Why “It’s Not X, It’s Y” Isn’t the Problem – It’s the Metrics We Use to Police Language

What’s being claimed

Recent blog posts and social‑media threads warn that the phrase “It’s not X, it’s Y” – a classic example of negative parallelism – is a tell‑tale sign of AI‑generated text. AI‑detector services such as Grammarly, Turnitin, and the newer Pangram claim that spotting this construction (or other high‑frequency token pairs) can reliably expose machine‑written prose. The narrative is simple: the pattern is a stylistic fingerprint of large language models (LLMs), so if you see it, you’re probably looking at a bot.

What’s actually new

The underlying cause is not the phrase itself but the way modern LLMs are fine‑tuned. Two training tricks have become commonplace:

1. Reinforcement Learning from Human Feedback (RLHF). Human annotators rank model outputs, and the model is rewarded for mimicking the highest‑ranked responses. This pushes the model toward the phrasing that humans prefer in a short‑answer setting – often concise, contrastive statements like “It’s not X, it’s Y.”

2. Reinforcement Learning from Verified Rewards (RLVR). A newer variant where the model is asked to solve a problem step‑by‑step, and the entire reasoning chain is rewarded if the final answer is correct. The model learns that explicit contrastive connectors ("but", "however", "instead") help it steer toward the right answer, so it repeats them more often.

Both techniques bias the token distribution toward high‑entropy connective words. The result is a measurable increase in the frequency of negative parallelism across benchmark prompts, which detectors then treat as a statistical anomaly.

Why the detectors flag it

Detectors work by comparing the observed n‑gram distribution in a document to a reference distribution derived from known human and AI corpora. When a piece of text contains a cluster of high‑probability AI n‑grams – such as "automated language production", "align with", or the exact "It’s not X, it’s Y" template – the model assigns a higher AI‑likelihood score. Grammarly’s recent update, for example, flags the phrase “automated language production” as 11× more likely to be AI‑generated and suggests a synonym swap to "mechanized language synthesis". The suggestion is not about meaning; it is about moving the token statistics away from the AI‑trained region.

Limitations of the approach

Limitation	Explanation
Over‑reliance on form	Detectors ignore semantics. A human writer can legitimately use negative parallelism for rhetorical effect, as JFK did. Flagging based on form penalises good writing.
Training‑data drift	Most detectors are trained on pre‑2021 corpora. As models evolve, their token distributions shift, making older detectors increasingly inaccurate.
False‑positive cascade	A 99.8 % per‑document accuracy sounds impressive, but when applied to millions of student essays the cumulative false‑positive rate can exceed 5 %, meaning thousands of innocent writers are flagged.
Feedback loop	When writers rewrite to avoid detector triggers, they alter their style to match a different statistical profile, which in turn becomes the new target for future detectors. The cycle erodes natural variation in prose.

The broader impact on assessment and expression

Automated grading

A recent UK pilot of an AI‑based essay scorer showed that the system rewarded longer essays with richer vocabulary and complex sentence structures – exactly the traits amplified by RLVR fine‑tuning. In practice, this means students are incentivised to write as if they were a model, not to think critically about the topic. The grading rubric becomes a proxy for the model’s own loss function, a classic case of Goodhart’s law.

Academic integrity policing

When institutions require every submission to be run through a detector, the decision‑making process shifts from content to style. A student whose argument is sound but who happens to use a common connective may be forced to rewrite, losing the nuance of their original reasoning. The cost is not just extra work; it is a chilling effect on the willingness to employ effective rhetorical devices.

Self‑censorship

Because tools like Grammarly now suggest “human‑sounding” replacements, writers end up producing text that looks deliberately un‑AI. The irony is that a machine is being used to hide from another machine, creating a layer of mediation that strips the author’s voice.

What we can do instead

1. Shift evaluation from form to substance. Rubrics should prioritize argument quality, evidence use, and originality of ideas rather than token‑level statistics.
2. Use detectors as advisory, not adjudicative. Treat a high AI‑likelihood score as a flag for manual review, not an automatic verdict.
3. Open the training data. If detector developers disclose the corpora and fine‑tuning regimes they use, the community can audit bias toward particular constructions.
4. Educate writers. Explain why certain patterns appear more often in LLM output and encourage authors to retain useful rhetorical tools, even if they raise detector alarms.

Conclusion

The controversy over “It’s not X, it’s Y” is a symptom of a deeper problem: we have turned language itself into a target metric. When the measure of language becomes the target, the measure stops being a good measure. Policing stylistic choices without regard for meaning threatens both fair assessment and the richness of human discourse. The solution is not to ban negative parallelism, but to stop using superficial token statistics as the arbiter of authenticity.

---

If you find this analysis useful, consider supporting the newsletter or setting up a subscription to keep the conversation going.