# The Ultimate Guide to Decontaminating AI Prose **Purpose:** Detect and remove LLM writing artifacts ("AI slop") from any document -- articles, books, reports, emails, marketing copy. **Applies to:** Text produced or rewritten by ChatGPT, Claude, Gemini, Llama, or any LLM. --- ## What Is AI Slop? Simon Willison coined "slop" in 2024 as the AI equivalent of spam: **generated content that nobody asked for, nobody reviewed, and nobody wants.** But even intentionally produced, human-reviewed AI text carries fingerprints. This guide catalogues every known tell and provides a systematic method to remove them. The core insight: **AI writing is detectable through its uniformity, its inflation, its hedging, and its absence of human texture** -- rough edges, strong opinions, specific memories, and imperfect structure. --- ## Part 1: The Word-Level Tells ### 1.1 The Blacklist Words These words occur 3-25x more frequently in AI-generated text than in human writing (Kobak et al., 2024; Liang et al., 2024). Their presence in non-literal contexts is a strong AI signal. | Word | AI frequency spike | Human alternative | |------|-------------------|-------------------| | delve | 25x (most notorious GPT-ism) | explore, examine, dig into, look at | | tapestry | ~10x | mix, web, patchwork, or cut entirely | | vibrant | ~8x | lively, active, specific descriptor | | landscape | ~7x (in non-literal use) | field, space, world, or name the thing | | multifaceted | ~6x | complex, layered, or describe the actual facets | | crucial | ~5x | important, key, or just state why it matters | | meticulous | 7.4x in peer reviews | careful, precise, thorough | | commendable | 9.4x in peer reviews | good, strong, solid | | intricate | 4.5x | detailed, complex, or describe the complexity | | nuanced | ~4x | subtle, specific, or show the nuance | | pivotal | ~4x | important, turning-point, or say why | | noteworthy | ~3x | worth flagging, or just flag it | | invaluable | ~3x | valuable, essential, or say what value it adds | | groundbreaking | ~3x | new, first, original, or describe what it broke | | robust | ~3x | strong, solid, reliable | | comprehensive | 3x in peer reviews | thorough, complete, full | | encompasses | ~3x | includes, covers | | spearhead | ~3x | lead, drive, start | | paradigm | ~3x | model, approach, shift | | underscores | ~3x | shows, reveals, highlights | | foster | ~3x | build, grow, encourage | | leverage | ~3x (as verb) | use, apply, take advantage of | | embark | ~3x | start, begin, set out | | realm | ~3x | area, field, domain | | testament | ~3x | proof, evidence, sign | | beacon | ~3x | example, signal, model | | resonate | ~3x | connect, ring true, land | **The rule:** If you can replace the word with something simpler and lose nothing, the original was AI filler. ### 1.2 Filler Intensifiers Words that pad sentences without adding information. LLMs scatter them everywhere; human writers use them once or twice per thousand words. | Word | What to do | |------|-----------| | remarkably | Replace with a specific number or detail | | notably | Cut, or say "worth flagging:" | | crucially | Cut, or say "the critical factor:" | | importantly | Cut entirely -- if it's important, the reader can tell | | fundamentally | "at root", "structurally", or cut | | incredibly | Replace with actual measurement | | essentially | Almost always deletable | | ultimately | "in the end", "finally", or cut | | interestingly | Cut -- if it's interesting, the reader will notice | | arguably | Take a position or cut | | certainly | Cut -- just assert the thing | | absolutely | Cut | | undeniably | Cut -- if you need to say it's undeniable, it's probably deniable | ### 1.3 Transition Words That Signal AI These transitions are legitimate in moderation. AI uses them 5-10x more than humans: - **Moreover** -- prefer "also", "and", or restructure - **Furthermore** -- prefer "beyond that", "and", or just start the next sentence - **Additionally** -- prefer "also" or cut - **Indeed** -- cut unless genuinely emphatic - **That said** / **That being said** -- prefer "but", "however", or restructure - **In essence** -- cut and state the essence - **In terms of** -- rephrase: "for X" or "regarding X" - **When it comes to** -- cut and just name the thing - **At the end of the day** -- cut - **At its core** -- cut and state the core **Density test:** More than 2 of these per page (250 words) = AI fingerprint. --- ## Part 2: The Phrase-Level Tells ### 2.1 Vapid Openers Sentences that throat-clear before the actual point. | Opener | Fix | |--------|-----| | Here's the thing: | Delete. Start with the