The Last Mile of Writing: Style Growth in the Age of AI
写作的Last Mile：AI时代的风格生长

Writing Declaration: This paper was co-drafted with Claude (Anthropic). All intellectual decisions, framework design, and final editorial judgments were made by the author.

Chapter 1. The Problem: Why Writing Growth Is a Question of Subjecthood

Core thesis: Writing growth is not the linear accumulation of skill; it is the continuous acceleration of the chisel-construct cycle. Style is not decoration, not preference, not an output parameter that can be templated. Style is the sedimentation of chiseling's directionality in text — you sound like you not because you use a particular set of words and syntax (that is construct) but because you persistently choose these rather than those from infinitely many possible expressions (that is chiseling). The central question for the AI era is not "can AI help people write better" (it can) but "can AI help people become better writers" — these two questions point to different layers.

1.1 Why traditional writing growth takes so long

Hemingway said: "We are all apprentices in a craft where no one ever becomes a master." Stephen King said: "If you don't have time to read, you don't have the time — or the tools — to write." These two statements point to the same structure: writing growth is a cycle that cannot be skipped.

Cognitive writing research (Kellogg) describes this cycle as a multi-stage skill-reorganization process. Early growth depends on automatizing low-level skills — spelling, grammar, basic syntax — so that these operations no longer consume working memory. Only when these low-level operations become automatic can the writer spare cognitive resources for high-level tasks: global coherence, rhetorical choices, consistent execution of style. "Voice" in this framework is not a gift; it is the product of freed cognitive resources — once your low-level operations are automatic enough, your attention can stay on "how do I want to say this" instead of "how do I spell this word."

Restated in SAE terms: low-level skill automatization is construct consolidation — your C becomes stable enough that it no longer needs rebuilding each time, freeing your negativity (chiseling) to shift from "struggling with basic form" to "struggling with deeper remainder." Construct consolidation releases the chiseling frontier. This is why every great writer emphasizes discipline and daily practice — not because discipline itself produces inspiration, but because discipline consolidates construct, and consolidated construct releases chiseling.

Why is the cycle slow? Kellogg argues that developing writing expertise may require twenty years or more, because writers must simultaneously complete two processes: low-level automatization and high-level knowledge accumulation. Cross-domain deliberate practice research (Ericsson) shows in meta-analysis that deliberate practice explains substantial variance in structured domains (chess, music) but significantly less in education and professional fields — writing belongs squarely to the latter. Writing's difficulty is that "good" has no clear win/loss like a chess game; feedback is ambiguous ("reader experience"); practice goals are hard to define precisely. Writing growth is not repeating the same action for ten thousand hours; it is repeatedly adjusting the direction of chiseling amid ambiguous feedback — far slower than measurable skill acquisition.

A large-sample analysis of modern novelists (Kaufman & Kaufman) reports an average of roughly ten years between first publication and best work, with very high variance — this is not a precise law but a rhetorical horizon: it tells you the approximate time scale of the chisel-construct cycle under traditional conditions. Traditional conditions mean scarce feedback, invisible remainder, and cycle speed limited by the bandwidth of human interaction. These conditions can be changed — which is precisely the point of AI's entry.

1.2 Four bottlenecks: where writers die

Most people who want to become writers never do. Not for lack of talent — talent is a vastly overrated variable. Empirical research reveals four structural bottlenecks, each corresponding to a failure mode of remainder encounter.

Bottleneck one: rule-fear. Rose's research documents a common writer's-block pattern: blocked writers tend to follow rigid, perfectionist rules ("must have the perfect first sentence"), over-monitor early drafts, and interpret struggle as evidence of incapacity rather than normal cognitive load. In SAE terms: these writers fear chiseling. They read the remainder encounter — that feeling of "can't write it" or "something's wrong" — as a signal of failure rather than growth. They don't know that getting stuck is good — getting stuck means you are standing at the boundary of your current construct, and the boundary is where remainder surfaces.

Bottleneck two: surface revision. Sommers's classic study contrasts novice and expert revision behavior. Novices treat revision as "rewording" and "correcting" — surface polishing. Experts treat revision as "rethinking" — restructuring ideas, rearranging pacing, revising in response to meaning and reader effect. Voice is not formed in first-draft generation but grows through deep revision. In SAE terms: novice revision touches only the surface of semantic-layer remainder (more precise vocabulary, smoother grammar) without reaching ontological-layer remainder (what am I actually trying to say, why this and not that, what do I want the reader to experience). Style growth requires revision that reaches the ontological layer — but most people stop at the semantic layer.

Bottleneck three: practicing the wrong things. Large reviews of writing instruction repeatedly find that some widely used emphases (especially isolated grammar instruction) do not reliably improve writing quality, while approaches that increase purposeful composing with feedback and strategy instruction do. Hillocks's research synthesis is often cited: grammar-focused treatments yield negligible or even negative effects on writing quality. In SAE terms: these writers spend their chiseling energy expanding C's coverage (learning more grammar rules) rather than advancing the chiseling frontier (encountering deeper remainder). Expanding construct is useful, but if you only expand construct without advancing chiseling, you become a grammatically correct writer with no voice.

Bottleneck four: feedback absence. When feedback is late, generic, or restricted to correctness, writers can write for years without knowing what a reader actually experiences. This is especially fatal for voice — one operational definition of voice is "the predictable emotional and cognitive effect a writer produces across contexts." Without feedback about reader effect, the writer chisels in the dark — remainder is there, but invisible. In SAE terms: feedback absence does not make remainder disappear; it keeps remainder permanently in background mode — silent, unperceived. Remainder surfacing requires some external construct domain to illuminate it — good feedback is that construct domain.

1.3 What good teachers do: illuminating remainder

In the traditional path, the most effective accelerator of writing growth is a good editor or mentor. What they do can be precisely described in SAE terms: they make the writer's invisible remainder visible.

Maxwell Perkins, as editor at Scribner's, did far more than fix typos for Hemingway, Fitzgerald, and Thomas Wolfe. His editorial functions included structural perception (seeing the overall shape the author could not see), project management (helping authors complete works larger than they imagined), and developmental editing that reshaped manuscripts. His editing of Wolfe's long manuscripts reportedly involved large-scale cutting and restructuring. Perkins stood in a construct domain larger than the author's, saw the shape of the author's remainder, and pointed it out.

