What follows is a summary generated by ClaudeAI. The summary represents a single, extended, session, discussing what makes AI written text sound poor to human readers. What resulted was (in my opinion anyways) a very interesting exchange that ended up leading to a rather significant improvement in writing quality. Now, the subject may seem a bit odd to you because it's specific to the characters in my book and so you won't have all the context (until you read the book of course) - BUT... the concept, I think is fascinating.
Lastly, I have left the summary as is. It is a little on the long side, but, as I said, I found the result interesting.
We started by asking whether AI could detect its own tells. The honest answer was no — and neither can any of the dedicated detection tools (Originality.ai, GPTZero, Copyleaks). But the question led somewhere more interesting than detection. It led to asking why AI prose reads as AI prose, and whether understanding the mechanism could improve the writing itself.
That turned out to be a much more useful question.
What AI prose actually does wrong
The problem with AI-generated prose isn't that it's bad. It's that it's undifferentiated. It gravitates toward a small rotation of safe, mid-register words. It produces even sentence rhythms. It reaches for abstraction when it should reach for specificity. The prose is smooth — but the smoothness is unintentional.
This matters because smoothness isn't always wrong. Transitions should be smooth. Connective tissue should be smooth. The moments where prose is doing logistical work — moving characters between rooms, establishing a new scene — those are moments where the writing should disappear. But emotional peaks, character interiority, sensory detail — those are the places where generic register actively costs you. If the prose is smooth there, it's doing real damage.
The diagnostic question we landed on was: can I point to a choice in this paragraph that only THIS narrator would make? If not, the paragraph has drifted.
We built a four-tier watch list. Tier 1 was the high-frequency defaults — "the quality of," "specific/specifically," "particular/particularly," "someone who," "simply." Not banned words. Diagnostic ones. When three cluster in a paragraph, the paragraph needs attention. Tier 2 was connective tissue — the transition phrases AI uses to create explanatory flow: "which meant that," "because" in interior monologue, excessive "as if," "she realized that." These produce a smooth causal narration that explains a character's reactions to the reader instead of letting the reader experience them. Tier 3 was emotional vocabulary — the twenty words AI rotates through for all emotional work: "a flicker of," "the weight of," "shifted," "settled over her," "carefully," "quietly," "a sense of." All accurate. None specific to any particular character. Tier 4 was structural tells — paragraph-level patterns like even rhythm, the thesis statement sentence that tells you what a scene means, the emotional summary paragraph.
Running this against the full 61,000-word manuscript produced findings I didn't expect.
The manuscript had two populations. Chapters 1–8 had been heavily revised; my voice was dominant. Chapters 9–15 were less revised, more AI-generated. We treated the first set as the training set and the second as the test set.
The headline finding: "the specific [X] of" — an abstraction wrapper that names a category instead of rendering the thing — appeared zero times in chapters 1–8 and twenty-four times in chapters 12–15. Every instance was a sentence that told you about an experience instead of putting you inside it. "The specific weight of a man who had been waiting for sixteen years" versus showing his hands go still on the counter.
"Settled/settling" appeared 53 times doing five different jobs. The word had become inaudible through repetition. "Something" appeared 375 times — many legitimate, many lazy. Zero named emotions anywhere in the manuscript ("she felt anxious/calm/uneasy") — the voice had been holding its taxonomic discipline. Zero "she realized that" or "she couldn't help but" — the human revision had killed these completely.
Why the watch list isn't enough
The watch list tells you what's wrong. The strategy bank we built alongside it tells you what to do instead — structural moves keyed to six recurring scene types, with techniques from literary reference points translated into actions the AI could execute.
Here's where the first thing I got wrong turned out to be instructive.
I included Ishiguro's displacement technique in the strategy bank. The technique is real and it works: Ishiguro's narrator in Never Let Me Go over-describes something trivial nearby while the reader understands the real emotional content, and the disproportion IS the feeling. So I named it as the primary move for emotional peaks, and the AI immediately defaulted to it — three times in a row, in a single scene. The protagonist was looking at the character's hands, the counter, a worn patch on the kitchen floor. Displacing away from the conversation into benign objects.
I caught it immediately. The AI had a literary model and went straight to it. The problem was it had produced Kathy H. from Never Let Me Go when it should have been producing my protagonist. Kathy displaces because she's avoiding. My protagonist looks at physical details because she's reading. Her taxonomic mind doesn't redirect away from the difficult thing — it redirects through the nearest available evidence toward understanding. Same technique. Wrong character. The literary model had overridden the voice.
