30 Signs of AI-Generated Text — The Complete tropes.md Breakdown

AI-generated text has distinctive patterns. One or two appearances aren't a problem, but when multiple patterns appear simultaneously or repeatedly, it signals "this was written by AI."

tropes.fyi's tropes.md catalogs these patterns into 6 major categories with 30+ specific items in a single markdown file. Adding it to AI system prompts can guide models away from common AI prose.

1. Word Choice

The most immediately visible AI signals.

"Quietly" and magic adverbs: Overusing "quietly", "deeply", "fundamentally", "remarkably", "arguably" to give mundane descriptions subtle importance. "Quietly orchestrating workflows" implies something significant happening without anyone noticing.

"Delve" and synonyms: Once the most famous AI tell. "Certainly", "utilize", "leverage" (as verb), "robust", "streamline", "harness" belong to the same family. Appearing at abnormally high frequency in AI text.

"Tapestry" and "Landscape": Using grandiose nouns where simple words suffice. "Tapestry" for anything interconnected, "landscape" for any field. "Paradigm", "synergy", "ecosystem" are the same type.

"Serves As" avoidance: Using "serves as", "stands as", "marks", "represents" instead of simple "is/are". Caused by AI's repetition penalty pushing away from basic copula toward flashier constructions.

2. Sentence Structure

The most systematically occurring domain.

Negative Parallelism "It's not X — it's Y": The most commonly identified AI tell. Wraps everything as a surprising reframe, generating false profundity. This volume of such writing didn't exist before LLMs.

"Not X. Not Y. Just Z.": Dramatic countdown negating two things before revealing the point. Creates a false feeling of narrowing toward truth.

"The X? A Y.": Rhetorical question-instant-answer pattern. Poses questions nobody asked and immediately answers them. AI considers this the essence of great writing.

Superficial Analyses: Appending present participle (-ing) phrases to inject shallow analysis. "Highlighting its importance", "reflecting broader trends."

False Ranges "from X to Y": X and Y aren't on any actual scale. Loosely related things listed as if they're a spectrum.

Gerund Fragment Litany: Stringing subjectless gerund fragments after a claim. "Fixing small bugs. Writing straightforward features." Humans don't draft this way.

3. Paragraph Structure

Short Punchy Fragments: Very short sentences as standalone paragraphs for artificial emphasis. Result of RLHF training pushing toward "readability."

Listicle in a Trench Coat: Disguising numbered points as continuous prose. "The first... The second... The third..." paragraphs hiding list format.

4. Tone

Most subtle but most grating when accumulated.

"Here's the Kicker": False suspense transitions promising revelation for points that need no buildup. "Here's the thing", "Here's where it gets interesting" are the same type.

"Think of It As...": Teacher-mode default assuming readers need analogies. AI frequently generates analogies less clear than the original concept.

False Vulnerability: Performative self-awareness pretending to break the fourth wall. Real vulnerability is specific and uncomfortable; AI's is polished and risk-free.

Grandiose Stakes Inflation: Inflating every point to world-historical importance. A blog post about API pricing becomes a meditation on civilization's fate.

Vague Attributions: Attributing claims to nameless authorities — "experts", "industry reports" — without specific sources.

Invented Concept Labels: Attaching abstract problem-nouns (paradox, trap, creep) to domain words. "Supervision paradox", "acceleration trap" — sounds analytical but is groundless.

5. Formatting

The most visually immediate AI tells.

Em-Dash Addiction: Compulsive overuse for dramatic pauses. Humans use 2-3 per piece naturally; AI uses 20+.

Bold-First Bullets: Every bullet point starting with a bold phrase. Extremely common in Claude and ChatGPT output; almost nobody formats this way manually.

Unicode Decoration: → arrows, smart quotes — special characters not easily typed on standard keyboards. Claude particularly favors the → arrow.

6. Composition

Structural patterns at the whole-document level.

Fractal Summaries: Applying "what I'll say, what I'm saying, what I said" at every level. Every subsection gets its own summary.

Dead Metaphor: Fixating on one metaphor and repeating it 5-10 times throughout. Humans introduce a metaphor, use it, and move on.

One-Point Dilution: Restating a single point 10 different ways across thousands of words. Repeating the same idea with different analogies to appear "comprehensive."

Signposted Conclusion: "In conclusion", "To sum up" — explicitly announcing the conclusion. Skilled writing lets readers feel the conclusion without being told.

Core Principle

It's about frequency and combination. Using "delve" once is fine. But using "delve" + "It's not X — it's Y" + 20 em-dashes + bold-first bullets in the same piece — that's AI writing.

The irony: this file itself was written with AI assistance, including a disclaimer: "AI for AI, humans for humans."