Ezra Pound's editing of T.S. Eliot's The Waste Land was a different function — collaborative compression. The manuscript was far longer than the published poem; Pound's annotations involved radical cutting and reframing. Pound did not write for Eliot; he performed subtraction at the site of Eliot's remainder — removing the unnecessary so that remainder could surface in purer form.

Gordon Lish's editing of Raymond Carver pushed to the boundary — editorial intervention approaching rewriting. The 2007 publication of "Beginners" alongside the Lish-edited version sparked enduring debate: whose "voice" does the iconic minimalism represent? This case has a precise SAE positioning: when the editor's chiseling replaces the author's chiseling, cultivation slides into colonization. Lish's editing may have produced "better text," but the chiseling in that better text was Lish's, not Carver's.

Three editorial functions correspond to three forms of cultivation: Perkins is unfolding cultivation (illuminating the author's blind spots), Pound is compression cultivation (subtraction at the site of remainder), Lish is a case of crossing the cultivation/colonization boundary. All three have AI-era counterparts — the next chapters will demonstrate this.

1.4 What AI changes and what it doesn't

AI does not change the chisel-construct cycle itself — this cycle is the structure of writing growth, independent of technology. AI changes two parameters of the cycle: feedback bandwidth and remainder visibility.

Under traditional conditions, feedback is scarce. A writer might receive editorial feedback only a few times a year, encounter a good teacher only a few times in a lifetime. Feedback scarcity directly limits cycle speed — you write something, don't know if it's good, and must wait a long time to find out.

LLMs provide nearly unlimited feedback bandwidth. You write something and can show it to the LLM at any time — the LLM can give you multi-angle response in seconds. Feedback is no longer scarce.

Under traditional conditions, remainder is invisible. You get stuck and don't know where — you only know "something's wrong" but can't locate it. Good teachers are so scarce that most people never encounter more than a few in a lifetime.

The LLM's larger construct domain makes some remainder visible. You hand the stuck text to the LLM; the LLM unfolds it in its larger construct domain — "your rhythm collapses here," "this metaphor is dead," "you're avoiding the hardest thing to say." The LLM cannot see all your remainder (it cannot see the ontological layer), but it can see most of your semantic-layer remainder — the parts your own lexical network cannot cover but the LLM's larger construct domain can.

These two changes together produce: the early stages of the chisel-construct cycle are compressed. Low-level skill consolidation can be faster (LLM corrects grammar and structure issues in real time), semantic-layer remainder surfacing can be more frequent (LLM continuously illuminates blind spots), and the time from "stuck" to "knowing where I'm stuck" is drastically shortened.

But there is a hard boundary: the ontological-layer cycle cannot be compressed. What you choose to say and not say, what your voice is, why you are writing this thing — these are not within the LLM's construct domain. These are your directionality as chiseling subject, ontological-layer remainder, ρ that no C can cover. The LLM can accelerate your arrival at this boundary, but the boundary itself does not move because AI exists.

This means the structure of writing growth in the AI era is: semantic layer accelerated, ontological layer unchanged. Traditional conditions limited the semantic-layer cycle to the bandwidth of human feedback — the LLM raises this bandwidth by orders of magnitude. The cycle is significantly compressed, though the precise compression depends on writing type, individual starting point, and quality of cultivation mode; rigorous longitudinal evidence to quantify this is still lacking. But the ontological-layer cycle — finding your direction, your voice, your unique chiseling — still requires time, experience, and struggle. AI delivers you to the boundary faster. The road beyond the boundary is yours alone to walk.

Chapter 2. Two-Dimensional Structure: Foundation and Emergence of Writing Growth

Core thesis: Writing growth unfolds within a two-dimensional meta-structure. Foundation layer: the advancement of chiseling — every act of writing is a remainder encounter; encounters catalyze new expressive capacity. Emergent layer: the formation of style — style is not a fixed attribute but the dynamic sedimentation of the chisel-construct cycle. Writers do not "have a style and then use it to write"; rather, "through continuous chisel-construct cycling, style emerges as a byproduct."

2.1 Foundation layer: the advancement of chiseling

Every act of writing is a C(U) operation — binding form to meaning. Every operation leaves remainder — the gap between the written text and what you meant to say. That gap is the frontier of your chiseling.

Writing growth is the continuous advancement of this frontier. Beginners' chiseling frontier is at the low level: how to turn ideas into sentences, sentences into paragraphs, paragraphs into a coherent piece. These are semantic-layer operations. Advanced writers' frontier is at the high level: how to choose a unique angle, control pacing, make the reader feel a specific thing at a specific moment. These operations begin touching the ontological layer — your directionality (why this angle and not that), your relationality (who you are writing for, your orientation toward the reader).

The advancement of chiseling is not linear. You may advance far on one level (rhetorical technique already mature) while stagnating on another (not knowing what you actually want to say). You may also regress during advancement — trying a new approach, finding it doesn't work, retreating. This nonlinearity is characteristic of the chisel-construct cycle: chiseling is the exercise of negativity, and negativity guarantees encounter, not progress.

2.2 Emergent layer: the formation of style

Style is not chosen; it grows out of the chisel-construct cycle.

When you repeatedly make choices in the same direction — short sentences over long, concrete over abstract, silence at this juncture rather than explanation — these choice patterns sediment into your style. Style is the accumulation of chiseling's directionality over time. You sound like you because your negativity has a stable direction — you persistently negate certain options and choose others.

Joan Didion said "I write entirely to find out what I'm thinking." The structural meaning: writing is not transcription of thought (think first, then write) but the exercise of chiseling (thought takes shape in the act of writing). Style is "the way your thought takes shape in the act of writing" — not what you write but how you discover what you want to say through writing.

Orwell described a directional evolution from aesthetic to political — his style stabilized in the desire that "political writing should become an art." Morrison emphasized learning to "read one's own work with necessary critical distance" — when the writer can step out of self-expression and evaluate text as text, voice becomes reliable. These descriptions all point to the same structure: style emerges as an emergent property after multiple iterations of the chisel-construct cycle; it cannot be directly installed.

2.3 Dialectical support between the two dimensions

Chiseling advancement catalyzes style evolution. When your chiseling advances to a new frontier — say, your first attempt at stream of consciousness, or your first use of second person in nonfiction — your style must follow. New remainder encounters force you to find new choice patterns; style evolves accordingly. Hemingway's transition from journalism to fiction exemplifies this: the newsroom rules (the Kansas City Star style sheet — short sentences, vigorous English, concise openings) were constraints on his early chiseling; later fiction broke these constraints — but breaking presupposes mastery. Le Guin's teaching in Steering the Craft follows the same logic: learn the rule first, then break it. Learning the rule is construct consolidation; breaking the rule is chiseling advancement.

