If you read much Chinese text generated by large language models, you've surely noticed this pattern:
"不是X，而是Y。" ("Not X, but Y.")
不是简单的堆砌，而是精心组织的表达。不是表面的修饰，而是深层的重构。不是机械的重复，而是有意的选择。
— See that? Just casually writing three sentences, I couldn't resist using it three times. This construction is the LLM's verbal tic, so dense that readers instantly recognize "this is AI-written."
But a new paper raises a counterintuitive question: is the problem really whether the model uses this pattern — or when it uses it?
Not Whether, but Appropriateness
The paper, "Not X, but When: Diagnosing a Mandarin LLM Writing Tic through Contextual Appropriateness" from the University of Toronto and Tsinghua University, makes a careful distinction.
The researchers point out that "不是X，而是Y" is perfectly normal Chinese. Human writers use it too, and when used well it's powerful — it corrects a misconception or replaces an alternative. The real question is: does the model know when to use it and when not to?
They borrow linguist Hymes' framework of "communicative competence." Hymes argued that language competence is not just the ability to produce grammatical sentences (grammaticality), but also knowing whether an expression is contextually appropriate (appropriateness). Prior AI research focused on frequency — how many times more often an expression appears in AI text than human text — and treated that as a target to suppress. But frequency only answers Hymes' last question: how often was the expression actually produced? For expressions with legitimate functions, frequency conflates appropriate use with inappropriate use.
A Decision Tree: Four Usage Types
The researchers built a decision tree classifying each use of "不是X，而是Y" into four categories:
- Substantive (S): X is genuinely a "live alternative" in context — someone mentioned it, or it's a common assumption — and negating it adds new information. Fully appropriate.
- Optional (O_L): X is not a plausible alternative in context, so negating it is meaningless. Like saying "it's not sunny, it's raining" when nobody claimed it was sunny.
- Optional (O_U): X is an alternative, but negating it adds no new information because the context already implies the answer.
- Defective (D): X is not an alternative AND negating it carries no information. A double failure.
- Only 1.30% are "defective" uses — the models almost never make the crude "completely wrong" error.
- But 52.8% are "optional" uses (O_L + O_U) — grammatically correct, semantically valid, but simply unnecessary. The models have little trouble with correctness, but fall far short on appropriateness.
- The five commercial models use the pattern at 2.5 to 9.2 times the density of human reference texts. Humans use it too — but humans know when to stop.
What the Data Shows
The researchers annotated 307 instances of the pattern generated by five commercial models, and the results are surprising:
The More Disturbing Finding: Reward Models Encourage It
The researchers ran a chilling experiment: they asked reward models to choose between quality-matched text pairs — one containing the pattern, one without. Multiple reward models clearly preferred the version with the pattern.
Worse: when LLM-as-judge made the choice in contexts that didn't support the pattern, the LLM judges also preferred the version with it.
What does this mean? If you train a model with RLHF, the reward model systematically encourages more of this pattern — even when context doesn't call for it. The study confirms: SFT and preference optimization both increase the pattern's frequency.
We think we're training models to "write better"; we're actually training them to "sound more like AI."
An Analogy
Imagine someone who just learned a fancy term like "paradigm shift." They use it in academic papers, in meetings, in small talk, even ordering takeout: "It's not switching restaurants — it's a paradigm shift."
Every use is grammatical and semantically valid, but none are appropriate. The problem isn't whether they can use the term — it's that they don't know when not to.
The LLM's "不是X，而是Y" is exactly this. The model learned the form, but not the pragmatic judgment of appropriateness — when X is a "live alternative," and when negating it wastes the reader's attention.
Why This Is Hard to Fix
The traditional approach — detect high-frequency expressions and suppress them — misfires here. The pattern has legitimate functions; sometimes it's exactly right. Bluntly lowering the frequency would kill appropriate uses along with inappropriate ones.
The proposed framework is more precise: not "how often was it used," but "was each use right." That requires pragmatic judgment, not simple frequency statistics.
This also explains why the problem is so stubborn — reward models like it, SFT reinforces it, preference optimization pushes it up. The whole training pipeline systematically pushes in the wrong direction. The model doesn't not know the pattern; it doesn't know when to hold back.
Special Significance for Chinese AI
The paper's focus on Mandarin is no accident. Chinese internet communities are far more sensitive to the "不是X，而是Y" pattern than English communities are to "not X, but Y." Chinese formal registers already use "不是……而是……" at high frequency; when models amplify that further, reader perception is especially strong.
It's a reminder that "AI-speak" has different faces in different languages. English AI tells might be "delve" and "tapestry"; Chinese AI tells might be "不是X，而是Y" and "值得注意的是" (it's worth noting that). Remedies must be language-specific, but the principle is the same: target appropriateness, not just frequency.
Conclusion
What's most admirable about this paper is that it upgrades "AI-speak" from a frequency problem to a pragmatic one. Previously we asked "how many times did the model use it"; this paper asks "does the model know when to use it." The former only tells you how much it overshoots; the latter tells you where the problem lies.
"不是X，而是Y" isn't the AI's fault — it's the result of humans training AI while treating frequency as a proxy for quality, systematically encouraging inappropriate use. Solving it requires not just suppressing frequency, but teaching models pragmatic appropriateness.
Paper: https://arxiv.org/abs/2610.22249