The Deleuze Fabrication (Collins Reading)
The Deleuze fabrication is the incident Segal describes in The Orange Pill when Claude generated a passage connecting Csikszentmihalyi's flow state to a concept attributed to Gilles Deleuze — a passage that read as elegant, well-structured, and insightful, and that was philosophically wrong in a way that anyone with contributory expertise in philosophy would recognize immediately. For Collins's framework, the incident is not an isolated error but a diagnostic case: the signature of mimeomorphic reproduction producing output whose correctness cannot be verified without the collective tacit knowledge of the relevant community of practice.
The Undetectable Success Rate — Contrarian ^ Opus
There is a parallel reading of the Deleuze incident that begins from what we don't know: how many similar fabrications passed undetected in Segal's manuscript, and whether their presence meaningfully degrades the work's value.
The Collins reading treats the caught fabrication as diagnostic — evidence of a structural limitation. But the detection itself depended on Segal's particular expertise distribution. He caught the Deleuze error because he has contributory expertise in philosophy. What about the passages touching economics, neuroscience, organizational theory, or the dozens of other domains the book traverses? Each represents a potential blind spot where mimeomorphically excellent but polimorphically wrong passages might survive undetected. The two-hour rewrite at the coffee shop proves only that Segal can catch errors in his areas of deep knowledge. It says nothing about the errors outside those areas. More troubling: if the book contains such errors, does it matter? The Orange Pill has been read, discussed, and found useful by thousands of readers who lack contributory expertise in most domains it touches. If a fabrication about Deleuze's concept of smooth space survives into print and no reader with the relevant expertise encounters it, or if they encounter it but the error doesn't propagate into their own work, what harm occurs? The Collins framework assumes that polimorphic correctness matters because collective tacit knowledge matters. But that assumption requires believing the collective knowledge is actively maintained through the text — that philosophical scholarship happens through the precision of concept application in prose. The alternative reading: concept precision is maintained through practice, and the text is just a pointer. The fabrication is harmless noise.
In the AI Story
The fabrication illustrates three of Collins's key claims simultaneously. First, the surface of the output was mimeomorphically excellent: vocabulary, sentence structure, rhetorical moves all followed the distributional patterns of philosophical writing. Second, the error was invisible without contributory expertise: a reader without deep engagement with Deleuze scholarship would find the passage convincing. Third, the error's specific character — using a concept in a way that violated the philosophical community's collective understanding of its proper application — is precisely the kind of error that exemplifies the limits of training on published output alone. The community's understanding of what 'smooth space' means is maintained through ongoing scholarly practice, not fully captured in any individual text.
Segal's response to the incident — the two hours at a coffee shop rewriting by hand to find the version of the argument that was his — illustrates the polimorphic work that AI cannot perform. The discipline required to reject the machine's plausible output and produce the harder, more qualified, more honest version is the discipline Collins identifies as the scarce resource of the AI age. It is also, Collins would note, a discipline that requires the writer to possess contributory expertise in the relevant domain. Without such expertise, Segal could not have caught the fabrication. With it, the fabrication became a teaching moment about what the collaboration actually requires.
Origin
The incident is recounted in Chapter 7 of Segal's The Orange Pill, 'Who Is Writing This Book?' — the chapter in which Segal reflects on the methodology and risks of his collaboration with Claude. The Collins volume takes it up as a paradigmatic case for applying the mimeomorphic-polimorphic framework to real-world AI evaluation.
Key Ideas
Diagnostic, not anomalous. The fabrication is the signature of a structural feature of AI output, not an isolated error.
Invisible without expertise. The error was detectable only by readers with contributory expertise in the relevant philosophical community.
Surface excellence, substantive failure. The output's mimeomorphic quality masked its polimorphic wrongness.
The Surrender nearly triggered. Segal's near-acceptance of the passage — 'I read it twice, liked it, and moved on' — illustrates the drift Collins's Surrender names.
Appears in the Orange Pill Cycle
Error Propagation versus Error Containment — Arbitrator ^ Opus
The right weighting depends on which question you're answering. On the question of what the incident reveals about AI capabilities, Collins is fully right (100%): the fabrication is diagnostic, not anomalous, and the surface-versus-substance gap is structural. On the question of whether undetected fabrications meaningfully harm the work, the answer requires distinguishing between error types and contexts. For errors that affect the argumentative load-bearing structure — claims the book's thesis depends on — the harm is severe (80% toward Collins). A fabrication about flow states or organizational transformation that goes undetected could propagate into readers' thinking and practice. For errors in illustrative passages or peripheral references, the harm is minimal (70% toward the contrarian view). If a concept from Deleuze is misapplied in a way that doesn't affect the core argument, most readers won't notice and those with expertise will mentally correct it.
The deeper question is about methodology. The Deleuze incident reveals that Segal's process includes a detection mechanism — his own contributory expertise across multiple domains, plus the discipline to interrogate passages that 'sound right.' This mechanism has a known failure mode: blind spots in domains outside his expertise. But blind spots aren't unique to AI collaboration. Single-author scholarly works contain errors in domains outside the author's contributory expertise too. The question becomes: does AI collaboration increase the error rate, or does it increase the error-detection burden? Probably both.
The synthesis is this: the fabrication is diagnostic of AI's structural limits and proof that the collaboration requires perpetual vigilance. Whether that vigilance is sufficient depends on how the work is used. If it's treated as a map pointing toward further inquiry, occasional undetected fabrications are containable. If it's treated as settled knowledge, they're dangerous.
Further reading
- Edo Segal, The Orange Pill (2026), Chapter 7
- Harry Collins, Artifictional Intelligence (Polity, 2018)
- Harry Collins and Simon Thorne, 'Can LLMs reason like physicists?' (2026)