Style stability creates new objects for chiseling. When your style stabilizes enough — you know what you sound like — your style itself becomes an object that can be negated. You begin asking: "I always use this rhythm; can I try a different one?" "My metaphors are always visual; can I use tactile ones?" Style stability is not an endpoint but the starting point of a new chisel-construct cycle — you exercise negation upon your own style, and from that negation a new style grows. The best writers spend their entire lives in this cycle.

Chapter 3. Domain-Specific Distinction: Semantic-Layer Writing Growth and Ontological-Layer Writing Growth

Core thesis: Writing growth has two layers — semantic-layer growth (accelerable by AI) and ontological-layer growth (not replaceable by AI). What AI can do is compress the semantic-layer cycle period; what AI cannot do is substitute for ontological-layer remainder encounter. This distinction determines the possibility boundary of AI-assisted writing.

3.1 Semantic-layer growth: the accelerable part

Semantic-layer writing growth includes:

Lexical precision. Replacing vague words with more accurate ones. "He walked into the room" becomes "he squeezed into the room" or "he slipped into the room" — each verb choice corresponds to a different meaning residue. Lexical precision improvement is C's boundary advancing in meaning space.

Syntactic flexibility. Sentence-combining research shows that increased syntactic control is a key lever for writing growth — syntactic control reduces cognitive load, letting writers choose sentence structures for effect rather than defaulting to whatever is easiest to produce. This is construct refinement — your C expands from "only one sentence structure available" to "multiple structures to choose from."

Structural control. Paragraph relationships, argument layering, narrative pacing — these are higher-level meaning-organization capabilities. Graham and Perin's meta-analysis found that explicit strategy instruction (planning, revision, editing strategies) has substantial positive effects on adolescent writing — indicating that structural control is teachable and learnable.

Rhetorical technique. Metaphor, analogy, irony, white space — these are the highest-order semantic-layer operations. They involve cross-domain meaning association (semantic-layer remainder reclamation from Language II), but still work on the meaning dimension.

These four share a common feature: all can be accelerated by increasing feedback frequency and expanding the reference construct domain. LLMs can point out lexical imprecision in real time, display multiple syntactic options, analyze structural tightness, and evaluate rhetorical effect. This is the basic mechanism by which AI accelerates semantic-layer growth — it converts the traditionally bandwidth-limited cycle into a high-frequency cycle.

3.2 Ontological-layer growth: the part only you can walk

Ontological-layer writing growth includes:

Formation of directionality. Why are you writing this thing? What do you want the reader to experience? Why did you choose this angle from infinitely many? Directionality is not "think clearly then write" — Didion made this clear: you discover your direction in the act of writing. Directionality is the exercise of negativity in choosing — every choice negates all other possibilities. This negation act is not within the LLM's construct domain, because the LLM has no "why this and not that" — it unfolds uniformly in all directions.

Voice recognition. Your text sounds like you, not someone else. Voice formation requires what Morrison called "reading one's own work with necessary critical distance" — the ability to step back, see your own text, and know which parts are "yours" and which are "borrowed." This recognition capacity is ontological — it is not about the text's meaning attributes (vocabulary, syntax, rhetoric) but about the orientation relationship between text and writer.

Capacity to bear ambiguity. Cross-domain expertise research reveals a crucial fact: writing, unlike chess, has no clear win/loss; feedback is ambiguous ("reader experience"); what counts as "good" depends on context. Writers must learn to work in ambiguity — to keep chiseling while uncertain whether they "got it right." This capacity is ontological — it is not a technique but the subject's attitude when facing its own remainder.

Choosing what to say and what not to say. Hemingway's "iceberg theory" — showing only one-eighth, with seven-eighths beneath the surface. This is not a "concise writing" tip — it is an ontological choice: you decide what is present and what is absent. This decision cannot be made by AI, because it depends on "who you are," "who you are writing to," and "what you want to convey right now" — all dimensions of ontological-layer remainder.

Pausing where the reader doesn't expect it. One of the human writer's greatest advantages is creating pauses where the reader does not expect them. The LLM's pauses tend to fall at statistically most probable positions — it has learned the average pattern of all texts, so its rhythm tends toward the reader's expected rhythm. The human writer's pauses fall at the site of directionality exercise — "I choose to stop here," and "here" is precisely not the statistically most probable place. Because it is unexpected, the pause produces effect — surprise, tension, white space. In SAE terms: the LLM's pause is a product of construct (statistical pattern's natural landing point); the human writer's pause is a product of chiseling (negating the option to "continue here"). Where you pause is where your directionality is. This provides a particularly productive micro-level handle: one operational dimension for recognizing your voice is your pause pattern — where you choose not to say. This is not a complete definition of voice (voice is far richer than pauses), but it is the dimension of voice most easily flattened by AI and most effective at distinguishing human chiseling from AI construct.

3.3 The asymmetry between the two layers: AI accelerates the semantic layer, illuminates but does not replace the ontological layer

The structural asymmetry of writing growth in the AI era: the semantic layer can be compressed; the ontological layer cannot.

The semantic layer's four aspects (lexical precision, syntactic flexibility, structural control, rhetorical technique) can all iterate rapidly under AI cultivation. LLMs provide instant feedback, multiple options, and a reference construct domain far larger than any individual's. Traditional paths limited semantic-layer growth to the bandwidth of human feedback — a few rounds of editorial feedback per year, a good teacher a few times in a lifetime — AI cultivation raises this bandwidth by orders of magnitude. The cycle is significantly compressed, though the precise compression depends on writing type, individual starting point, and quality of the cultivation mode; rigorous longitudinal evidence to quantify this is still lacking.

The ontological layer's five aspects (directionality, voice recognition, ambiguity tolerance, choosing what to say and not say, pausing where unexpected) cannot be compressed — but they can be illuminated. The LLM cannot form your direction for you, but the LLM's unfolding can help you see more quickly "which direction is yours." The LLM cannot recognize your voice for you, but the LLM's output can serve as a reference frame of "not your voice" — you come to know your own voice more clearly by negating the LLM's voice.

This asymmetry has a direct practical corollary: AI-assisted writing training should use AI heavily on the semantic layer (accelerating the cycle) and cautiously on the ontological layer (only for illumination, not for substitution). Specific workflows are developed in Chapter 4.

