anti-ai-tell

name	anti-ai-tell
description	Make written prose not read as AI-generated, even when it is. Three-tier discipline. Tier 0 prevents AI-ese at generation time (voice constraints, banned vocabulary, rhythm rules). Tier 1 is a mechanical lint pass (lint.py) catching measured tells - banned words, em-dashes, negative parallelism, uniform rhythm, copulative avoidance, vague attribution. Tier 2 is the judgment pass that actually matters - visible reasoning, anti-sycophancy, concreteness, committed stance, intellectual friction. Use BEFORE shipping any external-facing or human-read prose - investor notes, public READMEs and issue bodies, marketing copy, blog posts, docs, customer email. Triggers - write, draft, edit, rewrite, copy, email, post, blog, README, announcement, "sound human", "doesn't sound like AI", "humanize", ship external prose.
triggers	doesn't sound like ai|sound human|humanize|anti-ai-tell|ai tell|reads like ai|written by ai|external copy|marketing copy|blog post|readme|announcement|ship this prose|edit this draft
effort	medium
version	2026.06.11-v2

anti-ai-tell — prose that doesn't broadcast its origin

The doctrine, in three findings

1. Surface tells are necessary but no longer sufficient. Readers and tutorials already strip em-dashes, colons, and "It's not X; it's Y." What survives the scrub is the absence of a mind at work. Source: Eve Fairbanks, The Atlantic, "The Biggest Tell That Something Was Written by AI" (2026-05-29), https://www.theatlantic.com/technology/2026/05/how-to-tell-ai-writing/687345/ — "writing that can't really be argued with, because it has no underlying deliberative reasoning process, no train of thought." And the editor's view: "every element is equally off: The tone is bland; individual word choices are baffling; the structure lacks sense." Jason Koebler (404 Media) puts the same gestalt in one line: "The tell for AI isn't rhythm, wording, or fact errors. It's that problems with all these elements exist equally & at once."

2. The vocabulary tell is real and measured. Corpus studies put hard numbers on AI-ese: delves at 28× expected frequency, underscores 13.8×, showcasing 10.7× (Kobak et al., Science Advances 2025, arXiv 2406.07016); meticulous 34.7×, intricate 11.2×, commendable 9.8× in LLM-assisted peer reviews (Liang et al., ICML 2024, arXiv 2403.07183). Adjectives are the most stable single signal. The lists drift by model era (delve → showcasing → emphasizing), so data/vocabulary.json is era-tagged and versioned.

3. Against robust detectors, only distributional change works. Word-swapping defeats humans and frequency-based detectors; perplexity-ratio detectors (Binoculars, arXiv 2401.12070: >90% detection at 0.01% FPR) key on text being too probable and too smooth. The counters that measurably work are lexical diversity + clause reordering (DIPPER, arXiv 2303.13408: DetectGPT 70.3% → 4.6%) and variance injection — surprising word choices, mixed sentence lengths, non-templated structure. Conveniently, those are the same things that make prose read human to people.

Operator directive (2026-06-06): portfolio prose must not read as AI-written, even when it is.

---

Scope — composes with ELITE voice, does NOT replace it

• Internal candor (ELITE voice): Linear tickets, strategy memos, session reasoning, anything operator-only. Em-dashes and antithesis are fine there.
• External / human-read (this skill): investor + partner + recruiting copy, public GitHub (READMEs, issue + PR bodies), marketing, blog, docs, customer email, outward-facing living-doc sections.
• When in doubt: if a stranger could read it and judge the author, it's external. Apply the full ruleset.

---

Tier 0 — generation-time constraints (prevent, don't repair)

Embed these in the system prompt / instruction block BEFORE drafting. Cheaper than editing AI-ese out afterward. The block lives in adapters/SYSTEM_PROMPT.md for non-Claude clients.

1. Banned vocabulary: never use the hard_ban list from data/vocabulary.json (delve, tapestry, showcase, underscore, meticulous, realm, game-changer...). Treat strong_flag words as rationed: one per document, with cause.
2. Rhythm: mix sentences under 10 words with sentences over 20. Never three consecutive sentences of similar length. Vary paragraph length; a one-sentence paragraph is allowed to land a point.
3. No em-dashes. No "not X, but Y" constructions. Use commas, periods, parentheses; make the positive assertion directly.
4. Plain copulas: write "is", "are", "has" instead of "serves as", "stands as", "boasts", "represents".
5. Concreteness: every abstract claim gets a number, a name, or an example within two sentences. If none exists, flag [NEEDS EXAMPLE] rather than padding.
6. No throat-clearing, no closing bows: delete "It's worth noting", "In today's world", "In conclusion". Start where the point starts; stop where it ends.
7. Take a position: where the material supports a judgment, state it and defend it. No symmetric "on the one hand" hedging unless genuinely undecided — and then say why undecided.
8. Register match: these rules tune for directness; formal/academic targets keep necessary passives and looser sentence-length rules. Genre overrides the defaults, deliberately.
9. Voice grounding: when a house voice exists, paste 2-3 paragraphs of real prior writing as a few-shot anchor.
10. Self-review tail: end the instruction with "Before finishing, re-scan the output for em-dashes, banned words, and three similar-length sentences in a row; fix what you find." A cheap built-in second pass.

