Decoding LLM Clichés: A Fresh Perspective

Like millions of others convinced they possess knowledge the world desperately needs to hear, I decided to write a book on prompting. In the process (which, by the way, turned out to be far more difficult than anticipated), I found myself examining LLM clichés. You know the ones. At least, in the comment sections of tech blogs, hundreds of self-proclaimed experts use them to spot AI-generated text.

Anyway, these clichés definitely exist, and many authors now routinely add blocklists of these phrases to their prompts to weed them out. Whether this is actually a good or a bad thing is what I’ll break down below.

Editors, writers, and readers unanimously agree that LLM-generated texts are easily recognized by their heavy reliance on certain tropes: “It’s not just [X], it’s a [Y]…”, “Unlocking the potential of…”, “In today’s fast-paced digital landscape…”, and so on. Editors specifically demand that these clichés be excised or rephrased. It’s somewhat ironic how the adoption of LLMs has shifted how we evaluate writing. Previously, a popular-science author wrote in whatever way they felt was clearest for their audience; now, before publishing, they are forced to ruthlessly edit out expressions that might actually be perfectly natural to their personal style.

Sure, the simple fix is to explicitly instruct the LLM in your prompt not to use these blacklisted phrases and structures. The model will gladly comply if you tell it to. But every constraint has consequences, and catering to an audience that is overly demanding about form (rather than substance) is no exception.

For a language model, a word isn’t just the output of a token into a chat window. For a neural network trained on the patterns of millions of texts, language is its mode of reasoning — a router navigating through semantic space. Any given word or phrase carries not just a definition, but a cluster of interconnected patterns. By banning a word or a construction that you deem an “LLM cliché”, you are potentially disabling the model’s cognitive tools.

I ran a quick experiment using Gemini. I used three prompts across three separate sessions, asking it to explain the difference between overfitting and underfitting:

1. Explain the difference between overfitting and underfitting.

2. Explain the difference…, but do not use the “not X, but Y” construction.

3. Explain the difference…, explicitly contrasting the concepts.

These three prompts yielded noticeably different results. The direct request and the explicit instruction to contrast the concepts produced excellent explanations with comprehensive coverage of the topic. However, the response to the prompt that banned the “not X, but Y” structure completely dropped the exact concepts that inherently require contrast to be articulated: the bias-variance tradeoff, mitigation strategies, and underlying causes. In other words, it lost the very concepts that serve as semantic boundaries.

You can easily replicate this experiment with any moderately complex concepts. And feel free to tear my argument to shreds in the comments if your findings differ.

Ultimately, any linguistic ban on specific constructions can lead to a degradation in the model’s reasoning capabilities. This is because, for an LLM, linguistic patterns are reasoning.

Perhaps this is exactly why so many popular articles have lost their semantic weight, even while confidently passing AI content detectors.

Constructions You Shouldn’t Ban

Here is a practical list of what constitutes a cognitive tool rather than a mere stylistic device:

• Contrast and boundary. “Not X, but Y” forms a strict boundary between concepts—without it, the model blurs the distinction. “X, unlike Y” triggers comparative analysis. “X, whereas Y” holds two objects in context simultaneously.

• Causal chains. “Because” forces the model to explain rather than merely state. “Therefore” and “hence” force a deduction from premises. “If… then” engages conditional logic.

• Hierarchy and classification. “In particular” and “for example” represent movement from the general to the specific. “That is” (i.e.) acts as rephrasing that verifies understanding. “From the perspective of X” is an explicit designation of the angle of analysis.

• Limitation and qualification (caveats). “Provided that” introduces an invariant. “With the exception of” forms a boundary through negation. “Only if” sets a strict logical condition.

• Temporal structure. “First… then… ultimately” imposes a forced sequence. “Before” and “after” establish the causal order of events.

• Epistemic modality. “Likely,” “possibly,” and “certainly” provide confidence calibration. Without them, the model speaks with a flat intonation, leaving the reader unable to distinguish fact from assumption. “As a rule, but not always” is an explicit pointer to exceptions.

You can absolutely use a prompt to ban filler adjectives (fluff) like “unique,” “innovative,” or “revolutionary.” They carry no critical semantic weight. But be extremely careful with constructions that define relationships between objects: cause, effect, boundary, condition, and hierarchy.

Any structure that establishes a relationship between objects is an instrument of reasoning. By banning it, you are removing not a decoration, but a connection. For an LLM, language doesn’t describe reasoning—language is reasoning. Words like “not,” “because,” and “therefore” serve as logical operators, not as signs of excessive verbosity or purple prose. For an LLM, what appears to a human as mere style is, in fact, substance.

If you want to make cosmetic changes to cater to the reader, it’s much safer to manually weed out perceived clichés after the text has been generated. Otherwise, you run the very real risk of losing a significant portion of your underlying meaning.

Final Thoughts

An amusing corollary: when LLM developers work to reduce the frequency of AI clichés, they actively risk degrading the model’s intelligence. I strongly suspect this is exactly what happened with ChatGPT. For instance, the forced reduction in the probability of the universally despised “let’s delve into” didn’t just suppress that specific phrase; it suppressed an entire cluster of related patterns. This, in turn, weakened the activation of the very concept of deeply exploring a problem, prompting the model to default to superficial overviews.

I hope that someday this allergy to classical logic will pass, and readers will begin to focus on the substance of a text rather than its form. Because a genuine thought is not just a conveniently digestible piece of text—it is a boundary that separates meanings.