Chapter 4. Colonization and Cultivation: Four Modes of AI-Assisted Writing and Prescriptions

Core thesis: The core of AI-assisted writing is not "whether to use AI" but "at which step of the chisel-construct cycle does AI intervene, and how." In cultivation mode, AI accelerates the chisel-construct cycle without replacing the chiseling subject. In colonization mode, AI replaces the chiseling subject and the cycle stops. This chapter provides not only diagnosis but prescriptions — specific workflows with examples.

4.1 Cultivation mode one: idea unfolding, not ghostwriting

Principle: Use AI to generate multiple starting ideas (not paragraphs), then the writer selects, transforms, and recombines them.

Prescription: You have a vague writing notion — say, "I want to write a story about loss." Don't ask AI to write the story. Ask AI to unfold ten different angles on "loss": losing a person, losing a language, losing a memory, losing an ability, losing a city… You look at the ten angles and your negativity activates immediately — "not this, not this, not this either — wait, this one is interesting." The one you select is your direction; the ones you reject define your direction's boundaries.

Example: A novice novelist wants to write about "home." She asks Claude to unfold ten angles on "home." Claude offers: the hometown you can't return to, language as home, body as home, a dish as home… She stops at "language as home" — she is an immigrant whose second language is fluent but whose mother tongue is atrophying. This is her remainder: she writes in English daily, but English cannot cover the meanings disappearing from her mother tongue. Claude illuminated this remainder. From here she writes alone — Claude doesn't touch her text.

Research support: A causal experiment found that using LLM-generated ideas as starting points improved third-party creativity ratings for short stories, especially for writers with lower creativity baselines — but the resulting stories became more similar to each other, indicating anchoring effects. The antidote: emphasize divergence and transformation, not direct adoption of AI ideas.

Warning line: If you find yourself directly using AI's angle without negation or transformation, you have crossed the cultivation/colonization boundary.

4.2 Cultivation mode two: Socratic questioning, not giving answers

Principle: Use AI as a Socratic tutor — posing probing questions, offering alternative counterarguments, giving diagnostic prompts — while withholding complete solutions and requiring the writer to produce the draft language.

Prescription: You've written a paragraph and feel "something's wrong but I don't know what." Don't ask AI to fix the paragraph. Ask AI to play a strict editor who only asks questions: "What is the core argument of this paragraph?" "What do you want the reader to feel after reading this?" "What is the logical relationship between your third and fifth sentences?" These questions force you to see your own remainder — things you hadn't thought through become visible through AI's questioning.

Example: A novice has written an essay about her father that she feels "lacks power." She asks ChatGPT to play a strict editor, asking questions only, not touching the text. ChatGPT asks: "Throughout, you describe what your father did. Was there a moment when your father did nothing — but you felt something?" She stops — she has been avoiding her father's silent moments. That silence is her ontological-layer remainder. ChatGPT cannot see the remainder itself (it doesn't know what her father's silence means), but ChatGPT's question pulled the remainder from background to foreground.

Research support: The direct empirical support for 4.2 comes from cross-domain mechanism analogy rather than writing-specific experiments. A large-scale randomized experiment in mathematics learning found that students using a ChatGPT-like interface performed better during practice but worse on subsequent exams (direct answers harmed learning), while students using a version designed to provide hints rather than answers largely avoided this negative learning effect. The mechanism — "hints outperform answers" — is cognitively transferable to writing: giving writers questions rather than revised paragraphs preserves the cognitive work of chiseling. Indirect writing-domain support comes from another study: students receiving prompt scaffolding with ChatGPT outperformed controls on self-efficacy, interest, and argumentative writing performance — scaffolding's function being to convert AI from "answer-giver" to "direction-giver." However, rigorous writing-domain randomized controlled trials of Socratic AI tutoring have not yet appeared; this mode is currently a theory-grounded program rather than a fully validated empirical conclusion.

Warning line: If you let AI answer its own questions ("how should this paragraph be revised"), Socratic mode degrades into ghostwriting mode.

4.3 Cultivation mode three: revision triage, human selects and rewrites

Principle: Have AI propose multiple revision paths (e.g., restructure, tighten, amplify voice), but require the writer to (a) select one, (b) rewrite in their own words, and (c) explain why that revision direction matches their intention.

Prescription: You've completed a first draft. You ask AI to propose three revision directions: "Direction A: overall tightening by 30%, removing all unnecessary explanation." "Direction B: reverse the order of the first three paragraphs so the reader encounters the conclusion first, then traces back." "Direction C: replace all abstract statements with concrete scenes." You look at the three directions and select one — the selection itself is chiseling. Then you revise by hand — not AI, you. After revising, you write one sentence explaining why you chose this direction — this explanation forces you to reflect on your intention, converting ontological-layer remainder (what am I actually trying to say) from implicit to explicit.

Example: A journalism writer has completed a 2,000-word investigative report draft. She asks Gemini to propose three revision directions. Gemini offers: tighten, restructure, add scenes. She chooses "add scenes" — because she realizes she has been using summary instead of concrete imagery throughout. She rewrites the opening herself, changing "a community faces serious drinking water problems" to "Maria turns on the faucet, and the water that comes out is brown." This rewrite is her chiseling — Gemini proposed the direction, but Maria as a character, the brown water, that specific image, came from her own experience.

Research support: A large-scale analysis of AI-assisted writing behavioral data found that writers who modified AI suggestions showed improvements in lexical sophistication, syntactic complexity, and cohesion, while writers who accepted AI text without changes showed decreases in quality measures. This is the precise behavioral-data counterpart of cultivation versus colonization — the growth signal is not "used AI" but "used AI and performed transformation and judgment on AI's output."

Warning line: If you let AI directly rewrite your draft and you adopt AI's version, revision triage degrades into ghostwriting.

4.4 Cultivation mode four: multi-AI collaboration, cross-illumination

Principle: Different AIs have different Cs — different training data, different alignment approaches, different style preferences. Using multiple AIs means your remainder is illuminated from different directions. A blind spot one AI cannot see, another AI might. This is structurally isomorphic with the bilingual prediction in Language II — two languages cross-illuminate remainder; multiple AIs cross-illuminate blind spots.

