Six Pivots to Kill 'AI Taste': Materials, Choices, and Responsibility in Agent-Native Writing

Initial Approach: Treating AI Taste as a Language Problem

In July 2026, the author applied humanizer-zh and stop-slop skills to rewrite WeChat articles. These tools flagged opening padding, filler words, promotional tone, vague attribution, false intimacy, and formulaic sentence patterns, while suggesting direct verbs, rhythm adjustments, and removal of non-informative summaries. This step provided a concrete checklist to pinpoint why text felt "off" — e.g., a sentence that circles without action, repeated transitions, straw-man counters, and mandatory but empty section summaries.

However, aggressive rules (delete all adverbs, passive voice, dashes, or split every long sentence) sometimes harmed accurate expression in Chinese technical writing. A smooth complex sentence chopped into five short ones lost AI taste but broke readability. The author retained the problem-awareness of these rules but abandoned mechanical zeroing: a word, punctuation, or long sentence only signals "look here again," not proof of a problem.

Second Pivot: Structure Still Thinks for the Author

On August 25, 2026, a reverse calibration using real drafts and final versions showed AI taste had moved beyond single sentences. Some subheadings packed object, conflict, and conclusion upfront, answering before the body began. Articles used consecutive "not A but B" frames, each paragraph correcting the reader. Concepts were forced into equal-length definitions; case studies became company name lists. After facts were told, an English slogan was appended as if closure required elevation. Individually each sentence was grammatical, but together they felt extruded from a standard template.

DxC then changed its checking order: first pass examines structure only — what task a section carries, how materials relate, why one segment leads to the next. Fixes mean delete, merge, reorder, or rewrite a whole unit; swapping connectives or synonyms does not count. Second pass handles language: read aloud, check whether abstract words hide actions, whether multiple conditions are crammed into one sentence, whether references drift, whether long/short sentences have become a mechanical beat. Final three paragraphs undergo a deletion test: if removing them loses no facts, constraints, or actions, they were likely just announcing "I'm done."

A trap emerged: requiring every article to show obvious structural change led to breaking already sound structures. Now DxC uses whitelist editing — locate a specific problem, articulate the edit reason, then act; untouched parts stay as-is. A practical test for structural issues: strip the polished connectives and ask why two sections are adjacent. If you cannot answer, the relationship needs repair, not the wording.

Third Pivot: The Writer Cannot Reliably Self-Certify

Two-pass checking solved the inspection object, but the writer still certifies their own pass. This mirrors the Agent execution problem described in the series' second post: semantic judgment is probabilistic; even a clear Skill may be unread, skipped, or summarized with a single "checked."

On August 30, 2026, DxC added independent body proofreading. After the primary writer finishes structure and language checks, a separate proofreader in a clean context re-reads raw materials, Brief, outline, author context, and current body, delivering location-specific judgments. Any body change invalidates the old proofread; the new version must be re-proofed from scratch.

The CLI verifies a different fact layer: whether check items are missing, sequence is complete, and proofread receipts match the current body. Body fingerprint changes invalidate old receipts. This division continues the theme from earlier posts: natural-language interaction preserves open judgment; the sequence, required fields, and confirmations that forms once provided must be rebuilt elsewhere. Semantic quality goes to the model and independent reading; code only guards provable completion conditions.

Code can prove check-item and receipt completeness and body fingerprint match. It cannot prove the article reads naturally, let alone compute an objective "human-ness score." Independent proofreading catches the primary writer's blind spots, but the proofreader's semantic judgment still requires author review.

Fourth Pivot: The Input Lacks Author Evidence

Independent proofreading only solves who checks. If input lacks author evidence, it can only clean existing drafts, not conjure an author. humanizer-zh-adrop provided a key insight. Beyond checking two-part/three-part structures, promotional language, and empty tails, it protects existing author expression: direct judgments, irregular pauses, natural long sentences, complete processes, constraints, failures, and human reviews. If materials lack first-person experience, it refuses to invent a vignette just for "human feel."

Evaluating human-writing later, the team adopted principles of material sufficiency, action priority, information progression, and natural closure. Its scanning script spots some textual shapes but cannot judge factual strength, author evidence, or whole-piece voice. Static checks still only yield clues.

DxC then added a narrow boundary for historical articles: only when at least three different past articles show directionally consistent expressions can they be generalized as the current task's temporary writing DNA (expression traits). Even then, history only calibrates rhythm, structural tendencies, and judgment style — never supplies facts, ready-made positions, or copies old sentences to fake the author's voice.

This boundary blocks two deviations: generic rules molding every author into the same "natural," and history imitation molding the author into their past self. The current article needs the author's choices on this material batch.

More dangerous is fabricated humanity. When raw materials contain no anxiety, dialogue, loss, or cognitive shifts, the model invents "I was panicked then" or "Only later did I realize." These sentences carry emotion and turning points and read more human — but that person is not the author, just a temporary role the model fabricated.

When materials are insufficient, the reliable responses are mundane: keep asking, shorten the article, or narrow the claim. Do not fill factual gaps with rhetoric.

Deepest Layer: Completeness Without Accountability

Materials are thin, yet the article writes a complete experience. The author made no selections, so the text assigns pros, risks, challenges, and future to every object. Paragraphs lack real relations, so format inserts causality, hierarchy, and conflict. Evidence supports only local observations, yet the conclusion is framed as an industry trend.

Such articles are usually thorough. That thoroughness has no provenance.

You cannot see who observed what, which result changed the judgment, which materials were discarded, which areas remain unknown. Every sentence plays safe, but the whole piece has no one willing to stand behind it.

This is the deep AI taste the author now recognizes: after the author vanishes from the text, an over-complete shell remains.

Human flavor is no longer about colloquial density. It first means materials belong to a concrete person. That person has chosen among materials, is willing to explain where judgments came from, and admits where evidence stops.

Such articles need not have stories, emotions, or first-person pronouns. The author remains visible in the selections and boundaries.

Current Workflow: Materials First, Then Structure and Language

Today DxC starts body drafting by checking what materials can support. Are there actual observations, actions, results, constraints? Is there a judgment that genuinely changed? If not, decide whether to gather more material or shrink the article. Never first use synonym rewriting to inflate sparse facts into long copy.

Once a full draft exists, structure relations are addressed first. Does each chapter fulfill its outline duty? Do cases answer the same observation question? Does the conclusion overreach evidence? After structure holds, read the whole text aloud, fixing only already-located issues. Natural long sentences, necessary lists, ordinary metaphors, and the author's stable colloquialisms are not auto-deleted just because they hit statistical markers.

The candidate draft then goes to independent proofreading. The proofreader looks for concrete locations: where the text circles in place, where a template manufactures a correction, where emotions absent from materials are inserted, where strong conclusions appear with no owner. Code verifies proofreading completeness and version correspondence; it does not adjudicate naturalness.

A final portion still requires author confirmation. Which experiences are theirs? Which judgments are they willing to publish? Which sentence can they stand behind? Where should "I don't know" stay? This part has no automation shortcut.

This method comes from DxC and the author's own writing practice. It is not a universal detection standard for all authors, styles, and models. Humans also write boilerplate; models can do high-quality editing when materials are sufficient and boundaries clear. Forty-four check items, version binding, and a few calibration samples only make the process more complete and easier to audit — they do not prove the problem is universally solved.

AI can help spot empty talk, rebuild structure, and make proofreading more thorough. It cannot conjure a scene a person lived through, the price they paid, or the judgments they are willing to own.

True "human flavor" is letting the author reappear in materials, choices, and boundaries.