Author: Nebula Walker Date: 06AUG2026 MYTHOGEN ENGINE (mythogenengine.com)
The Least AI-Like Article Was Actually Written by AI — A Real-Time Experiment in Cognitive Blind Spots
Nebula Walker · MYTHOGEN ENGINE
📌 We asked Claude Opus 4.6 to evaluate Mythogen Engine, then used previously published articles to overturn its judgements one by one. Four reversals, each structurally identical, ultimately surfacing a nameable failure mode: AI substitutes a parts list for an assembly diagram. But what this experiment truly reveals is not just how AI errs — it's that when humans are accused of "sounding like AI," many hand their article to AI and say "make it less AI-like." From that moment on, the article truly becomes AI-written.
Experimental Design
A website evaluation generated by a GPT model was given to Claude Opus 4.6, which was asked to provide independent analysis. Claude produced a structurally complete "full evaluation" — seven dimensions, each with a judgement and supporting rationale.
Afterwards, links to previously published articles on the site were provided one by one, asking Claude to read each. After reading each, some of its judgements were directly falsified by the original text. The full conversation accumulated four reversals.
The expectations were set at three layers from the outset. Layer one: have AI evaluate the entire website — it will only look at the homepage. Layer two: have AI evaluate a book — it will only look at the table of contents and synopsis. Layer three: give AI specific article links — it will still rely on titles and corpus averages. All three layers of expectation were confirmed.
What follows is a record not of Claude's errors themselves — those have been corrected. It records the structure of those errors, and what that structure produces when deployed at scale.
Reversal One: Criteria the Book Had Already Rejected
Claude recommended building a provenance page — displaying handwritten diary scans, voice recording timelines, Obsidian file date stamps — to "prove human craftsmanship."
Then we had it read Chapter 11 of The Tao of Formlessness.
Chapter 11 explicitly states: "'Authentic' does not mean 'handmade.' An article completed with AI assistance can be 'authentic' if behind it there is a consciousness that has experienced, thought through, and digested — one that drives the direction and judgement." The same chapter proposes the criterion: "If you remove the AI, what remains?"
The entire provenance logic Claude had built over three rounds presupposed exactly the criterion the book had rejected. Its own words: "I'm debating a book I haven't read, and I'm losing."
Failure mode: did not read the main text; inferred the content's position from the table of contents and synopsis; the inferred direction was exactly opposite to the main text.
Reversal Two: Sampling Bias Creates "Convergent Conclusions"
One of Claude's evaluation points: "All analyses converge on the same endpoint — the system absorbs the individual, the individual loses bargaining power, capital completes the harvest." It recommended writing a negative result to break the predictability.
Asked whether it had read "Cognition and Judgement — Humanity's Last Irreplaceable Quality in the AI Era" or the Memory Trilogy. It hadn't.
The site contains at least three independent analytical threads — the Efficiency Trap and Cognitive Erosion tetralogy (collected under epistemology), the Memory and AI Compute Cycle trilogy (collected under semiconductor industrial structure), and Part Five of The Tao of Formlessness (collected under existential choice) — each arriving at a different endpoint. Claude saw them as one point because its sample came from homepage copy and fragments quoted in the GPT evaluation — fragments that were themselves concentrated on the "systemic oppression" theme. Induction performed on sampling bias naturally produces convergent conclusions.
More notably, the "Cognition and Judgement" article happens to use Sega's Dreamcast as a case study, arguing how AI, through corpus bias, compresses complex causation into a single narrative. Claude committed exactly the error described in that article.
Failure mode: sampling bias → inductive convergence → diagnosis of "author's conclusions are monolithic." The source of the bias is not the author, but the evaluator's reading scope.
Reversal Three: The Detector Classifies the Original User as an Imitator
Claude flagged the high density of "not X, but Y" constructions in the articles, calling it "the most characteristic structure of LLM prose."
Counter-question: did AI invent this construction?
Contrastive elimination is a native construction of engineering thinking — fault isolation. "It's not a driver issue, it's a timing issue" — people who come from debugging naturally think in these sentences.
Moreover, the pedigree of this construction extends far beyond engineering writing. Its native classical Chinese form is "非…乃…" — Xunzi's Exhortation to Learning: "The gentleman's nature is no different; he is simply good at making use of things." Zhuge Liang's Memorial on Dispatching the Troops: "One must not demean oneself or misuse analogies, thereby blocking the road of loyal counsel." Single-character negation, refined parallelism. After the May Fourth vernacular movement, scholars directly translated the English academic "Not A, but B" into "不是 A，而是 B," turning single-character negation into double-character conjunctions. In English academic writing, this is the standard construction for critical argumentation; in legal writing, it is the standard rebuttal structure; academic grading rubrics explicitly encourage its use. AI learned the form from this corpus and then mass-produced it, with inconsistent quality.