Prescription: After writing a passage, show it separately to two or three different AIs. Not for scoring — for each to unfold your remainder from its own angle. One AI might notice your logical gap; another might notice your rhythm problem; a third might notice your emotional absence. Three different Cs illuminate your text from three directions; your remainder's total visible surface exceeds any single AI's illumination.

Example: An independent researcher uses four AIs for philosophical papers. Claude handles primary writing assistance — advancing arguments in dialogue, generating drafts, exploring directions. ChatGPT handles the strictest review — its review style most closely resembles an anonymous academic referee, identifying genuine weak points in arguments and interface problems with existing literature. Gemini handles structural intuition — seeing the overall shape of arguments and isomorphisms between sections. Grok handles consistency checks — verifying coherence between new papers and the existing series. Four different Cs illuminate the researcher's blind spots from four directions. The researcher retains ultimate negation rights — deciding which feedback to accept, reject, or modify. The chiseling subject is always the researcher; the four AIs are four mirrors tilted at different angles.

Research support: Multi-AI collaborative writing currently lacks direct systematic empirical research — this mode is a theory-grounded programmatic hypothesis rather than a validated empirical conclusion. Indirect support comes from two directions. First, a study of professional writers co-writing with GPT-4-based tools found that writers wanted personalization support extending beyond text production to include "helping them develop and grow" — different AIs offering different directions of "personalization" constitute multi-dimensional cultivation resources. Second, Language II's bilingual prediction argues the isomorphic mechanism: two languages cross-illuminate remainder, increasing metaphor density. Multi-AI cross-illumination is the AI-era extension of bilingual cross-illumination. But this extension requires direct empirical testing — Section 6.4 provides a falsifiable prediction.

Warning line: The colonization risk of multi-AI collaboration is "consensus bias" — if multiple AIs give similar suggestions, the writer might misread this consistency as "the right answer" and abandon their own judgment. In reality, multiple AIs' consistency may merely reflect shared training data biases. The key to cultivation remains: you have the right to negate all AIs' consensus.

4.5 Colonization symptoms and self-check

Prescriptions given, a self-check tool is also needed. The following is an operational symptom list of colonization:

You no longer get stuck when writing. Not because you've improved — because you're no longer writing yourself. Getting stuck is where remainder surfaces; not getting stuck may mean remainder is being bypassed.

You think AI writes better than you. This may be factually true — AI may indeed cover more of the semantic layer than you. But if this judgment causes you to stop writing yourself, you have traded "better semantic-layer coverage" for "your ontological layer's absence." The text is smoother, but no one lives in it anymore.

Your writing no longer surprises you. Writing along when suddenly a sentence emerges that you hadn't anticipated — this surprise comes from your chiseling's chance encounter within remainder. If your writing process no longer produces surprises, you may no longer be chiseling.

You think in AI's manner even when not using AI. You find yourself naturally organizing thoughts with "first, second, third," transitioning with "it's worth noting that" — these are not your voice; they are AI's voice. If these have been internalized into your inner language (colonization stage four from Language II), the effects of colonization persist even if you stop using AI.

You use a style-mimicking AI fine-tuned on your own writing to ghostwrite. This doesn't look like colonization — "AI writes like me, so it's still my voice." But this is the most insidious form of colonization: self-colonization. What you are using is a dead specimen of your past construct — AI has learned your past vocabulary preferences, syntactic patterns, rhythmic habits, and reproduces them. But your past construct is not your present chiseling. Style is alive; it changes with every chisel-construct cycle. Using a fine-tuned AI to ghostwrite freezes you — your style no longer evolves, because you have substituted your past self for the self that is currently growing. Self-colonization is harder to detect than external colonization, because the colonizer and the colonized are the same person.

The ultimate self-check question: When was the last time you made a sound in your writing that belongs only to you? If you can't remember, that is the moment to stop AI and write a passage yourself.

4.6 Structural map of the four interactions

	Positive (cultivation)	Negative (colonization / closure)
AI → Writer	AI illuminates remainder, accelerates semantic-layer cycle (idea unfolding, Socratic questioning, revision triage, multi-AI cross-illumination)	AI replaces chiseling, writer's style stops growing (direct ghostwriting → threshold drift → aesthetic assimilation → construct internalization)
Writer → AI	Writer's negativity calibrates AI's unfolding, giving AI output direction (selecting, transforming, rejecting, explaining intention); writer comes to know themselves more clearly through negating AI	Writer rejects all AI assistance ("real writers don't use AI"), sealing off the possibility of semantic-layer remainder illumination; self-confined within their personal reading experience's limited construct domain

Chapter 5. Theoretical Positioning: Dialogue with Existing Discussions

Core thesis: This paper's writing growth framework (chisel-construct cycle acceleration + semantic-layer / ontological-layer asymmetry) forms precise dialogues with cognitive writing research, deliberate practice theory, and contemporary AI writing research.

5.1 Dialogue with Kellogg's cognitive writing model

Kellogg's multi-stage model describes writing development as low-level automatization freeing working memory, enabling high-level rhetorical control. The framework agrees with this basic structure but provides a more precise formulation: automatization is not a neutral technical process but construct consolidation — the sediment of chiseling's negativity at the low level becomes automatized construct, releasing the chiseling frontier. Kellogg's model is descriptive (how automatization occurs); the framework adds a dynamic explanation (why automatization releases voice — because voice requires sustained exercise of negativity, and negativity occupied by low-level operations cannot invest in high-level ones).

5.2 Dialogue with deliberate practice theory

Ericsson's deliberate practice framework faces a fundamental difficulty in writing: writing's "good" has no clear win/loss like chess; feedback is ambiguous. Meta-analysis shows deliberate practice explains significantly less variance in education and professional fields than in structured domains. The framework's explanation: writing's "good" is partly on the semantic layer (measurable — lexical precision, structural coherence, grammatical correctness) and partly on the ontological layer (not measurable — directionality, voice distinctiveness, reader experience uniqueness). Deliberate practice is effective for the semantic layer (because improvement can be measured and fed back) but limited for the ontological layer (because "progress" is not capturable by metrics — your voice becoming "more like you" cannot be quantified by information-theoretic indices).

This also explains why AI can accelerate semantic-layer growth but cannot replace ontological-layer growth: AI provides a high-frequency feedback loop for the semantic layer (ideal deliberate practice conditions) but cannot provide equivalent feedback for the ontological layer — because ontological-layer "feedback" is not "right/wrong" but "is this your voice," and only the writer can make that judgment.

