The previous three posts in this series established a sequence. The publishing market is not suffering a volume problem. It is suffering from a cognition verification problem. The mechanisms that once connected readers to trustworthy work have collapsed, and the reorganization filling that vacuum rewards incumbents and personalities rather than writers. The cultural loss is specific and ongoing. What the market cannot do on its own is rebuild the accountability infrastructure that made trust signals possible.
That leaves a question the market is not yet asking clearly: What does cognition verification actually look like at the level of the manuscript?
This post examines how to answer it.
Let’s start with what could be easily mistaken for verification.
AI detection tools currently operate at the prose level. They examine sentence structure, token probability distributions, stylistic consistency, and a range of linguistic features that differ statistically between human- and machine-generated text. Some of them are reasonably accurate under controlled conditions. None of them is reliable enough for consequential decisions, as the false positive problem documented in the Epoch Times piece illustrates. An author whose legitimate work triggers a detection tool faces reputational damage with no meaningful recourse because the tool measures the wrong thing.
The wrong thing is style. Style is a surface property. It is also increasingly a learnable property for language models. The gap between human and AI prose at the stylistic level is narrowing and will continue to narrow. Any verification system built on stylistic detection is built on ground that is actively eroding beneath it.
The right thing to measure is structure. Not because structure is harder to fake, though it is, but because structure is where authorial cognition actually lives. A novel’s architecture is the record of a mind making hundreds of interdependent decisions across tens of thousands of words, each one constrained by what came before and constraining what comes after. That record is not stylistic. It is cognitive. And it is legible, if you know how to read it.
Post 2 described the structural failure modes of AI-generated fiction in clinical terms. This post takes the inverse position: what structural properties constitute positive evidence of sustained authorial cognition?
There are several, and they cluster around a single underlying phenomenon. Call it narrative intentionality: the degree to which every significant element of a manuscript is in a purposeful relationship with every other significant element, governed by a coherent authorial intention that persists across the entire work.
Narrative intentionality is not the same as technical competence. A manuscript can be competently written at every local level and still lack it. It is also not the same as thematic consistency, which can be achieved by repeating motifs without genuine developmental logic. Narrative intentionality is closer to what experienced editors mean when they say a novel knows what it is. Every scene is doing more than one thing. Character decisions carry consequences that ripple forward. The ending was always the ending, even if the reader couldn’t see it coming. The work has a governing intelligence behind it that was present from the first page to the last.
This is detectable. Not with certainty, and not by algorithm alone, but with the kind of structured analytical framework that a trained reader applies when evaluating a manuscript at depth. The framework examines specific properties: the load-bearing relationship between scenes, the developmental consistency of character psychology across dramatic pressure points, the degree to which thematic content is embodied in action and consequence rather than stated in dialogue or narration, the structural integrity of the second half relative to promises made in the first, and the coherence of the ending as a destination the manuscript was always moving toward rather than a terminus it arrived at by exhaustion.
None of these properties is impossible for a language model to approximate in short form. All of them become progressively harder to approximate as length increases because approximating them requires holding an intention across the entire work, which is precisely what current architecture cannot do.
This is where FictionMark enters the argument, and where I need to be direct about what it is and what it is not.
FictionMark is a structural manuscript analysis service. It was built to answer the question serious writers and serious publishers have always needed answered: Does this manuscript work as fiction at the architectural level? That question predates AI. It is the question developmental editors ask, that MFA workshop instructors ask, that the best literary agents ask when they are deciding whether a manuscript has the bones to survive revision and publication.
What AI has done is make that question newly urgent for a reason it was never urgent before. Structural analysis now serves a dual function. It serves its original purpose: identifying where a manuscript’s architecture succeeds and where it fails, giving the author actionable guidance for revision. And it serves a new purpose: generating evidence of sustained authorial cognition that stylistic detection cannot generate.
A manuscript that passes structural analysis at the level FictionMark examines has, by definition, demonstrated narrative intentionality across its full length. That demonstration is not proof of human authorship in the legal sense. The treatment of AI-assisted work under copyright law remains a genuinely unsettled question, and I am not a lawyer. What structural analysis provides is different and, in some ways, more useful than legal proof. It provides a credible, substantiated trust signal. It tells a reader, a publisher, or an editor that a mind was governing this work from beginning to end.
That is the instrument the market has not yet built into its accountability infrastructure. Not a detection tool. A cognition analysis tool.
The distinction between detection and cognition analysis matters enough to be re-stated precisely.
Detection is binary and backward-looking. It asks, “Was this generated by AI?” The answer is increasingly unreliable, increasingly gameable, and increasingly beside the point as human-AI collaboration becomes normalized in writing as it has in every other creative field. A writer who uses AI for research, for brainstorming, for generating rough material that they then substantially reshape, is not doing something categorically different from a writer who uses a research assistant, a writing group, or an editor who rewrites sentences. The question of origin is genuinely complex, and the industry’s attempt to treat it as simple is producing a false-positive problem and the reputational hazard that comes with it.
Cognition analysis is continuous and forward-looking. Is there evidence of sustained authorial cognition in this work? The answer to that question does not depend on the tools the author used. It depends on whether a governing intelligence was present across the entire manuscript, making intentional decisions in service of a coherent whole. That question is answerable, it is answerable with structured analysis rather than statistical pattern-matching, and it does not produce false positives against legitimate human authors whose prose happens to be clean and consistent.
It also cannot be gamed by running a manuscript through additional AI passes. Structural coherence is not a stylistic property that can be added after the fact. It is either present in the architecture, or it is not, and if it is not, no amount of prose-level refinement will install it.
A word on the limits of what I am claiming, because intellectual honesty requires it.
FictionMark is not infallible. Structural analysis at depth is a judgment, not a measurement, and judgment carries error. A manuscript with genuine architectural coherence might be analyzed by a framework that misses the logic governing it. A manuscript with significant structural problems might be produced by a deeply intentional human author working through their first novel. The tool is not a verdict. It is a substantiated assessment, and the market currently lacks such assessments entirely.
The alternative to an analysis of imperfect cognition is not perfect detection. It is the current situation: no mechanism for verification at all, a market reorganizing around whoever accumulated trust capital before the flood, and a cultural loss accruing silently in the mid-list where serious literary risk-taking has always lived.
An imperfect instrument calibrated to the right question is more useful than a precise instrument calibrated to the wrong one.
Post 5 moves from the manuscript to the ecosystem. If cognition analysis can rebuild trust signals at the level of the individual work, what happens when the reader-author relationship scales beyond discrete texts into the persistent cognitive environments described in Post 3? That is where the governance argument lives, and it is the question the industry is least prepared to ask.
E.S. Martell is a cognitive psychologist, AI governance researcher, and science fiction author. He publishes the After Biology Substack and runs Second Initiative Press. FictionMark.com is his AI-powered manuscript analysis service.