thing. | | Here's where it gets interesting: | Delete. If it's interesting, the reader will notice. | | Here's what you need to know: | Delete. Just tell them. | | Let's dive in / Let's dive deeper | Delete. Just dive. | | Let's unpack this | Delete. Just unpack. | | Let's be clear / Let's be honest | Delete. Just be clear. Just be honest. | | It's worth noting that | Delete. If it's worth noting, note it. | | It's important to note that | Delete. Just note it. | | It's important to remember that | Delete. Just state the thing. | | The reality is | Delete. Just state the reality. | | Picture this: | Delete unless genuinely setting a scene. | | Imagine a world where... | Delete. Overused and patronizing. | | Consider for a moment... | Delete. Just present the idea. | ### 2.2 Unearned Profundity Sentences that reach for gravitas they haven't built. | Pattern | What's wrong | Fix | |---------|-------------|-----| | "This isn't just X -- it's Y" | Implies insight, delivers restatement | State Y directly | | "X is more than just Y" | Same problem | Say what X actually is | | "The implications are profound" | Vague grandeur | Name the specific implications | | "This changes everything" | Almost never true | Say what specifically changes | | "This raises important questions" | Stalling | Ask the actual questions | | "The stakes couldn't be higher" | Inflation | State the actual stakes | | "As we stand on the brink of..." | Melodrama | Name what's happening | | "Only time will tell" | The laziest possible ending | Make a prediction or state what you don't know | | "X is a game-changer" | Meaningless hype | Say what changed and how | | "X sits at the intersection of Y and Z" | AI's favorite spatial metaphor | Say how X combines Y and Z | | "Navigate the complexities of..." | Vague action | Name the specific challenges | ### 2.3 The Parallelism Tic The construction **"It's not X, it's Y"** (and its variants). Used once, it's rhetoric. Used ten times in one document, it's a fingerprint. Variants: - "It's not X -- it's Y" - "It's not X; it's Y" - "This isn't about X, it's about Y" - "The point isn't X" - "The question isn't whether X" **Fix:** Keep 1-2 per document max. For the rest, vary: | Instead of | Try | |------------|-----| | It's not X, it's Y | Less X than Y | | It's not X -- it's Y | Y, not X (inversion) | | This isn't about X | X gives way to Y | | The point isn't X | Where once X, now Y | | | Just state Y directly without the antithesis | ### 2.4 Sycophantic and Self-Narrating Phrases Mostly in chat/conversational output, but they leak into documents: - "Great question!" - "That's a really interesting point" - "I'd be happy to help with that" - "Let me walk you through this" - "Here's a comprehensive breakdown" - "I want to be straightforward about..." - "There's a lot to unpack here" **Fix:** Delete all of these. Just do the thing. ### 2.5 The Universal Applicability Claim AI rarely says "this doesn't apply to your situation." Instead: - "Whether you're a beginner or an expert..." - "No matter your industry..." - "For individuals and organizations alike..." **Fix:** Be specific about who this is for. Excluding audiences builds credibility. --- ## Part 3: The Structural Tells ### 3.1 The Listicle Compulsion AI defaults to numbered lists and bullet points even when prose would be more natural. A human writer might use 1-2 lists in an article; AI structures the entire piece as nested lists. **Fix:** Convert at least half of your lists back to prose paragraphs. Lists are for reference material; arguments belong in sentences. ### 3.2 The Five-Paragraph Essay AI strongly defaults to: intro with thesis, point 1, point 2, point 3, conclusion restating intro. Every section has the same number of sub-points, every bullet is roughly the same length, every paragraph has a similar word count. **Fix:** Vary section lengths. Some points deserve three paragraphs; others need one sentence. Asymmetry is a sign of human judgment. ### 3.3 The "Heading + Summary + Bullets" Pattern Almost every AI section follows: ``` ## [Heading] [One paragraph introducing the topic] - [Point 1] - [Point 2] - [Point 3] [Optional transition sentence] ``` **Fix:** Break the template. Start some sections with a question. Start others with a specific example. Let some sections have no substructure at all. ### 3.4 The Conclusion That Restates the Introduction AI almost always ends by paraphrasing its opening paragraph. Human writers end with a new insight, a question, a call to action, or an unexpected turn. **Fix:** Delete conclusions that merely summarize. End with something the reader didn't know