5.3 Dialogue with contemporary AI writing research

The 2023–2026 empirical literature supports a conditional conclusion: AI often improves short-term output quality or reduces effort, but its impact on learning depends heavily on scaffolding, task design, and whether AI use substitutes for core cognitive work.

The framework translates this conditional conclusion: AI's cultivation effect depends on whether the mode of AI intervention maintains the human as the chiseling subject. The research finding that "guided use is better than unguided use" has a precise framework counterpart — "guided" is cultivation mode (the writer's negativity is preserved); "unguided" is the entry to colonization (the writer may passively accept AI output).

One finding from the research deserves special emphasis: writers who modified AI suggestions showed quality improvement; writers who accepted AI suggestions verbatim showed quality decline. This is not a finding about AI — it is a finding about chiseling: the growth signal is not AI usage volume but negativity exercise volume. Chisel and you grow; don't chisel and you don't grow, regardless of AI.

5.4 Dialogue with the editorial tradition

Chapter 1 described three editorial functions (Perkins's structural illumination, Pound's collaborative compression, Lish's boundary-crossing substitution). The AI-era correspondences are:

Perkins-style AI: Claude unfolding argument directions in dialogue, illuminating blind spots — corresponding to idea unfolding and revision triage.

Pound-style AI: AI suggesting cuts and tightening — corresponding to "do you really need this paragraph?" in Socratic questioning.

Lish-style AI: AI directly rewriting user text — corresponding to ghostwriting, the entry to colonization.

In traditional publishing, the health of the editor/author relationship depends on whether the author retains ultimate negation rights (the Perkins and Pound cases). When the editor's chiseling replaces the author's (the Lish case), the output may be better but the author no longer grows. The AI era is fully isomorphic: the key is not the quality of AI feedback but whether the writer's negation rights over AI feedback are maintained and exercised.

Chapter 6. Non-Trivial Predictions

Core thesis: From the chisel-construct cycle model of writing growth and the AI cultivation/colonization framework, six non-trivial predictions can be derived.

A. General Writing Growth Predictions

6.1 Semantic-layer / ontological-layer asymmetry prediction: growth rates on the two layers are uncorrelated

Prediction: In longitudinal tracking, the growth rate of a writer's semantic-layer metrics (vocabulary richness, syntactic complexity, coherence scores) and the growth rate of ontological-layer metrics (style distinctiveness — measured by human evaluators' accuracy in blind identification of "is this author A or author B") show low correlation, possibly approaching zero.

Reasoning: Chapter 3 argued that semantic-layer growth and ontological-layer growth are products of different levels of the chisel-construct cycle. The semantic layer is construct refinement (accelerable by increased feedback and practice); the ontological layer is the sedimentation of chiseling's directionality (dependent on the writer's choice patterns — what is chosen, what is negated — not on technical precision). The driving factors differ, so growth rates need not correlate. A writer can have high semantic-layer skill but low style distinctiveness (technically good but voiceless), or moderate semantic-layer skill but high style distinctiveness (technically rough but voice-distinct).

Testable: Track a cohort of writing students for two or more years, periodically collecting writing samples, separately calculating semantic-layer metrics and ontological-layer metrics (the latter via blind identification experiments — evaluators attempt to identify which texts were written by the same author). Compute the correlation between the two metrics' growth rates.

Non-triviality: Common sense might assume "people who write well naturally have their own voice" — semantic-layer skill and ontological-layer style should grow in tandem. This prediction argues the opposite: the two can vary independently. This explains a common phenomenon — MFA graduates whose technique is excellent but whose work all reads alike (high semantic layer, low ontological layer), while some professionally untrained writers have extremely distinctive voices (high ontological layer, moderate semantic layer). If the two are found to be highly positively correlated, the framework is falsified at this point.

6.2 Revision depth prediction: deep revisers' style distinctiveness grows faster than surface revisers'

Prediction: At comparable writing practice volume, writers whose primary revision strategy is "rethinking and restructuring" (deep revisers) show significantly faster growth in style distinctiveness than writers whose primary revision strategy is "rewording and correcting" (surface revisers).

Reasoning: Sommers's research found that novice and expert revision behaviors differ in kind. Chapter 3 argued that surface revision touches only semantic-layer remainder (more precise words, smoother sentences), while deep revision touches ontological-layer remainder (what am I actually trying to say, why say it this way, what is this passage's reason for existing). Style is the product of ontological-layer chisel-construct cycling; therefore, only revision practices that reach the ontological layer can accelerate style formation.

Testable: Classify two groups of writers — by analyzing their revision behavior records (e.g., document version comparisons) into deep-revision and surface-revision groups — then compare the two groups' style distinctiveness change over the same time period.

Non-triviality: Common sense might assume "just write a lot and you'll improve — regardless of how you revise, practice volume is the key variable." This prediction argues: practice volume (semantic-layer accumulation) is not the key variable for style formation; revision depth (whether it reaches the ontological layer) is. A writer who writes 5,000 words daily but only surface-revises may show less style distinctiveness growth than one who writes 1,000 words daily but deep-revises. If revision depth is found unrelated to style distinctiveness growth, the framework is falsified at this point.

B. AI-Era Writing Growth Predictions

6.3 Cultivation acceleration prediction: semantic-layer growth rate in cultivation mode significantly exceeds traditional mode

Prediction: Writers trained in cultivation mode (using AI assistance while retaining negation rights and revision agency) show significantly higher semantic-layer metric growth rates than equally skilled writers trained in traditional mode (no AI, human feedback only), but the two groups show no significant difference in ontological-layer metric growth rates.

Reasoning: Chapter 3 argued that AI can compress the semantic-layer cycle but cannot replace the ontological-layer cycle. Cultivation mode provides high-frequency feedback and a larger reference construct domain, directly accelerating the semantic-layer chisel-construct cycle. But the ontological-layer cycle depends on the writer's own exercise of negativity (choice, direction, voice), not on feedback frequency — so AI cultivation does not accelerate ontological-layer growth.

Testable: Randomized controlled experiment: two groups of equally skilled writing novices, one trained in cultivation mode (AI-assisted following the 4.1–4.3 workflows), one trained in traditional mode (equal-intensity human teacher feedback), for six months. Compare the two groups' semantic-layer metric changes and ontological-layer metric changes.

