If you spend enough time reading on the web today, you start to feel an unsettling sense of linguistic déjà vu. Every third essay opens with a sweeping observation about an “ever-evolving landscape.” Every critique pivots on a dramatic reframe (“It’s not X—it’s Y”). Every product announcement promises to “quietly orchestrate” your life while honoring the “rich tapestry” of modern work.
We are drowning in synthetic cadence. Generative models haven’t just automated prose; they’ve standardized a dialect.
Enter tropes.fyi, a living taxonomy of AI writing habits curated by designer and engineer Ossama Al-Qarni. It is less a standard style guide and more a field guide to machine-generated mannerisms, cataloging the subtle, pervasive patterns that instantly signal a piece of text was spat out by a large language model.
The Mechanics of Synthetic Prose
The obvious tells were identified early on. In 2023, anyone with an internet connection noticed the sudden, viral explosion of the word delve. Overnight, an obscure verb became the universal prefix for every introductory paragraph on LinkedIn.
But as Al-Qarni’s index demonstrates, vocabulary is only the surface level of the problem. The real tell of synthetic text isn’t a single word; it’s the underlying structural rhythm—what the guide calls the “architecture of artificial profundity.”
Consider the structural habits cataloged on the site:
- Negative Parallelism: The compulsive reliance on “It’s not X—it’s Y.” It’s an easy mechanism to manufacture instant depth by pretending to subvert expectations.
- Tricolon Abuse: Grouping ideas into rhythmic sets of three (“identity, payments, and compute”) until the prose reads like a metronome rather than a human voice.
- The “Serves As” Dodge: Replacing simple copulas (“is”, “are”) with grander alternatives like “stands as” or “serves as a testament to.”
- Bold-First Bullets: Formatting every single list item with a bolded header phrase, a habit inherited directly from Reinforcement Learning from Human Feedback (RLHF) preferences tuned for scannability over voice.
When these tropes compound within an essay, the text loses its human friction. It becomes frictionless, sterile, and ultimately unreadable.
Why Models Write This Way
These patterns aren’t random quirks; they are the direct mathematical byproduct of how modern models are trained.
Because LLMs predict the next most probable token, they naturally gravitate toward statistical averages. When you combine base probabilistic modeling with RLHF—where human evaluators consistently rate structured, agreeable, and authoritative-sounding responses higher—you end up with a model that defaults to a very specific persona. It writes like a overly enthusiastic middle-management consultant who just finished a workshop on public speaking.
The model avoids simple declarative statements because simple statements carry risk and lack the surface polish of complex clauses. Instead, it reaches for “False Ranges” (“from innovation to transformation”) and “Grandiose Stakes Inflation” to make mundane observations feel world-historical.
The ironic result is that in an effort to make AI outputs sound helpful and polished, alignment training created a distinct dialect that human readers now instinctually recoil from.
The Curation Imperative
When the cost of generating text drops to zero, the value of written work shifts entirely to voice, precision, and editing.
Tools like tropes.fyi (and its raw, system-prompt-ready markdown repository) matter because they give us a vocabulary to debug our own prompts and editing processes. Feeding these anti-patterns directly into an LLM’s system instructions helps strip away the performative fluff, forcing the model closer to plain, direct English.
Yet, relying solely on negative constraints is only half the battle. You can ban “delve,” forbid double hyphens, and blacklist “tapestry,” but if the underlying thought lacks a clear perspective, the model will simply invent new ways to be bland.
Good writing—whether produced by a human sitting at a mechanical keyboard or a human steering an inference engine—requires specificity. It requires real examples, logical leaps, asymmetric observations, and the willingness to take a clear stance without wrapping it in three layers of polite hedge-phrasing.
The web doesn’t need more “robust frameworks for navigating complex landscapes.” It needs sharp arguments, explicit attribution, and prose that sounds like it was written by someone with actual skin in the game.
(via tropes.fyi)