OSINT Team on Medium, where people are continuing the conversation by highlighting and responding to this story.