This was the first real ceiling. The watch list worked for diagnosis. The strategy bank worked for individual fixes. Neither solved the deeper question: if we can identify these tells so easily in review, why can't we use the same knowledge to prevent them during drafting?
The honest answer was: partially, but there's a ceiling. The watch list functions as negative constraints, and the AI will avoid the flagged patterns. But those patterns aren't random tics. They're where the AI's probability distribution lands when it needs to do a specific job. Block "the economy of someone who had been cutting apples for a long time," and the AI will produce "the precision of someone who had been cutting apples for a long time." The furniture moves. The room doesn't change.
The real fix isn't knowing what not to write. It's knowing what to write instead. And that requires something the AI doesn't have.
The empathy gap — and what it actually is
When I described this problem to the AI, we had a conversation about what was missing. A human writer imagining a specific character cutting an apple draws on accumulated sensory memory — what practised hands look like, how a knife moves when someone isn't watching the blade, the specific way a person occupies a kitchen they've stood in for decades. The AI doesn't have this. But it does have extraordinary pattern-matching across prose that describes exactly these kinds of moments.
The gap isn't really empathy. It's selection.
The AI can generate twenty plausible details for what a character's hands look like. The problem is it can't reliably pick the one that's alive. A human writer doesn't generate twenty and pick — they see one, the right one, because their embodied experience filters before the writing begins. The filtering is invisible because it happens before the words.
This reframing changed what I thought needed to be built. The question stopped being "how do I teach the AI empathy" — you can't, and the question is the wrong shape — and became "how do I narrow the AI's selection space to the details that are already established and working in the chapters where the voice is right?"
The character documents I had built were architectural. They described characters' relationships, narrative function, philosophical positions. What was missing was what I started calling Layer 2: not what characters ARE, but what they LOOK LIKE when they're being what they are. Physical behaviour signatures.
Building Layer 2 from the training set
The method was simple: go through chapters 1–8 and pull every physical detail for each character. Not descriptions. Behaviours. Not what they look like in a general sense — what they do specifically, what they reach for, how their hands move.
For one character, this produced a concentrated physical portrait. The worn patch on the counter where he stands. The hand on the counter edge — the resting hand, always present while the other works. The apple peel in a single unbroken strip. Both hands going still when he's thinking about what not to say. The half-sentence where the body finishes what the words don't.
Every one of these details was already in the manuscript. The problem was they were scattered across eight chapters and not consolidated in a way the AI could access during drafting. When writing a new scene for this character, the AI couldn't reach for "both hands still" — it reached for "the economy of someone who had been cutting apples for a long time" because the abstraction is always available and the specific detail requires knowing which detail is already established. Layer 2 was simply the established details, organised.
We tested it immediately. A scene from chapter 13 — the character's biggest disclosure scene, with the densest concentration of abstraction wrappers — got redrafted with the Layer 2 document loaded.
"The specific weight of a man who had been waiting for sixteen years" became "Both hands were still now. She had never seen both hands still at the same time."
The abstraction wrapper was gone. The physical detail was grounded in established continuity. The improvement was real and immediate. And then the Ishiguro problem happened again.
The AI had the character's physical signatures. It knew what to find. But it still defaulted to displacement — the protagonist looking at his hands, the counter, the worn patch — because the strategy bank was still pointing toward Ishiguro as the primary move for emotional peaks. The character's Layer 2 had told the camera what to look at. It hadn't told the camera how to look.
The observer layer
The human insight that broke this open was: it can't be done in isolation. We improved the character's description, so we describe him better — but there needs to be an influence for the protagonist. She's the one pointing the camera. We need to know where she would point it, and then how the subject would look.
The character's physical signatures (what he looks like) were only half the equation. The other half was the protagonist's relationship to what she's seeing in a specific moment. The same detail — both hands still on the counter — means something different when she's filing successfully than when she's angry. The AI was defaulting to the Ishiguro model because it didn't have a document telling it how this narrator's camera behaves under specific emotional pressures.
So we built the observer layer. Seven distinct camera modes, extracted from how the protagonist actually behaves across the training set:
Filing successfully — systematic sweep, negative-definition stack, controlled rhythm.
Filing failing — obsessive precision on adjacent details, broken rhythm.