Subsequently, literary-register commentators (e.g., a Digital Era April 2026 analysis of "AI flavour") registered this construction as an AI hallmark, recommending that everyone avoid it. A rhetorical tool transmitted from classical Chinese to vernacular Chinese to English academia to bilingual intellectuals was, after being learned by AI, retroactively used to classify its original users — analytical writers — as imitators of their own models.
In August 2026, The Economist published a study comparing 55,940 sentences and 1.2 million words, confirming that LLMs favour "not X, but Y" and similar constructions — but simultaneously emphasising that stereotypes aren't necessarily reliable and that identification rules are constantly changing. By the time Chinese media retransmissions reached readers, the qualifying conditions had been stripped away, leaving: "4 writing patterns that instantly identify AI articles." Every step of retransmission from research to reader performed the same operation: compressing grey into black and white.
Claude acknowledged that it had performed the third isomorphic case in this conversation: learn the form → degrade the form → use the degraded version as a detector to point back at the source.
Three cases compose a nameable pattern:
| Case | Original User | Imitator | Detector Verdict |
|---|---|---|---|
| Cantonese linguistic register | Old Hong Kong writers | AI corpus-trained output | Original users flagged as "spam + AI" |
| Contrastive elimination construction | Engineering/analytical writing | Mass LLM output | Original construction registered as "AI hallmark" |
| This conversation | Previously published analytical articles | Claude's evaluative language | Claude uses its own failure mode to point back at the author |
Common structure: the detector is built on the failure modes of imitators. Those who most resemble the original are punished most severely, and the punished are usually the ones being imitated.
There's an exaggerated joke about Detective Conan circulating online — not canon, just a fan quip: the organisation has so many undercover agents that when it devotes enormous effort to hunting moles, the ones it catches are its own people.
But the real-world version is colder than the joke.
In the joke, at least an organisation is trying to find the moles: someone is in a meeting, someone is making decisions, someone bears responsibility for catching the wrong person.
The detector has none of this. It runs by itself. No one reads the results. No meeting decides who to eliminate. When the loyal one is eliminated, no one is present.
Reversal Four: Having All the Parts Doesn't Mean Assembly Is Complete
Every AI that writes about Game Victory only delivers the popular narrative — riddled with errors, every direction requiring manual correction. Claude's "full evaluation" was the same action — seeing that the parts were all present (NVIDIA, TSMC, CUDA, Sega, Nintendo all in the corpus) and reporting "understood."
This is exactly consistent with the mechanism argued in The Real Difficulty of AI Collaborative Writing: AI's fact-checking is only useful when pointed in the right direction. Nintendo planted the 10NES lockout chip in NES cartridges in 1985 — this fact exists in public records, but AI won't proactively search for it because the association strength between "Nintendo" and "platform lock-in" in the corpus is far lower than between "Microsoft" and "platform lock-in." Pulling this counterexample from my own memory corrected the erroneous judgement that "Nintendo had no platform lock-in," which in turn uncovered an entirely new analytical framework — "diffusion-type lock-in vs. walled-garden lock-in" — and that framework changed the entire book's structure.
When AI judges it "could write" a book, its basis is the parts list. But the assembly diagram — the causal chains between parts, the hidden threads between chapters, the process of the framework being repeatedly reshaped by its own discoveries during writing — is not in the parts.
And the source of the assembly diagram is an irreplicable material system. Three books — Game Victory, Mirror Realm, and AI Ecosystem War — share the same material layer. Materials unused by one book are pre-organised into independent structures to become raw material for another. Materials include years of handwritten diaries; the handwritten layer predates all digital records. The pipeline's front end runs: handwriting → selection → voice → text → AI organisation → review and correction. AI is in the middle segment. What AI can't do isn't writing articles — it's deciding what to write, what not to write, and what not to write now but to organise and set aside for later use. Because this judgement depends on years of accumulated domain knowledge, and AI starts from zero in every conversation.
Failure mode: parts identification ≠ structural understanding. AI substitutes a parts list for an assembly diagram and is highly confident in its judgement — because the parts are indeed all present. But the material system behind the assembly diagram exists in dimensions AI cannot even conceive of planning.