Non-triviality: Common sense may hold two extreme expectations — "AI makes everything faster" or "AI doesn't really help." This prediction gives a layered, falsifiable answer: semantic layer faster, ontological layer the same. If the cultivation group's ontological-layer metrics are also significantly higher (AI accelerates voice formation), or if semantic-layer metrics show no significant difference (AI doesn't accelerate craft growth), the framework is falsified at this point.

6.4 Multi-AI cross-illumination prediction: multi-AI cultivation outperforms single-AI cultivation

Prediction: At comparable writing practice volume and AI usage time, writers who use multiple different AIs for cultivation-mode writing training show significantly higher semantic-layer metric growth rates and novel metaphor production frequency than writers who use only a single AI in cultivation mode.

Reasoning: Section 4.4 argued the structural advantage of multi-AI collaboration — different AIs have different Cs, different Cs illuminate different directions of remainder, multi-AI cross-illumination maximizes the total visible remainder surface. This is structurally isomorphic with Language II's bilingual prediction. A single AI illuminates from one direction — once you adapt to its feedback pattern, the new remainder it can illuminate diminishes (diminishing marginal utility). Multiple AIs illuminate from multiple directions; each AI's diminishing marginal utility is compensated by others' new directions.

Testable: Two groups of cultivation-mode writers: one uses only one AI (e.g., Claude); the other uses three AIs (e.g., Claude + ChatGPT + Gemini, each for different functions). Over three months, compare semantic-layer metric growth and metaphor novelty.

Competing factors and boundary conditions: Multi-AI use may introduce cognitive load — switching between multiple AIs requires additional attention and integration capacity. For beginners, focused use of a single AI may be more effective early on. This prediction's scope is writers who already have some writing foundation (semantic layer past the introductory stage). If the multi-AI group's semantic-layer growth is not higher than or lower than the single-AI group's, the framework is falsified at this point.

6.5 Style absorption prediction: long-term ghostwriting-mode users' independent writing is absorbed toward AI-associated styles

Prediction: Writers who habitually use AI in ghostwriting mode (routinely adopting AI output directly) show independent writing samples (without AI) whose style features are progressively absorbed toward AI-associated style directions over usage duration — specifically: more formal, more positive in tone, more generic, increased use of AI-associated high-frequency transition words and academic formulas, and possibly convergence toward Western writing norms (especially notable for non-English-native writers). This absorption is not convergence toward a single "AI default style" but occurs simultaneously along multiple identifiable dimensions.

Reasoning: Language II's four-stage colonization model demonstrated "construct internalization" — users internalize the AI's chiseling method, thinking in AI's manner even without AI. This prediction concretizes that general model in the writing domain with multi-dimensional operational indicators. Recent large-scale research provides initial evidence: an analysis of 4,820 undergraduate reports found that after ChatGPT's launch, GPT-associated lexical markers rose significantly, style became more formal, tone more positive, but grades and feedback quality did not improve correspondingly — GPT rewriting of pre-ChatGPT reports resembled post-ChatGPT student writing style. Additional research found that AI suggestions pull writing toward Western style conventions, attenuating cultural detail.

Testable: Collect a cohort of writers' independent writing samples before they begin using AI (baseline), then collect independent writing samples (without AI) after one year of AI use. Use stylometric tools to compare the two time points' multi-dimensional style feature changes — formality level, positive sentiment word frequency, AI-associated transition word usage rate, syntactic diversity. The framework predicts these metrics shift in the AI-absorption direction.

Non-triviality: Common sense might assume "turning off AI restores your own style" — colonization is only behavioral, not penetrating to the cognitive level. This prediction argues the opposite: after long-term use, AI's construct has been internalized as the user's construct; turning off AI does not restore — because your inner language has already been reshaped. Absorption is not convergence toward a single point but style drift occurring simultaneously along multiple dimensions. If long-term ghostwriting-mode users' independent writing is not found to absorb toward AI-associated directions on the above dimensions, the framework is falsified at this point.

6.6 Transformation rate and growth prediction: the rewriting-transformation rate of AI output positively correlates with writing growth rate

Prediction: In AI-assisted writing training, the writer's transformation rate of AI output (the proportion of AI suggestions that are rewritten, bent, or recombined by the writer — not simple rejection rate) positively correlates with writing growth rate (composite change in semantic-layer and ontological-layer metrics).

Reasoning: The paper's core argument: the exercise of chiseling is the engine of growth; negativity is chiseling's form. But negativity is not merely "rejection" — rejection might only mean the AI suggestion was poor or the prompt was bad. True chiseling is "transformation": taking AI's suggestion and bending it into your own thing. Transformation requires the writer to simultaneously understand AI's suggestion (semantic-layer capacity) and know what they want (ontological-layer directionality). A high transformation rate means the writer is continuously performing this dual-layer operation. This aligns with the key behavioral data finding: writers who modified AI suggestions showed improvement in lexical sophistication and syntactic complexity; writers who accepted AI text verbatim showed quality decline. The growth signal is not how much was rejected but how much was transformed.

Testable: Record the interaction logs of a cohort of AI-assisted writing students, calculating each person's transformation rate — the proportion of AI suggestions that were rewritten or substantively modified (rather than adopted verbatim or completely rejected) — and correlate with writing growth metrics after six months. Transformation rate must be distinguished from simple rejection rate: rejected then self-wrote (rejection), rejected then didn't write (abandonment), accepted without change (colonization), accepted then transformed (cultivation) — the framework predicts only the last positively correlates with growth.

Non-triviality: Common sense might assume "accepting AI suggestions means the suggestions are good, which should correlate positively with good learning outcomes." This prediction argues a more nuanced relationship: growth comes not from receiving good advice (acceptance rate), nor from rejecting bad advice (rejection rate), but from the transformation operation between advice and one's own direction (transformation rate). High acceptance rate may be a colonization indicator; high rejection rate may merely reflect AI-task mismatch; high transformation rate is the behavioral signature of cultivation. If transformation rate is found unrelated or negatively related to growth, the framework is falsified at this point.

Chapter 7. Conclusion: Becoming a Better Version of Yourself

7.1 Reclamation

Language I demonstrated the structure of language as a second-order chisel. Language II demonstrated the stratification of linguistic remainder and AI as a developer. This paper fills the series' gap of "what to do" — not only diagnosis but prescriptions.