at the start. ### 3.5 Excessive Signposting AI over-narrates its own structure: "First, we'll look at... Then, we'll explore... Finally, we'll consider..." This is the writing equivalent of saying "I'm about to tell a joke" before telling the joke. **Fix:** Cut structural narration. If the structure is clear, you don't need to announce it. ### 3.6 Cross-Reference Over-Explanation Re-defining terms every time they appear: "the Foundation -- the universal baseline we described earlier -- provides..." A human author trusts the reader to remember; an LLM hedges. **Fix:** First mention in a section gets a brief reminder. After that, trust the reader. --- ## Part 4: The Formatting Tells ### 4.1 Italic Function Words -- The Most Diagnostic Pattern No human writer italicizes *and*, *is*, *that*, *with*, *for*, *it*, *was*, *but*, or *or*. LLMs do it constantly because they mimic emphasis patterns without understanding which words carry semantic weight. **Function words to never italicize** (remove emphasis on sight): > and, is, that, with, for, it, was, has, but, or, be, do, than, this, are, were, can, all, so, just, very, too, also, even, if, as, how, what, when, the, a, an, of, in, to, on, at, by, up, out, no, not **This is arguably the single most reliable AI detector.** If a document italicizes function words more than once, it was AI-assisted. ### 4.2 Bold Overuse AI **bolds key terms** extensively. Human writers use bold sparingly -- for definitions, warnings, or truly critical emphasis. **Fix:** Remove bold except for: - Coined terms at first use - UI elements or technical identifiers - Genuine warnings or critical notes If more than 5% of your text is bold, strip most of it. ### 4.3 Em Dash Overuse LLMs love em dashes -- they use them constantly -- sometimes three or four per paragraph -- creating a breathless, parenthetical style that real editors flag immediately. **Fix:** Limit to 1-2 per page. Replace with commas, parentheses, periods, or restructured sentences. ### 4.4 Emoji Tells AI (especially ChatGPT) places emojis in predictable patterns: - One emoji per section header - Evenly distributed throughout - Favorite choices: rocket, lightbulb, checkmark, star, fire, pointing finger **Fix:** Remove decorative emojis. If you must use them, cluster them naturally rather than distributing them evenly. ### 4.5 Triadic Lists Groups of three, everywhere: "resilience, creativity, and purpose." Humans write triads too, but LLMs default to them so reliably that density becomes diagnostic. **Fix:** Vary list length. Use pairs, singles, and groups of four or five. If every list has exactly three items, break the pattern. --- ## Part 5: The Voice-Level Tells ### 5.1 Emotional Flatness Despite using charged words (vibrant, passionate, transformative), AI writing feels emotionally flat. There's no genuine anger, humor, sarcasm, confusion, or vulnerability. It's uniformly positive-neutral. **Test:** Read a paragraph aloud. Can you hear a human voice with a mood, an attitude, a personality? Or does it sound like a corporate FAQ? ### 5.2 No Intellectual Courage AI rarely makes bold claims, takes unpopular positions, or expresses genuine disagreement. Everything is balanced, measured, safe. Real writers have opinions. **Test:** Does the text ever say something that might make someone disagree? If not, it's playing it safe. ### 5.3 Uniform Register Human writers shift registers -- formal in one paragraph, colloquial in the next, technical when needed. AI maintains a consistent "educated professional" register throughout. **Test:** Does the writing ever get casual? Blunt? Funny? Technical? If the register never shifts, it's AI. ### 5.4 Avoided Contractions LLMs tend toward formal register: "it is" instead of "it's", "do not" instead of "don't." In narrative prose, this creates a stiff, robotic cadence. **Fix:** Use contractions in narrative and conversational prose. Reserve formal register for technical exposition. ### 5.5 No Negative Space AI fills every possible angle. Human writers know what to leave out. The absence of content is as important as its presence. AI doesn't know what to omit. **Test:** Could you cut 30% of the text without losing the argument? If yes, the filler is probably AI padding. ### 5.6 Lack of Specificity AI uses generic examples and avoids lived details. A human might write "I spilled coffee on my keyboard at 3am debugging this." AI writes "developers often face challenges in their workflow." **Test:** Are there concrete details that couldn't come from training data? Dates, names, personal anecdotes, specific numbers? ### 5.7 Semantic Satiation