The Illusion of "Banned Words"

Lists of suspicious transition words ("Furthermore", "In conclusion", "Therefore") only work for a few weeks until models adapt their vocabulary. The true giveaway isn't in individual words, but in the deep architecture of the text.

Everyone looks for giveaway words that expose AI-generated text.

"Furthermore"
"Therefore"
"In conclusion"

That kind of list works... for a few weeks. Then the model updates, learns new habits, and swaps one connective phrase for another.

Individual words reveal very little. The deep structure of the text tells the real story.

When you understand how an LLM writes, you notice far deeper habits that persist regardless of which model is used.


1. The Text Never Takes a Risk

One of the easiest symptoms to spot is permanent neutrality. Language models were trained via RLHF to minimize conflict. This produces prose that avoids stating anything with conviction.

AI Pattern (Lukewarm Neutrality)

"Kubernetes can bring benefits, but also challenges, depending on the organizational context."

Technically correct, completely forgettable.
Human Take (Risk Included)

"Kubernetes overcomplicated operations that used to be trivial."

Direct, opinionated, and sparks genuine debate.

When a text spends 2,000 words without taking a stance, there's usually a model trying not to offend any potential reader.


2. Every Paragraph Serves the Exact Same Function

Another habit is structural. Observe AI-generated articles: most follow the exact same mechanical loop.

The Infinite AI Paragraph Loop

1. Introduce idea
2. Explain idea
3. Reinforce idea
4. Summarize idea
5. Transition

This cycle repeats dozens of times. Humans don't write like this. Sometimes a paragraph exists solely because the author recalled an interesting aside. Sometimes it abruptly halts because they've said enough.


3. The Model Is Afraid to Leave Gaps

LLMs hate implicit information. They try to over-explain everything:

  • Defining obvious concepts.
  • Over-explaining analogies.
  • Repeating context and over-justifying assertions.

The result is text that feels like it's constantly teaching a complete novice. Experienced writers do the exact opposite: they trust the reader to connect the dots.


4. Information Density Is Abysmal

This is one of the strongest indicators. Finish reading a 2,500-word article and ask yourself:

"How many genuinely new ideas appeared here?"

In many AI-generated articles, the answer is surprisingly small. The model rewrites the same information using different sentence structures. There is lexical variation; there is no intellectual progression.


5. The Text Explains Rather Than Demonstrates

AI loves abstract claims. Humans demonstrate through concrete scenes and real observations.

Abstract Explanation (AI)

"Efficient communication is critical for tech teams."

Concrete Demonstration (Human)

"Just sit through a sprint retrospective where forty minutes are spent debating a single Jira word."

One explains; the other demonstrates. Models produce far more explanation than observation.


6. Analogies Never Break

When a model starts a metaphor, it stays rigidly consistent to the end. If it starts comparing software to civil engineering, it will meticulously discuss foundations, bricks, walls, and finishes.

People rarely do this. They pivot mid-way, abandon a metaphor, or mix references unpredictably. It looks like a stylistic flaw; in reality, it's a deeply human trait.


7. Everything Feels Over-Optimized

There is a difference between writing to communicate and writing to maximize statistical smoothness metrics. LLMs do the latter.

  • Everything is perfectly balanced.
  • Sentence lengths are uniform.
  • Transitions are suspiciously frictionless.

After a few minutes of reading, a strange feeling sets in: there is no friction. And memorable writing always has friction.


8. The Author's Voice Disappears

Perhaps the ultimate symptom. After reading ten articles by the same human author, you recognize their writing: their obsessions, comparisons, humor, and pet peeves.

With AI content, the opposite happens: you recognize the model, not the author.

Takeaway

Traces of Thought vs. Patterns of Writing

That is why so many companies publish hundreds of technically flawless articles that nobody remembers a week later. Grammar, vocabulary, and transition words are just the surface layer. Language models learn patterns of writing, while human beings leave traces of thought.