Reading someone familiar — catches one deviation from baseline.
Reading someone new — wide sweep, dense prose, failed filing at the end.
Angry or hurt — camera SHARPENS, stays on the person, facts as weapons.
Body before mind — records the aftermath of an involuntary action.
Recognition — revisits established details, re-reads them through the new lens.
Mode 5 was the critical discovery. When this protagonist is angry, she does not displace to benign objects. She gets more precise. The anger strips away the social layer and she reads the person who hurt her with devastating accuracy. "She lived with you for four years. She had a baby with you." Facts as weapons. The taxonomic mind under pressure doesn't displace — it accelerates.
This was the fix for the Ishiguro problem. The observer layer told the AI which mode was active, which determined the camera's behaviour, which determined the emotional register of the prose. The strategy bank still existed, but it was now filtered through the character's specific relationship to observation. Ishiguro's displacement technique was available — but only for Mode 1, cautious filing, where restraint is the right register. Not for the scene in chapter 13, which was Mode 5.
What changed
We redrafted the same scene with both layers loaded — the observer modes telling the camera how to move, the character's Layer 2 telling it what to find.
The angry passage now stayed on the character instead of displacing to the counter. The anger didn't just sharpen the reading of him — it expanded to the full architecture of her childhood. The house, the garden, the name, the apple, the Thomas book. The objects became evidence of what he did, not a retreat from the conversation. Same physical details from the same Layer 2 document. Completely different emotional relationship, because the observer mode was now correct.
The physical-to-abstract ratio in chapters 1–8 was 14:1. In chapters 9–15 it was 6:1. With both layers loaded, the test drafts were approaching the training-set ratio.
Once the architecture was validated, we built it for the complete cast and for the recurring locations. Each character layer followed the same extraction method — not character descriptions, but physical behaviour signatures from chapters where the human voice was already working. Each location layer followed the same principle: not what the place is, but what this narrator sees when the camera arrives, grounded in the specific vocabulary from the training set.
The drift analysis by character produced its own findings. One character's worst drift was structural, not vocabulary: a late chapter had decoded him. "The held quality was care. The restraint was tenderness." Eight chapters of accumulated mystery discharged in three clauses. The training set never does this — the protagonist reads his surface with increasing precision but never explains him to the reader. Another character had the steepest proportional drift: 7% abstraction in training, 36% in test. He nearly vanishes from the narrative in the later chapters, and the few appearances get described abstractly rather than physically. Another character showed zero abstractions in training and five in test — not in the word count, but in the structure. The test set explained her unreadability rather than enacting it.
What the system can't do
The Layer 2 system closed the gap between abstraction and physical rendering. The remaining gap is between competent physical rendering and the alive, unpredictable detail that makes a reader stop and reread a sentence.
The AI can now reach for "both hands still on the counter" instead of "the specific weight of a man who had been waiting." That's a real improvement. But the strange, specific, surprising detail — the one that could only come from this narrator in this moment looking at this person — still requires human selection. The AI generates from a narrowed field. The human writer sees the one detail that's alive and discards the rest.
The Layer 2 system doesn't simulate empathy. It simulates the output of empathy — the specific physical observations that a human writer produces when they imagine what it's like to be inside a character's perception. By consolidating those observations from the chapters where the human voice is dominant, the system gives the AI a reference library of what this character actually looks like that it can pattern-match against in new scenes.
The human still writes the details that aren't in the library yet. The AI writes better drafts from the details that are.
The loop
The conventional framing of AI in fiction is that AI generates and humans edit. What this process suggests is a different model: the human's best writing becomes the training data for the AI's drafting, which produces better raw material for the human's next revision.
The loop is:
Human writes and revises chapters until the voice is right. The physical details, observer modes, and character behaviours from those chapters are extracted and organised. The AI uses those extracted patterns to draft new scenes with narrower, more character-specific selection. The human revises the AI's draft, adding the unpredictable details the AI couldn't find. Those new details enter the Layer 2 documents for the next drafting cycle.
Each cycle tightens the collaboration. The AI's ceiling rises as the reference library grows. The human's revision load decreases as the drafts get closer to voice. Neither replaces the other. The gap narrows but doesn't close — and the irreducible human part, the alive detail, the surprising choice, the thing only this writer would see — remains the thing that makes the fiction worth reading.
That's not a problem with the system. That's the point of it.