Tier 1 — mechanical lint (run lint.py, fix everything it flags)

python3 lint.py DRAFT.md            # docs/READMEs (markdown allowed)
python3 lint.py DRAFT.txt --prose   # emails/posts (markdown flagged too)
python3 lint.py DRAFT.md --json     # machine-readable, CI-gateable

ID	Tell	Measured basis
S1	Em-dash overuse	Fairbanks names it the #1 stripped tell
S2	Negative parallelism ("not just X, but Y")	Russell et al. ACL 2025 via WP:AISIGNS
S3	Uniform sentence rhythm (CV < 0.45; runs of 4)	Muñoz-Ortiz 2024: humans have scattered length distributions
S4	Evenly paced paragraphs (CV < 0.30)	Fairbanks: "uniform in length, with evenly paced paragraphs"
S5	Throat-clearing openers	practitioner consensus
S6	Connective filler (moreover, furthermore...)	Kobak common-set excess
S7	Closing-bow wrap-ups ("In conclusion...")	WP:AISIGNS formulaic conclusions
S8	Banned vocabulary, era-aware	Kobak r=28/13.8/10.7; Liang 34.7×/11.2×/9.8×
S9	Copulative avoidance ("serves as", "boasts")	Geng & Trotta: >10% drop in plain is/are
S10	Rule-of-three compulsion	WP:AISIGNS triad habit
S11	Vague attribution ("studies show")	WP:AISIGNS; name the source instead
S12	Markdown artifacts in prose (opt-in)	practitioner consensus for plain-prose targets
S13	Density-watch clusters (crucial/comprehensive/notably...)	Kobak best-10 marker set, Δ=0.134

A clean lint pass means the cheap tells are gone. It does NOT mean the prose reads human. Proceed to Tier 2.

Tier 2 — the judgment pass (where the actual tell lives)

Five questions, asked of the draft as a whole. These cannot be linted; they are the Fairbanks layer.

1. Is the reasoning visible? Can a reader reconstruct why each claim follows? If a paragraph could be deleted without breaking an argument chain, it's decoration. Show the work: "X, because Y, despite Z."
2. Does it ever push back? Human thought has friction: a doubt voiced, a counterargument taken seriously, a premise questioned, an earlier sentence corrected. Sycophantic agreement with the reader's assumed view is the single deepest tell (models affirm users ~49% more than humans do — Science, via Fairbanks). One honest "this might be wrong because..." outweighs ten scrubbed em-dashes.
3. Is anything concrete? Names, numbers, dates, places, one specific anecdote. AI pads with abstractions because abstractions can't be wrong. Editors rate "inject one concrete example per abstract paragraph" the highest-value single edit.
4. Does it commit? A verdict, a ranking, a recommendation with a reason. Symmetric hedging ("there are pros and cons") is machine neutrality.
5. Would the author recognize themselves? Idiosyncrasy is the human signature: a pet phrase, an odd-but-apt metaphor (only fresh ones — never stock), a structural choice a style guide would frown at. Language-flattening is itself a tell (Turk, The Atlantic, 2025-04-29).

Anti-perfection note: do NOT inject typos as fake authenticity. The honest version of the same signal is unpolished structure — a thought that visibly develops rather than arrives pre-assembled.

Workflow

1. Draft with Tier 0 constraints in the prompt (or rewrite an existing draft against them).
2. Run lint.py; fix every finding; re-run until clean.
3. Do the Tier-2 hand-pass; revise.
4. For high-stakes prose: read it aloud once. The ear catches uniform meter the eye forgives.

Maintenance

The vocabulary lists rot as models change. Refresh data/vocabulary.json from Wikipedia's WP:AISIGNS era tables and new Kobak-style corpus studies; bump the updated field. Era notes in the JSON show the drift history.

Run refresh_vocab.py --snapshot new_terms.txt --dry-run --date 2026-09-01 to preview a vocabulary refresh deterministically before it lands.

Cross-media (visual) discipline

Prose is one medium. AI-generated UI, images, logos, and documents converge on the same look for the same reason prose does: the model emits the centre of the training distribution. The fix is the same shape — surface deny-list (mechanical) plus a Tier-2 judgment pass — and one new idea: a positive style fingerprint that injects a chosen point of view, because you can't lint your way to taste.

Full catalogue of 79 sourced visual tells: research/visual-ai-tells.md.

Tool	What it does
visual_lint.py FILE	Tier-1 visual linter for CSS/Tailwind, markdown, .docx, SVG. Flags indigo-500, Inter, untouched shadcn, gradient-text, glassmorphism, emoji headers, markdown bleed, raster-in-SVG. Reads fingerprint.json as an allow-list.
fingerprint.py	The style fingerprint: derive <file> extracts slots from existing work, gaps lists what to ask, seed prints the starting point, validate checks it. Seeding flow is derive-then-confirm.
variance_floor.py SET	Population-level sameness detector. Scores a batch for variance collapse; never judges a single artifact (per-sample detection is provably unreliable — Sadasivan 2303.11156).

The style fingerprint (fingerprint.json) is the cure for genericness: one persistent off-mode anchor (accent, font, radius, voice, distinctiveness dial) that drives both generation and the linter's allow-list. When a design task has unset slots, ask the user — don't fill them with the centroid default.

Tier-2 visual judgment (no linter sees this): is there a hierarchy, a point of view, a rough edge? Perfect uniformity is itself the tell.

Plugin hooks (automatic enforcement)

The Claude plugin ships three hooks so the discipline fires without being invoked:

• PostToolUse (Write|Edit) → lints the touched file, warns on tells (non-blocking).
• SessionStart → injects the active fingerprint so generation carries the POV.
• UserPromptSubmit → on a design task with an incomplete fingerprint, nudges the agent to ask the user for a style anchor (the active-ask mechanism).

Slash commands: /aat-lint (lint a file/dir), /aat-fingerprint (derive-then-confirm).

Files