The 'AI Smell': Detectable Patterns in Machine-Generated Content

The 'AI Smell': Detectable Patterns in Machine-Generated Content

Late last year, I started writing a math blog and decided to experiment with using LLMs to polish and enhance my writing. The AI-generated text immediately felt significantly better than my own writing. It had richer vocabulary, more interesting sentence structures, and a certain polish that my drafts lacked. At the time, I genuinely believed this wasn't the typical 'AI-slop' that many people complain about.

Then, about three months later, I began seeing the exact same sentence structures appearing across the entire internet. What's fascinating is that this 'ai-smell' seems like an artifact that emerges across various AI-assisted tasks—patterns you can now easily recognize once you know what to look for.

Writing Patterns That Reveal AI Assistance

Beyond the obvious overuse of em-dashes, several distinctive patterns have emerged in LLM-generated writing:

1. Excessive Punchlines LLMs seem to have a particular affinity for crafting memorable, quotable statements. While human writers certainly do this too, AI tends to overdo it:

"Humans trust symmetry because it feels like intelligence made visible." "The Tiger fit the story. Jin-yong fit the physics." "Symmetry becomes a trap."

These statements function as rhetorical zingers, designed to leave an impression but often lacking substantive connection to the surrounding text.

2. Consecutive Short Sentences Another common pattern is the use of short, declarative sentences in quick succession:

"Yet the tilt is not an accident. It is the shape of the optimum." "Then AlphaEvolve arrived. It had no preference for symmetry. No aesthetic prior. No instinct to preserve harmony." "These examples are not decorative. They form a distributed argument."

This creates a staccato rhythm that feels artificial to experienced readers. Human writing typically varies sentence length more naturally.

3. The 'X is the Y of Z' Formula This construction appears frequently in AI-generated text:

"Cringe is the visible signature of moving along a gradient you chose."

While sometimes effective, overuse of this formulaic structure creates a recognizable pattern.

4. The 'not just X, but Y' Construction AI seems particularly fond of this rhetorical device:

"Solutions that do not merely satisfy the constraint but satisfy the aesthetic instincts"

This structure adds artificial emphasis but often doesn't contribute substantive meaning.

AI Website Design Patterns

The 'AI smell' extends beyond writing into web design, where several distinctive patterns have emerged:

1. The JetBrains Mono Font Obsession Many AI-generated websites default to JetBrains Mono, a monospaced font typically used for code. This creates a distinctive aesthetic that screams 'AI-generated' to designers.

2. Ubiquitous 'Step' and Bullet Points Websites created with AI assistance often feature identical step-by-step instructions or bullet points with the exact same formatting, particularly when using JetBrains Mono.

3. Homogeneous Button Design AI-generated websites tend to use remarkably similar button designs—often with rounded corners, specific shadow effects, and hover animations that follow common UI patterns found in training data.

4. Card-Based Layouts The card-based layout has become so common in AI-generated websites that it's become a telltale sign. These cards typically follow the same grid structures, spacing, and shadow patterns.

5. Blinking Dot Badge Components A particularly distinctive pattern is the use of a small blinking dot within badge components, often used to indicate 'new' or 'unread' status. This specific interaction pattern has become surprisingly common in AI-generated interfaces.

Why These Patterns Emerge

These 'AI smells' emerge because LLMs are trained on massive datasets that contain these patterns. The models don't truly understand context or creativity in the human sense; instead, they recognize statistical patterns in their training data and replicate them.

When many people use similar prompts or templates, the resulting content converges around these patterns, creating the recognizable 'AI smell.' This is particularly evident in creative tasks where the models attempt to mimic human style without truly understanding the underlying principles.

Implications for Content Authenticity

The emergence of these detectable patterns has significant implications for how we evaluate content authenticity:

1. Detection Tools: As these patterns become more widely recognized, detection tools will improve at identifying AI-generated content.

2. Prompt Engineering: Users will develop more sophisticated prompting techniques to avoid these patterns.

3. Human Collaboration: The most valuable content will likely emerge from human-AI collaboration, where humans guide the AI to avoid these telltale patterns.

4. Style Differentiation: As users become aware of these patterns, they'll actively work to develop distinctive styles that stand out from AI-generated content.

The Path Forward

I'm not against using LLMs or AI for creative tasks. These tools can certainly enhance productivity and provide inspiration. However, as these patterns become more widely recognized, we'll need to develop more nuanced approaches to AI-assisted content creation.

The most valuable approach may be to use AI as a starting point or tool for enhancement rather than relying on it for complete content creation. By understanding these 'AI smells,' creators can consciously work to avoid them, developing content that maintains human authenticity while still benefiting from AI assistance.

As AI continues to evolve, we'll likely see these patterns change and adapt. New 'smells' will emerge, and detection methods will improve. This ongoing evolution represents an interesting challenge in the relationship between human creativity and machine assistance.