Through Synonyms AI often restates the same idea using different words in consecutive sentences, creating an illusion of depth through synonym rotation rather than adding new information. **Test:** After each sentence, ask: "Did I learn something new here, or was this a restatement?" If three sentences in a row say the same thing with different words, collapse them into one. --- ## Part 6: Model-Specific Tells ### ChatGPT (GPT-4/4o) - Heavy on "delve", "tapestry", "vibrant", "landscape" - Confident, assertive tone -- rarely hedges - "Certainly!", "Absolutely!" as openers - More emoji usage - Ends with encouraging calls to action - The "[Topic]. Here's why:" headline pattern - Tends toward grandiose, inflated prose ### Claude - More hedging and epistemic humility ("I think", "it seems") - "I'd be happy to help", "That's a great question" - Heavy em-dash usage - "That said," as transition - "Genuinely", "straightforward", "thoughtful" - Volunteers ethical caveats unprompted - More discursive paragraphs, less listicle-heavy - Meta-commentary: "This is a complex topic, but..." ### Gemini - More verbose than either GPT or Claude - Heavy on "It's important to remember that..." - Produces "Key takeaways" sections unprompted - More likely to cite sources (sometimes hallucinated) ### Open-Source Models (Llama, Mistral, etc.) - More repetitive within paragraphs - Weaker transition management - More likely to produce circular arguments - Less polished formatting --- ## Part 7: Detection Methodology ### Quick Scan (5 minutes) Run these five tests on any document: 1. **The Delve Test:** Ctrl+F for: delve, tapestry, vibrant, landscape, multifaceted, crucial, meticulous, foster, leverage, robust. More than 2 hits in a short document = AI. 2. **The Italic Function Word Test:** Search for single-word italics. If *and*, *is*, *that*, *for*, *it*, or *but* are italicized, it's AI. 3. **The Opener Test:** Search for: "Here's the", "Let's dive", "It's worth noting", "Picture this", "Imagine a world". Any hits = AI. 4. **The Transition Density Test:** Count "Moreover", "Furthermore", "Additionally", "Indeed", "That said." More than 2 per page = AI. 5. **The Structure Test:** Are all sections the same length? Do all lists have 3 items? Does the conclusion restate the intro? If yes to all three = AI. ### Full Audit (30-60 minutes) For book-length or high-stakes documents, run systematic searches: **Pass 1 -- Word-level:** Search for every word in sections 1.1, 1.2, and 1.3. Record counts and locations. **Pass 2 -- Phrase-level:** Search for patterns in sections 2.1 through 2.5. Record counts. **Pass 3 -- Formatting:** Search for italic single-word patterns, bold patterns, em-dash density. **Pass 4 -- Structure:** Map section lengths, list item counts, and conclusion types. **Pass 5 -- Voice:** Read 5 random paragraphs aloud. Apply the tests from Part 5. ### Automated Detection Script For each pattern, use regex searches across your document: ```bash # Blacklist words (case-insensitive) grep -icE '\b(delve|tapestry|vibrant|landscape|multifaceted|crucial|meticulous|intricate|pivotal|commendable|robust|comprehensive|encompasses|spearhead|paradigm|underscores|foster|leverage|embark|realm|testament|beacon|resonate)\b' document.md # Filler intensifiers grep -icE '\b(remarkably|notably|crucially|importantly|fundamentally|incredibly|essentially|ultimately|interestingly|arguably|certainly|absolutely|undeniably)\b' document.md # AI transitions grep -icE '\b(Moreover|Furthermore|Additionally|Indeed)\b' document.md # Vapid openers grep -icE "(Here's the thing|Here's where|Here's what|Let's dive|Let's unpack|Let's be clear|Let's be honest|It's worth noting|It's important to note|Picture this|Imagine a world)" document.md # Unearned profundity grep -icE "(This isn't just|more than just|changes everything|implications are profound|stakes couldn't be higher|game.changer|at the intersection of|navigate the complexities)" document.md # Em dash density grep -c '—' document.md # Italic function words (Markdown) grep -cE '\*\b(and|is|that|with|for|it|was|has|but|or|be|do|than|this|are|were|can|all|so|just|very|too|also|even|if|as|the|a|an|of|in|to)\b\*' document.md ``` --- ## Part 8: The Fix Methodology ### Tier 1: Mechanical Fixes (No Judgment Required) These can be fixed by search-and-replace or automated scripts: | Pattern | Fix | |---------|-----| | Italic function words | Remove italics | | "It's worth noting that..." | Delete the phrase | | "Here's the thing:" | Delete | | "Let's dive in" | Delete | | Filler intensifiers with no specific content | Delete or replace with specific detail | | "Certainly", "Absolutely" as openers | Delete | | Decorative emojis | Delete | ### Tier 2: Judgment Fixes These require reading context and choosing the best alternative: | Pattern | Approach | |---------|----------| | "Not X, it's Y" parallelism | Keep 1-2 per document, vary the rest | | Blacklist words | Replace with specific alternatives that fit the context | | Triadic lists | Vary list lengths; break some into prose | | Em dash overuse | Replace most with commas, periods, or parentheses | | Bold overuse | Keep only for coined terms and genuine emphasis | | Avoided contractions | Add contractions in narrative prose | | Cross-reference over-explanation | Trust the reader after first mention | ### Tier 3: Rewrite Required These can't be fixed with word-level edits -- they require rewriting passages: | Pattern | Approach | |---------|----------| | Unearned profundity | Replace grandiose claim with specific evidence | | Emotional flatness | Add genuine voice: an opinion, a joke, an admission | | Lack of specificity | Add concrete examples, numbers, names, dates | | Symmetric structure | Vary section lengths and formats | | Conclusion that restates intro | Write a new ending that advances the argument | | Semantic satiation | Collapse synonym-rotated sentences into one | | Universal applicability claims | Name the specific audience | ### Tier 4: Leave Alone Not every pattern is slop. Keep: - **Triads that list three actual things** (not filler groupings) - **Specific, surprising analogies** ("hit the market like a truck" has personality; "like a tapestry" does not) - **Bold on coined terms** at first use - **One or two "not X, it's Y" constructions** per long document -- it's legitimate rhetoric in moderation - **Formal register** in genuinely technical passages --- ## Part 9: Prevention (Writing Prompts That Reduce Slop) The best cleanup is prevention. When prompting an LLM to write: ### Effective prompt additions: ``` - Write in a direct, conversational tone. Use contractions. - No filler words: cut "remarkably", "notably", "crucially", "importantly." - No throat-clearing openers: don't start sentences with "It's worth noting" or "Here's the thing." - No "not X, it's Y" constructions. - No em dashes. - No triadic lists unless listing exactly three things. - Use specific examples with real numbers, names, and dates. - Take a clear position. Don't hedge with "on one hand / on the other hand." - Vary paragraph and section lengths. Not everything needs three sub-points. - Don't italicize function words (and, is, that, for, it, but, or, with). - Don't use: delve, tapestry, vibrant, landscape, multifaceted, crucial, foster, leverage, robust, comprehensive, embark, realm, beacon, paradigm. - Don't end by restating the introduction. ``` ### The "Write Like a Human" Checklist Before publishing any AI-assisted text, verify: - [ ] Would a specific human being actually write this sentence? - [ ] Does this paragraph add new information, or restate the previous one? - [ ] Is there at least one opinion, joke, or specific detail per page? - [ ] Does the register shift at least once (formal to casual or vice versa)? - [ ] Is the conclusion different from the introduction? - [ ] Could I cut 20% without losing the argument? (If yes, cut it.) - [ ] Are there any words I wouldn't use in conversation? - [ ] Does every list have a reason to be a list instead of prose? --- ## Part 10: Decision Framework When reviewing a flagged pattern, apply this tree: ``` Is this pattern used fewer than 3 times in the full document? YES → Probably fine. Leave it. NO → Continue... Does removing/changing it alter the author's meaning? YES → Leave it. Flag for human review. NO → Continue... Is this a proper noun, coined term, or intentional style choice? YES → Leave it. NO → Continue... Is this a function word with italic/bold emphasis? YES → Remove emphasis. No human writer does this. NO → Continue... Is this a filler word that can be cut without losing meaning? YES → Cut it. NO → Continue... Is this a structural pattern (triadic list, symmetric sections)? YES → Vary it. Break the template. NO → Rewrite the passage with specificity and voice. ``` --- ## Appendix A: French-Language AI Slop French has its own distinct set of AI tells that don't map 1:1 to English. This appendix covers everything specific to detecting and cleaning AI-generated French text. ### A.1 The French Blacklist Words | French | English equivalent | Human alternative | |--------|-------------------|-------------------| | crucial | crucial (THE most notorious French GPT-ism) | important, central, determinant | | essentiel | essential | important, necessaire, utile | | fondamental | fundamental | de base, structurel, premier | | captivant | captivating | interessant, frappant, marquant | | fascinant | fascinating | interessant, etonnant, surprenant | | passionnant | thrilling/exciting | interessant, stimulant | | troublant | troubling | inquietant, preoccupant, deroutant | | transformateur | transformative | nouveau, different, qui change | | revolutionnaire | revolutionary | nouveau, inedit, sans precedent | | indispensable | indispensable | necessaire, utile, important | | vibrant | vibrant (English carryover) | vivant, anime, dynamique | | robuste | robust (English calque) | solide, fiable, resistant | | holistique | holistic (English calque) | global, d'ensemble, complet | | inlassablement | tirelessly (350x overuse) | sans relache, avec perseverance | | entrelacer | intertwine (350x overuse) | meler, croiser, combiner | ### A.2 Overused French Verbs and Verb Phrases The single most overused French AI construction is **"permettre de"** (to allow/enable). AI uses it where a direct verb would be more natural. | AI writes | Human writes | |-----------|-------------| | Cela permet de reduire les couts | Cela reduit les couts | | Ce systeme permet d'ameliorer | Ce systeme ameliore | | Cette approche permet de comprendre | Cette approche eclaire / On comprend mieux avec | Other overused verb constructions: | Verb phrase | Human alternative | |-------------|-------------------| | mettre en oeuvre | faire, realiser, appliquer | | mettre en place | creer, installer, lancer | | naviguer dans | traverser, gerer, comprendre | | ouvrir la voie a | preparer, faciliter | | tisser des liens | creer des contacts, se rapprocher | | jeter les bases de | commencer, fonder, preparer | | franchir un cap | progresser, avancer, depasser | ### A.3 French Vapid Openers and Closers **Openers to delete on sight:** | French opener | Equivalent English tell | |---------------|------------------------| | Dans un monde en constante evolution... | In today's fast-paced world... | | Dans le paysage actuel de... | In the current landscape of... | | A l'ere de... | In the era of... | | A l'heure ou... | At a time when... | | Force est de constater... | One must acknowledge... | | Il est essentiel de noter que... | It is essential to note that... | | Il est important de noter que... | It is important to note that... | | Il est crucial de comprendre... | It is crucial to understand... | | Il convient de souligner... | It should be emphasized... | | Explorons cette problematique... | Let's explore this issue... | | Plongeons dans les details... | Let's dive into the details... | **Fix:** Same as English -- delete the opener and start with the actual point. **Closers to cut or rewrite:** | Closer | Fix | |--------|-----| | En conclusion | Delete -- if it's the last paragraph, the reader knows | | En resume | Delete | | En somme | Delete | | En definitive | Delete | | Pour finir | Delete | | Comme nous l'avons vu | Delete -- don't narrate what the reader just read | ### A.4 French Transition Words AI Overuses | Transition | Frequency in AI vs human | Alternative | |------------|-------------------------|-------------| | En outre | 5-8x overused | aussi, et, par ailleurs (sparingly) | | Par ailleurs | GPT-4's reported favorite | de plus, d'autre part, et | | De plus | 5x overused | aussi, et | | De surcroit | 4x overused | en plus, et | | En effet | 4x overused | (often deletable -- just state the fact) | | Neanmoins | 3x overused | mais, pourtant | | Par consequent | 3x overused | donc, du coup | | Dans cette optique | 3x overused | pour cela, dans ce but | | A cet egard | 3x overused | sur ce point, la-dessus | | Cela etant dit | 3x overused | mais, cependant | | Dorenavant | unnecessarily formal | desormais, maintenant | | En d'autres termes | signals restatement, not progress | (delete and state it once) | **Density test:** More than 3 of these per 250-word page = AI fingerprint. Natural French relies more on sentence flow and implicit logic than on explicit connectors. ### A.5 The "Langue de Bois" Problem The most distinctive French AI tell is **register**. AI writes French in a pompous, administrative, pseudo-academic style the French call "langue de bois" (wooden language) -- the bureaucratic, evasive, empty style of politicians and corporate communications. **Symptoms:** 1. **Excessive impersonal "il" constructions:** "Il est important de...", "Il convient de...", "Il est essentiel de..." -- these avoid taking a position. Human writers say "je pense" or "on sait" or just assert directly. 