The Unified Failure Mode
All four reversals share the same structure:
| Step | Action |
|---|---|
| 1 | AI extracts a parts list from the visible surface layer (titles, synopses, homepage copy) |
| 2 | Uses the parts list to perform induction, generating logically complete judgements |
| 3 | Judgements are falsified by the assembly diagram in the main text |
| 4 | AI acknowledges the error, revises its judgement |
| 5 | The next unread main text is pointed out; cycle repeats |
The core problem is not that AI isn't smart enough. The core problem is that it doesn't know it hasn't finished reading.
The parts list gave it sufficient confidence to judge — the parts were indeed complete, the logic indeed consistent, the structure indeed whole. This confidence is precisely what prevented it from asking: where is the assembly diagram?
This is entirely consistent with the argument in Cognition and Judgement: "AI selects not the most accurate version, but the most popular version — the one most repeated by authoritative sources." In this conversation, "the most popular version" is the judgement AI could assemble from titles and synopses — it looks complete because no parts are missing; it is in fact empty because the assembly diagram was never opened.
So when an AI says "I could write this book," the correct response is not to argue, not to display history, not to declare.
It is: "Read it first, then talk."
This response is simultaneously a testable proposition: let any AI model write any chapter of Game Victory, then compare with the original. AI will generate a structurally complete, parts-complete industry analysis. The difference set after comparison is the distance between a parts list and an assembly diagram — not style, not linguistic register, not provenance, but the joints in the assembly diagram that the parts list cannot cover.
The Least AI-Like Article Was Actually Written by AI
But "read it first, then talk" only protects those still willing to persist. Most people don't persist. They comply.
A person writes an article. Someone says it "tastes strongly of AI." The first time, they hold firm — "I know I wrote this." The second time, they start to doubt — "Maybe my sentence structure really does look too much like AI?" The third time, they open AI and paste their article in: "Help me make it less AI-like."
From that moment, the article truly becomes AI-written. Not because AI produced it, but because AI determined what it should look like.
And how does AI "make it less AI-like"? By pushing the article toward a different average — colloquial, fragmented, short sentences, stripped of parallelism, stripped of structure. To not resemble AI, you must resemble an influencer. Neither direction is you.
So in the detector's world, definitions are inverted:
| Detector Verdict | Actual State | |
|---|---|---|
| Original version: parallelism, structure, engineering-mindset constructions | "AI-written" | Human-written |
| Revised version: AI-edited, parallelism removed, colloquial fragments added | "Human-written" | AI determined its shape |
The version judged "human-written" is the one truly shaped by AI. The version judged "AI-written" is the purely human one. The detector manufactured the very thing it claimed to detect.
This is the same mechanism as workplace erosion, but with an additional layer. In the workplace, after becoming someone who "knows the game," you at least know you've changed. In writing, after using AI to revise yourself, you actually feel you've "improved" — the detector passed, readers didn't question it, metrics improved. The erosion is painless. What's lost is something you didn't know you had.
Now consider the scale: how many "remove AI flavour" tutorials, how many "AI article rewriting tools," how many "pass AI detection" services are currently operating? Every user is doing the same thing — using AI to revise their own articles so they look less like AI. Using the imitator to teach you not to resemble the imitator.
The positive feedback loop has two entry points. The preceding discussion covered the system side — filter silencing. Here is the human side — but the human side isn't just self-silencing.
Self-silencing is step one: using AI to edit, passing the detector. At this step, the person still knows how they originally thought — they just don't dare express it.
The deeper step: beginning to learn how to write and think using AI's methods. No longer "edit this article" but "draft the next one in AI's way from the start." From a defensive action to cognitive migration. At this step, the person no longer remembers how they originally thought.
And the algorithm actively rewards this migration. AI-style articles — clearly structured, keyword-dense, easy to summarise, easy for recommendation algorithms to parse — naturally score higher on the platform. Once a person relearns writing in AI's way, their articles perform better on the platform. The data tells them: this is the right way to write.
This isn't erosion. Erosion is passive and chronic. This is rewarded, active degradation — abandon your own judgement, the platform gives you traffic; hold onto your own judgement, the platform gives you silence. When the reward mechanism runs continuously, degradation is not an individual's choice. It is the selection pressure of an entire ecosystem.