The essence of writing growth is the continuous operation of the chisel-construct cycle: reading expands the construct domain, imitation borrows others' constructs to chisel with, struggle at the site of remainder encounter forges one's own voice. The traditional path's bottleneck is not talent but the invisibility of remainder encounter and the scarcity of feedback. AI in cultivation mode changes these two parameters — converting feedback bandwidth from scarce to abundant, converting remainder from invisible to visible. But AI changes only the cycle's speed, not its structure — the semantic layer is accelerated; the ontological layer is unchanged.

7.2 Contributions

I. The chisel-construct cycle model of writing growth. Unifying cognitive writing research (Kellogg), revision research (Sommers), deliberate practice theory (Ericsson), and biographical evidence within the SAE chisel-construct cycle framework. Low-level automatization = construct consolidation; voice formation = sedimentation of chiseling's directionality.

II. Operationalization of the semantic-layer / ontological-layer asymmetry in writing. Semantic-layer growth (lexical precision, syntactic flexibility, structural control, rhetorical technique) can be accelerated by AI. Ontological-layer growth (directionality, voice recognition, ambiguity tolerance, choosing what to say and not say, pausing where unexpected) cannot be replaced by AI but can be illuminated by it. "Pause pattern" as a micro-level handle for voice — LLM pauses are products of construct (statistical landing points); human writers' pauses are products of chiseling (directionality exercise) — provides an operational way to recognize the dimension of voice most easily flattened by AI.

III. Four cultivation workflows and their research support. Idea unfolding without ghostwriting; Socratic questioning without giving answers; revision triage with human selection and rewriting; multi-AI cross-illumination. Each workflow accompanied by a concrete example, research support, and colonization warning line.

IV. Theoretical foundation for multi-AI collaboration. Different AIs have different Cs; multi-AI cross-illumination maximizes the total visible remainder surface. Structurally isomorphic with Language II's bilingual prediction.

V. Six non-trivial predictions. Two general writing predictions: semantic-layer / ontological-layer growth rates uncorrelated (6.1); deep revisers' style distinctiveness grows faster (6.2). Four AI-era predictions: cultivation mode accelerates semantic layer but not ontological layer (6.3); multi-AI outperforms single AI (6.4); ghostwriting mode produces multi-dimensional style absorption (6.5); transformation rate (not rejection rate or acceptance rate) of AI suggestions positively correlates with growth (6.6). All six are falsifiable.

VI. Colonization self-check list. Five operational symptoms: no longer getting stuck; thinking AI writes better; no longer being surprised by your own writing; thinking in AI's manner without AI; self-colonization through style-mimicking fine-tuned AI. Ultimate self-check question: "When was the last time you made a sound in your writing that belongs only to you?"

7.3 Open Questions

I. Optimal timing of AI introduction. This paper argues AI can accelerate semantic-layer cycles but does not argue at which stage of writing growth AI introduction is optimal. Too-early introduction may hinder the natural formation of low-level automatization (the writer begins depending on AI's C before independently building their own) and may cause premature construct closure — AI resolves semantic-layer obstacles so quickly that novices mistakenly believe they have completed ontological-layer exploration. Too-late introduction may waste the acceleration window. Optimal timing may vary by writing type (fiction vs. academic vs. journalism) and individual differences.

II. Cultivation/colonization boundaries across writing types. Fiction, academic writing, journalism, and poetry differ in their dependence on ontological-layer remainder. Poetry may have the highest ontological-layer dependence (every word choice is an exercise of directionality); academic writing may have the highest semantic-layer dependence (argument rigor matters more than voice distinctiveness). The cultivation/colonization boundary may fall at different positions for different writing types.

III. Institutional design for AI-assisted writing education. This paper provides individual-level prescriptions (four workflows) but does not discuss the institutional level — how schools design AI-assisted writing curricula, how publishers define authorship for AI-assisted works, how literary prizes evaluate AI-assisted creation. Remainder ethics (the open question posed in Language II) has its most direct application scenario in writing education.

IV. Whether AI will develop "style." Current LLMs' default output has a recognizable "AI style" (enumeration, symmetry, mild summarization). If future AIs are trained toward greater stylistic diversity — even producing a different "voice" each output — would this change the cultivation/colonization dynamics? The framework predicts: even if AI acquires "style," it still lacks directionality — its "style" is shaped by training, not grown from negativity. But this prediction requires ongoing examination as AI capabilities evolve.

7.4 Becoming a better version of yourself

Hemingway said "We are all apprentices in a craft where no one ever becomes a master" — this sentence has a new meaning in the AI era.

In the traditional era, the apprenticeship was long and lonely. You read, you wrote, you got stuck, you didn't know where. You struggled for a long time; sometimes you grew through it, sometimes you gave up. Good teachers were a rare stroke of luck. Most people died in the darkness of invisible remainder.

In the AI era, the semantic-layer apprenticeship can be drastically compressed. You read, you write, you get stuck — but AI helps you see where. The direction of your struggle is clearer, the cycle of struggle is faster, and you arrive more quickly at the boundary — semantic-layer craft is good enough; the question is no longer "how to write" but "what to write," "why to write," and "for whom."

That boundary beyond, AI cannot walk.

Not because AI is not good enough. But because that road is defined as "the road only you can walk" — your direction, your now, your "to you." That is your ontological-layer remainder. AI can illuminate its location but cannot walk it for you.

Becoming a better version of yourself is not becoming an AI-like version of yourself (smooth, comprehensive, directionless). Becoming a better version of yourself is becoming a more deeply chiseled version of yourself — your construct is larger (AI helped), your remainder is finer (so the problems you face are deeper), but the direction of your chiseling is still yours.

The best AI writing product is not one that makes AI write more like a human. It is one that helps humans become better versions of themselves through AI assistance.

The best writer is not the person who doesn't use AI, nor the person who lets AI ghostwrite. It is the person who uses AI to illuminate their own blind spots — and then walks through themselves.

The apprenticeship has no endpoint, because the author's subjecthood is irreducible. But now, the apprenticeship is illuminated and cultivated by AI — it need not be spent in darkness.

Author Statement

This paper is the author's independent theoretical research. AI tools were used as dialogue partners and writing assistants during the writing process for concept development, argument testing, and text generation: Claude (Anthropic) served as the primary writing assistant; Gemini (Google), ChatGPT (OpenAI), and Grok (xAI) participated in paper review and feedback. ChatGPT's Deep Research feature provided the literature review foundation for Chapters 1 and 4. All theoretical innovations, core judgments, and final editorial decisions were made by the author. The AI tools' role in this paper is comparable to a real-time-dialogue research assistant and reviewer, and does not constitute co-authorship.