Rule candidate: contrastive 'is X, not Y' framing (AI-tell)

[yzhao062]

Pattern

LLMs over-use "X is A, not B" / "X is not A but B" framings, especially in titles, abstracts, and topic sentences. The construction feels insightful (negation + reframe = perceived depth), but in scientific prose it usually:

1. Negates a strawman no reader actually held.
2. Adds words without adding information.
3. Reads as AI-generated because the pattern is so prevalent in LLM output.

Example of the pattern

PyOD 3: When the Agentic Layer Is an Interface, Not an Accuracy Gain

The "is an X, not a Y" structure is the diagnostic shape. Common variants:

• "X is not just A — it's B"
• "It is not A that matters, but B"
• "Less A, more B"
• "X is A, but more importantly B"
• "The point is not to A but to B"

Why it's an AI-tell

Negation + reframe is a chiasmus-like rhetorical move that LLMs over-apply because it scores high on perceived insight per token. Human writers reach for it sparingly, when the reframe is genuinely the point of the sentence. AI writers reach for it as a default opener for titles, abstracts, and section ledes.

BAD → GOOD direction (rough sketch; finalize during rule write-up)

BAD: "The benefit is not raw accuracy, but the interface it enables."
GOOD: "The benefit is the interface this layer enables." (Or, if the contrast is the point, write the previously expected outcome and the reframe as two sentences rather than collapsing them into one negation.)

BAD: "It is not the size of the model that matters, but the data it was trained on."
GOOD: "Data quality drives the result more than model size."

BAD (title): "When the Agentic Layer Is an Interface, Not an Accuracy Gain"
GOOD (title): "The Agentic Layer as an Interface" (lead with the positive claim; if the surprise is that it does not boost accuracy, say so in the abstract where it can be substantiated, not in the title).

Notes for the future rule

• Slot: the field-observed bucket (RULE-A through RULE-I). Next free letter is RULE-J.
• Regex sketch (illustrative, not the final detector): (is|are|was|were)\s+(not\s+just\s+|not\s+)?[a-zA-Z\-]+\s*[,—]\s*(but\s+|not\s+)
• High false-positive risk: legitimate contrastive uses do exist (e.g., correcting a documented prior claim). The rule body should describe when the reframe is the genuine thesis vs when the negation is a rhetorical filler.
• Adjacent rule: RULE-D (transition overuse) already covers "Additionally" / "Furthermore". This rule covers a distinct rhetorical move: contrastive negation framing.

Opening as an issue so I can fix later.