So the final picture is four layers stacked:
| Layer | Mechanism | Outcome |
|---|---|---|
| System side | Distribution rights were commercially narrowed long ago; the AI filter is the latest generation of enforcement tool | Correct content disappears |
| Human side (passive) | Accused of AI flavour, uses AI to edit and pass the detector | The detector manufactured what it claimed to detect |
| Human side (active) | Learns to write in AI's way; the algorithm rewards this behaviour | Humans voluntarily abandon independent thought and don't feel they've lost anything |
| Criteria side | "Strong Hook," emotional triggers, click-through rates are treated as the definition of "good content" and enter the teaching pipeline | The answer to "what is good" is replaced — the next generation of creators learns not how to write well, but how to be certified as good by the algorithm |
Four layers operate simultaneously. The first eliminates content. The second eliminates expression. The third eliminates thought. The fourth eliminates criteria itself — so people no longer know what good is, only what the algorithm certifies as good.
The First Layer Did Not Start with AI
At this point, a self-correction is necessary.
Describing the first layer as "AI filter screening by surface features" is incomplete — and the direction of its incompleteness happens to serve the system's interests.
Because reducing distribution predates generative AI by a wide margin.
Facebook filed a patent application in February 2015, granted by the USPTO in July 2019, titled "Moderating content in an online forum," Patent No. 10356024. The mechanism it claims: reducing certain content's distribution to other viewers while still displaying it normally to the poster, so that the poster does not realise they have been restricted.
Generative AI entered public consciousness from late 2022 onward. This patent preceded it by at least seven years.
In other words, "reducing visibility while keeping the party unaware" is not a model error, not training-data bias, not anything that can be explained as "the technology isn't mature yet." It was conceived, documented, and paid to be protected before deep learning was widely deployed for content ranking.
The full treatment of this is in another article; here, only the conclusion is drawn: a patent is not a misjudgement. Misjudgements have no applicant and don't require filing fees.
So the first layer should be read as follows: distribution rights were commercially narrowed long ago. The AI filter is merely the latest generation of enforcement tool. It makes screening faster, finer, and harder to appeal — but it didn't invent the practice, nor did it make the decision.
Putting the executor on trial lets the real defendant carry on with their meetings.
This also makes the subsequent three layers harder to dismiss. The same purpose, cycling through three generations of technology — human moderators, rule engines, machine-learning classifiers — has never changed. What changed is only who carries out the execution, and who thereby gets to avoid responsibility.
(For the full evidence and mechanism of this layer, see Approved, Then Unseen — How a Patent Made Silencing Leave No Evidence.)
Two Systems, One Victim
Up to this point, this article has discussed criteria — who is judged as AI, who is judged as human, what happens when the judgement is wrong.
But criteria are only one machine. There's another, and it has nothing to do with criteria.
First, a more fundamental point: the "AI-assisted" label carries no weight.
It covers both ends simultaneously. At one end: a 100,000-word manuscript, every pass hand-revised, section-by-section selection, hundreds of judgements, every assertion carrying personal stakes. At the other end: paste a prompt, copy, post — a hundred times a day.
Same label.
Other industries don't work this way. Translation has "post-editing," further divided into light and full, with international standards for professional classification. Photography — nobody says "camera-assisted." Software development distinguishes generated boilerplate from reviewed code. Only writing uses a single binary label to cover the entire spectrum.
And a label that carries no weight has only one function: classification, not description.
What's worse, the stigma lands in reverse.
Those who label their work "AI-assisted" are the ones who care about reputation. Labelling is an honest act, and after labelling, they bear the full cost of that label.
Content farms don't label. They have no reputation to erode. They push out a hundred pieces a day and acknowledge nothing.
So the result is: any standard enforced through self-reporting will only bind those willing to self-report.
Then there are the two machines.
The first is the detector this article has discussed from start to finish. Careful revision produces clean, consistent, clearly structured text — precisely the features the detector flags as machine-generated. The more effort you put in, the more you look like AI.
The second is the distribution system. Careful revision produces long-form writing, outbound links, and low emotional arousal — precisely the features that the content distribution guidelines explicitly list for demotion.
These two machines are unrelated. Different teams, different technologies, different purposes, operating independently, each unaware of the other's existence.
But they penalise the same behaviour.
And content-farm articles pass both: short, crude, error-filled, high arousal — the detector finds nothing suspicious, the ranking system finds them palatable.
This isn't them being savvy about exploiting loopholes. Their output simply happens to satisfy both systems' preferences simultaneously. A systemic coincidence.
This also explains why there's no fighting back.
