The Rise of 'LLM Smells': Recognizing the Patterns of AI-Generated Content

For many, the first encounter with Large Language Model (LLM) assistance feels like a revelation. The prose is polished, the vocabulary is expansive, and the sentence structures are sophisticated. However, as these tools have permeated every corner of the internet, a new phenomenon has emerged: the "LLM smell."

What begins as a perceived improvement in quality eventually reveals itself as a predictable pattern. When the same "sophisticated" structures appear across disparate blogs, LinkedIn posts, and landing pages, the polish starts to feel like a mask. We are now entering an era where the very markers of "good" writing are becoming the primary indicators of artificiality.

The Anatomy of AI Writing Smells

AI-generated text often suffers from a specific kind of homogeneity. While the models are capable of producing grammatically perfect prose, they frequently lean on rhetorical crutches that create a distinct, recognizable rhythm.

Rhetorical Tropes and Sentence Structures

One of the most pervasive patterns is Contrastive Negation. This is the "It’s not just X, it’s Y" or "Not X, but Y" formula. While effective in moderation, LLMs use this structure with an obsessive frequency to create a sense of depth or revelation.

Other common linguistic tells include:

  • The "Punchline" Habit: A tendency to end paragraphs with a short, profound-sounding sentence that attempts to encapsulate a complex idea into a neat bow (e.g., "Symmetry becomes a trap").
  • Consecutive Short Sentences: A rhythmic pattern of three or four brief sentences used to create artificial tension or drama.
  • The "X is the Y of Z" Formula: A penchant for metaphorical definitions, such as "Cringe is the visible signature of moving along a gradient you chose."
  • Predictable Lead-ins: Phrases like "The honest caveat:" or "The thing to internalize:" which signal to the reader that a key point is coming.

Vocabulary and "Corporate-Speak"

Beyond structure, certain words and metaphors have become hallmarks of AI slop. Terms like "blast radius," "load bearing" (used non-architecturally), and "smoking gun" appear with a frequency that far exceeds their natural usage in human conversation. More recently, users have noted Claude's obsession with the term "inside baseball" to describe niche topics.

The Visual Smell: AI-Generated Web Design

The "smell" extends beyond text into the visual realm. AI-assisted web development often produces sites that look professional at a glance but share an eerie sameness.

Common visual tells include:

  • Typography: An over-reliance on specific fonts like JetBrains Mono for a "modern tech" aesthetic.
  • UI Components: The ubiquitous use of the same card layouts, specific button styles, and the "blinking dot" badge component to signal activity.
  • Layout Patterns: A rigid adherence to "step-by-step" bulleted sections on almost every landing page.

While some argue that this sameness is actually a benefit—providing legibility and a standard user experience—others see it as a loss of idiosyncratic design that makes a brand feel human.

The Authenticity Paradox

The rise of LLM smells has created a strange psychological shift in how we communicate. As AI-generated content becomes the baseline for "polished" professional writing, humans are beginning to intentionally degrade their own output to signal authenticity.

"I've noticed my own writing has been getting a lot sloppier the past couple years just so that my writing doesn't get mistaken for AI."

From intentionally leaving typos in Slack messages to avoiding the use of em-dashes (which AI uses liberally), users are adopting a form of "digital camouflage." The goal is to prove a human was behind the keyboard, even if it means sacrificing the very polish that LLMs provide.

Strategies for Using LLMs Without the Smell

To leverage the power of AI without falling into the trap of homogeneity, technical writers and creators are suggesting a shift in how they use these tools. Instead of asking an LLM to "write" or "polish" a draft, use it as a structural editor.

  • Critique, Don't Generate: Ask the LLM to spot overused words, identify passive constructions, or critique the flow of a topic sentence.
  • Manual Integration: Avoid copying and pasting vocabulary adjustments. If an LLM suggests a better word, evaluate it against your own voice before integrating it.
  • The "Smell Test" Pass: Perform a final rigorous edit specifically to eliminate contrastive negations, excessive em-dashes, and artificial punchlines.

Ultimately, the danger of LLM usage isn't the lack of quality, but the lack of voice. When we outsource the act of writing, we outsource the unique imperfections that allow a reader to connect with a human mind.

Sources