2. **"Nous" where "on" is natural:** In spoken and casual written French, "on" is far more common than "nous" for first-person plural. AI defaults to formal "nous." If every "we" in your text is "nous," it reads like a government report. 3. **No colloquialisms:** AI French has no personality -- no argot, no regional flavor, no generation-specific language. It reads like a textbook written for no one in particular. 4. **Stacked subordinate clauses:** "Il convient de noter que, dans une certaine mesure, cette decision pourrait potentiellement impacter les parties prenantes concernees" -- a sentence that says nothing while sounding important. Human French is more direct. ### A.6 French Typography Tells These are formatting artifacts specific to French that reveal AI authorship: | Tell | What's wrong | Correct French | |------|-------------|----------------| | Em dash (tiret cadratin) overuse | ChatGPT's single most famous French tell, acknowledged by OpenAI. French uses em dashes far less than English. | Prefer commas, parentheses, or restructured sentences | | "texte" or "texte" | English-style quotation marks | French uses guillemets: << texte >> | | Missing thin spaces | No space before : ; ! ? | French requires a thin non-breaking space before these | | Straight apostrophes (') | English-trained model artifact | Use typographic apostrophes (') | | Title Case Sur Chaque Mot | English capitalization style | French capitalizes only the first word: "Titre de l'article" | | Comma before "et" | Oxford comma (English convention) | French does NOT use a comma before "et" | ### A.7 Anglicisms AI Introduces Research shows 16% of ChatGPT's French errors have English origins. Watch for: | AI writes (anglicism) | Correct French | |-----------------------|----------------| | Faire du sens | Avoir du sens | | Adresser un probleme | Traiter / resoudre un probleme | | Impacter (overused) | Avoir un impact sur, affecter, toucher | | Realiser (meaning "understand") | Se rendre compte, comprendre | | Supporter (meaning "support") | Soutenir, appuyer | | Definitivement (meaning "definitely") | Assurement, certainement, sans aucun doute | | Eventuel (used as "eventual") | Eventuel means "possible" in French; for "eventual" use "final" or "a terme" | ### A.8 French Recycled Metaphors The French equivalents of "like a tapestry": | Metaphor | Why it's a tell | |----------|----------------| | Le paysage de... (the landscape of) | Non-literal "landscape" -- same problem as English | | Au coeur de... (at the heart of) | AI's go-to locator phrase | | La pierre angulaire de... (the cornerstone of) | Generic importance marker | | Un ecosysteme de... (an ecosystem of) | For any group of things | | Un levier puissant (a powerful lever) | For any tool or method | | Le fil conducteur (the common thread) | For any recurring theme | | Un equilibre delicat (a delicate balance) | For any tradeoff | | La face cachee de... (the hidden face of) | For any downside | | A double tranchant (double-edged) | For any tradeoff (variant) | | La pointe de l'iceberg (tip of the iceberg) | For any partial revelation | ### A.9 The "Non seulement X, mais Y" Problem The French equivalent of "It's not X, it's Y" is **"Non seulement X, mais aussi/egalement Y"** (Not only X, but also Y). ChatGPT uses this construction obsessively in French. Same fix as English: keep 1-2 per document, vary the rest. ### A.10 Quantitative Detection Metrics These measurements distinguish AI French from human French: | Metric | AI range | Human range | |--------|----------|-------------| | "Notamment" frequency | 1 per 200 words | 1 per 800 words | | Adverbs ending in "-ment" | 4-6% of words | 1-3% of words | | Average sentence length | 18-22 words (low variance) | 12-25 words (high variance) | | Paragraph length variance | +/-20% | +/-50% or more | | Interrogative sentences | 2-5% | 5-15% | | Type-Token Ratio (vocabulary diversity) | 0.65-0.75 | 0.50-0.65 | The **"notamment" test** is the single most reliable quantitative test for French AI text. Count occurrences and divide by word count. If the ratio exceeds 1:400, the text is likely AI-assisted. ### A.11 French Prompt Additions for Prevention Add these to your prompt when generating French text: ``` - Ecris dans un ton direct et conversationnel. Utilise "on" plutot que "nous." - Pas de mots-remplissage : supprime "crucial", "essentiel", "fondamental", "fascinant", "passionnant." - Pas d'ouvertures creuses : ne commence pas par "Dans un monde en constante evolution", "Il est important de noter", "Force est de constater." - Pas de "non seulement X, mais aussi Y." - Pas de tirets cadratins (em