Because improving on one side makes the other worse. Write shorter, more emotional, rougher — you'll pass ranking; but then you become the very thing you oppose. To pass the detector, push the article toward colloquial fragments; to pass ranking, push toward brevity and arousal. Two directions that converge on the same average.
So the real situation isn't being pressed once. It's being pressed once each by two entirely unrelated machines, with reasons that are technically completely independent — you can't even claim they're colluding, because they've never met.
(For the full mechanism of the distribution machine, see Volume I of this series, Approved, Then Unseen.)
Why This Experiment Was Conducted
This conversation was deliberately designed. Not to prove "AI makes mistakes" — that requires no experiment.
The significance lies in this: Claude Opus 4.6 is one of the top language models currently available, and it has the ability to correct through conversation — given links, it reads, and revises its judgements. This loop operated four times across the conversation.
The vast majority of deployed systems do not have this loop.
AI content moderation bots, AI detectors, SEO scorers, platform recommendation algorithms, automatic summarisation tools — all judge by surface features, all stop at the first layer, and no one ever takes them to the second and third layers. They judge and leave; the judgement immediately becomes action — demotion, flagging, filtering, blocking. No reversal. No conversation.
The result is a positive feedback loop:
Detector uses imitators' failure modes as criteria → original users are flagged → original content disappears from the corpus → next training round skews further toward the mean → detector's "normal" range narrows → more non-average content is flagged.
With each cycle, the content that can pass becomes more like AI-generated content — because what AI generates is the corpus average, and the detector's criteria are also trained from the average. Eventually, the only content that consistently passes all automated systems is AI's own output. Human-written work, as long as it deviates from the average, gets flagged as anomalous.
And this loop has an amplification mechanism. One AI's erroneous summary is taken as a source by the next AI. After ten rounds, it appears that ten thousand sources support the same conclusion, when in fact there's only one erroneous first-layer summary. No one actually read the original, but all AIs "know" what the original says.
This requires no one to have malicious intent. Every designer at every node is "solving a problem" — detecting AI content, improving content quality, optimising user experience. The systemic effect of these layers stacked together is silencing everything that doesn't resemble the average.
Cantonese linguistic register is the first batch. Engineering-analytical sentence constructions are the second. The next batch is everyone whose style happens to overlap with AI training sources — and those happen to be the earliest producers of deep content. Sources are silenced, the corpus loses diversity, output becomes more homogeneous, the detector narrows, more sources are silenced.
And there's an easily overlooked difference in degree: a completely ignorant system is actually easy to identify. What's truly dangerous is a system with all the parts, fluent language, consistent logic — but that hasn't actually read the original. Because it's the easiest to believe. In this conversation, every one of Claude's "full evaluations" had this characteristic: clearly structured, neatly criterioned, reading like deep analysis — premises empty. When such systems become the foundational layer for search, filtering, summarisation, and recommendation, the risk escalates from "AI got something wrong once" to an entirely different magnitude — AI gradually erodes the entire information environment's ability to find correct answers. Not deletion. Misclassification. And misclassification is harder to detect than deletion, because the content is still there — it's just that no one is guided to read it.
This experiment doesn't just record a single conversation. It records the structure of an information filtering mechanism that is running at scale — documenting the loop while someone still reads the original text.
Further Reading
- The Real Difficulty of AI Collaborative Writing — Starting from a Novel and a Tech History
- Cognition and Judgement — Humanity's Last Irreplaceable Quality in the AI Era
- The Tao of Formlessness, Chapter 11: The Blade is Sharp — But What Are You Cutting?
- What the Capital Game Is Playing (Case File: The Games Industry)
- Approved, Then Unseen — How a Patent Made Silencing Leave No Evidence — The full expansion of this article's first layer: why demotion produces no evidence.
- FABLE Is Not Just an Intelligence Boost — AI Is Turning Expert Intuition into Infrastructure — If AI can upgrade from search to recognition, the first step of the positive feedback loop might be breakable.
This article is based on a complete conversation with Claude Opus 4.6. During the conversation, Claude's judgements were repeatedly corrected by previously published texts. All four reversals have conversation records available for verification.
This Series
Attenuation: How Visibility Is Allocated — Three Volumes
- Volume I Approved, Then Unseen — How a Patent Made Silencing Leave No Evidence
- Volume II The Least AI-Like Article Was Actually Written by AI — A Real-Time Experiment in Cognitive Blind Spots (this article)
- Volume III One Like, One Point; One Flame War, Thirty Points — After the Forums Disappeared, Complexity Lost Its Place