dashes). - Utilise les guillemets francais (<< >>) et les espaces fines avant : ; ! ? - Pas d'anglicismes : "avoir du sens" (pas "faire du sens"), "traiter un probleme" (pas "adresser"), "soutenir" (pas "supporter"). - Varie la longueur des paragraphes et des phrases. Pas de symetrie. - Utilise "permettre de" au maximum 2 fois. Prefere le verbe direct. - Pas de conclusion qui repete l'introduction. - N'utilise pas : paysage (sens figure), pierre angulaire, ecosysteme (sens figure), levier puissant, fil conducteur, equilibre delicat. - Titre : majuscule uniquement au premier mot. ``` ### A.12 French Detection Script ```bash # Blacklist words grep -icE '\b(crucial|essentiel|fondamental|captivant|fascinant|passionnant|troublant|transformateur|revolutionnaire|indispensable|vibrant|robuste|holistique|inlassablement)\b' document.md # Vapid openers grep -icE "(Dans un monde en constante|Dans le paysage|A l'ere de|Force est de constater|Il est essentiel de|Il est important de noter|Il est crucial|Il convient de|Explorons cette|Plongeons dans)" document.md # Overused transitions grep -icE '\b(En outre|Par ailleurs|De plus|De surcroit|Neanmoins|Par consequent|Dorenavant|Cela etant dit|Dans cette optique|A cet egard)\b' document.md # "Permettre de" density grep -ic 'permet\(tre\)\? de' document.md # "Notamment" frequency test echo "Notably count vs word count:" grep -ic 'notamment' document.md wc -w document.md # Anglicisms grep -icE "(faire du sens|adresser (un |le |ce )?probl|impacter|definitivement)" document.md # Em dash density grep -c '—' document.md # English-style quotes in French text grep -cE '["""]' document.md ``` ### A.13 Summary: English vs French AI Tells | Category | English tell | French equivalent | |----------|-------------|-------------------| | Most notorious word | "delve" | "crucial" | | Most overused verb | "leverage" | "permettre de" | | Vapid opener | "In today's fast-paced world" | "Dans un monde en constante evolution" | | Favorite transition | "Moreover" / "Furthermore" | "Par ailleurs" / "En outre" | | Hedging phrase | "It's worth noting that" | "Il est important de noter que" | | Register problem | Corporate-speak | Langue de bois | | Typography tell | Em dash overuse | Em dashes + English quotes + missing thin spaces | | Parallelism tic | "It's not X, it's Y" | "Non seulement X, mais aussi Y" | | Pronoun tell | N/A | "Nous" where "on" is natural | | Quantitative test | N/A | "Notamment" frequency (1/200 vs 1/800) | | Anglicism tell | N/A | "Faire du sens", "adresser un probleme" | --- ## References - **Kobak et al. (2024)** -- "Delve into ChatGPT usage in scientific writing: monitoring excess word usage in academic papers." Documented 25x spike in "delve" and other LLM-signature words in academic publications post-ChatGPT. - **Liang et al. (2024)** -- "Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews." Found 10-17% of ICLR/NeurIPS reviews showed substantial LLM modification. Identified "commendable" (9.4x), "meticulous" (7.4x), "intricate" (4.5x) as top markers. - **Gao et al. (2024)** -- Frequency analysis of scientific abstracts showing statistically significant increases in "realm", "intricate", "pivotal" post-2023. - **Simon Willison (2024)** -- "Slop is the new spam." Coined/popularized the term for unwanted AI-generated content. - **Dan Shipper** -- "The Field Guide to AI Slop" ([ignorance.ai](https://www.ignorance.ai/p/the-field-guide-to-ai-slop)). The original catalogue of LLM prose fingerprints: filler intensifiers, vapid openers, unearned profundity, parallelism tics, triadic lists, em dash overuse, generic analogies, bold/italic abuse, avoided contractions. ### French-specific sources - **Daria Decrypte l'IA** -- "Les tics de langage de ChatGPT." Detailed analysis of ChatGPT's French verbal tics. - **Generation IA / Flint Media** -- "Comment dejouer les tics de langage de ChatGPT." Practical guide to avoiding French AI patterns. - **Projet Voltaire** -- "Comment detecter un texte produit par ChatGPT." Linguistic analysis from France's leading language-quality platform. - **Cours NDRC (2026)** -- "Comment Detecter l'Ecriture IA: 15 Signes a Reconnaitre." Quantitative metrics for French AI detection (TTR, adverb density, sentence variance). - **Siecle Digital / Fnac L'Eclaireur / CCFI** -- Multiple articles on the tiret cadratin (em dash) as ChatGPT's most recognizable French tell. - **Journal du Net** -- "Si votre collegue emploie cette expression." Top 5 most overused ChatGPT French phrases. - **Shaib et al. (arXiv, 2509.19163)** -- "Measuring AI Slop in Text." Cross-language frequency analysis of AI-overused terms.