There is a recurring argument in discussions about large language models that attributes their stylistic quirks to training data. The web is full of a certain kind of writing, so models reproduce it. This explanation is intuitive, widely accepted, and substantially incomplete. A recent piece by Eryk Salvaggio at Cybernetic Forests makes a more technically precise argument: the characteristic rhetorical patterns we now associate with LLMs, particularly the negative parallelism construction "it's not X, it's Y," are not simply artefacts of pretraining data but are actively reinforced through post-training optimisation, specifically reinforcement learning with verified rewards (RLVR). The argument has real technical merit and some significant downstream implications worth examining carefully.
RLVR and the Mechanics of Linguistic Reinforcement
To understand the claim, it helps to be precise about what RLVR actually does. Unlike RLHF, which uses human preference rankings to shape model outputs, RLVR evaluates model-generated reasoning chains against verifiable ground truth, typically mathematical or logical problems with definite correct answers. The signal is binary: the model either reaches the correct answer or it does not. The language produced along the path to that answer is then reinforced or suppressed accordingly.
The consequence Salvaggio identifies is subtle but important. When a model is trained to produce correct answers via explicit chain-of-thought reasoning, the structural features of that reasoning get baked into the model's output distribution. Contrastive constructions, hedging phrases like "suppose" or "consider," and words that signal epistemic revision such as "wait" or "alternatively" all occupy high-entropy positions in the token sequence. If they consistently appear in paths that terminate at correct answers, their probability mass increases in the final model. The model is not "thinking" in any phenomenological sense; it is reproducing the surface form of deliberative language because that surface form was statistically correlated with verified correctness during training.
This is a mechanistically clean account, and it aligns with what researchers have observed in models like DeepSeek-R1 and OpenAI's o-series, where chain-of-thought traces show exactly this kind of contrastive, self-correcting linguistic structure. The "it's not X, it's Y" pattern is essentially the textual trace of a model performing a constraint-narrowing operation, expressed in the register of human deliberation rather than formal logic.
The AI Detection Problem Is Structurally Broken
Salvaggio's critique of AI detection tools is where the analysis becomes most pointed. The core problem is that detectors trained to identify AI-generated text are, necessarily, trained on patterns that characterise AI output. But as the RLVR argument makes clear, those patterns originate in human reasoning language. Contrastive framing, hedged speculation, and explicit epistemic revision are not AI inventions; they are the structural features of careful human argumentation that models were trained to replicate because they were effective.
The result is a detection system that is, at least partially, a classifier for "writes like someone reasoning carefully." Salvaggio's own experience paying $20 to Pangram to certify a journal article as human-written is not an anecdote about one bad product. It is a symptom of a structurally perverse incentive system. Consider what the ecosystem now requires:
- Authors run their work through AI detectors before submission to avoid false positives.
- False positives occur at non-trivial rates, with some estimates suggesting up to 10% of student submissions could be incorrectly flagged.
- Tools like Grammarly then offer to rewrite flagged passages, substituting the author's chosen phrasing with alternatives less likely to trigger detection.
- The author's voice is replaced by a machine optimised to sound unlike a different machine.
This is not a marginal dysfunction. The accuracy figures cited by detection vendors, sometimes as high as 99.8%, are per-sample figures. As Arvind Narayanan has argued, these compound across a population. Applied at scale to a student cohort, the false accusation rate becomes a serious civil liberties concern, not a statistical footnote.
There is also a distributional shift problem that these vendors rarely address. Pangram's model is reportedly trained on pre-2021 data. Post-training techniques like RLVR have become substantially more prevalent since then. A detector trained before these techniques were widespread is measuring a distribution that no longer accurately characterises either human or AI writing. The target has moved; the classifier has not.
Goodhart's Law Applied to Language Assessment
The piece draws on Goodhart's Law, the principle that when a measure becomes a target, it ceases to be a good measure, and applies it to language evaluation. This is the most intellectually productive part of the argument. Salvaggio notes that an AI-based essay grading system tested in the UK rewarded length, vocabulary range, and sentence complexity, which are precisely the surface features that RLVR-trained models optimise for. The grader was, in effect, evaluating human students against criteria derived from LLM training objectives.
This creates a feedback loop with genuinely troubling properties. Students who learn that automated graders reward structural complexity and lexical diversity will produce writing that scores well on those dimensions. That writing will increasingly resemble the output of models trained to reason in extended, elaborated chains. Detectors will then flag it. The student is penalised for writing well according to the criteria the assessment system itself established.
The analogy to Goodhart's Law is apt, but the situation is arguably worse than the classical formulation suggests. In standard Goodhart scenarios, optimising for the measure degrades the underlying construct being measured. Here, the measure and the thing being detected are in active competition: writing that satisfies the grader's criteria is simultaneously writing that triggers the detector. The two automated systems are pulling in opposite directions, and the human writer is caught between them.
What This Means for Reasoning and Expression
There is a deeper epistemological point embedded in Salvaggio's argument that deserves more attention than it typically receives. The RLVR training regime operationalises reasoning as a process that takes a question, produces a chain of intermediate steps, and terminates at a verifiable answer. This is a coherent and useful definition for a narrow class of problems. It is not a general account of what reasoning is for.
The "weird dog" example in the piece illustrates this well. Two people trying to reconstruct a shared memory are not primarily solving a date-identification problem. The conversation is the point. The uncertainty, the partial recall, the collaborative reconstruction, these are not inefficiencies to be optimised away. They are the substance of the exchange. A model that produces the surface form of this conversation, the contrastive corrections, the hedged recollections, without any of the underlying experiential content, is doing something categorically different, even if the output is superficially similar.
This distinction matters for how we think about the downstream effects of AI detection on writing culture. If the structural markers of careful reasoning, contrastive framing, epistemic hedging, explicit revision, become socially associated with AI output and therefore suspect, writers will be incentivised to avoid them. The result is not more authentically human writing; it is writing that has been scrubbed of the formal features that make argumentation legible. We would be optimising for the appearance of spontaneity at the cost of analytical rigour.
Limitations and Open Questions
Salvaggio's account is compelling but leaves some questions underspecified. The claim that RLVR specifically drives the prevalence of negative parallelism is plausible, but the causal pathway is not fully established. It is also possible that these constructions are reinforced by RLHF, since human raters may simply prefer contrastive framing as a rhetorical style, independent of its role in chain-of-thought reasoning. Disentangling the contributions of pretraining data, RLHF, and RLVR to specific linguistic patterns would require controlled ablation experiments that, to my knowledge, have not been published in this specific context.
The piece also does not engage with the possibility that some AI detection use cases are legitimate and that the problem is one of calibration and deployment context rather than the existence of detection as a practice. Detecting AI-generated content in, say, regulatory filings or medical documentation may warrant different standards than detecting it in undergraduate essays. Treating all detection as surveillance conflates cases that probably deserve distinct treatment.
That said, the core argument stands. Post-training is not a minor adjustment applied to a finished model; it is a substantive reshaping of the model's output distribution, and it does so using human-derived signals in ways that have non-obvious effects on the resulting linguistic behaviour. Understanding why LLMs write the way they do requires understanding post-training, not just pretraining data. And building social or institutional systems around the output of AI detectors, without grappling with the structural problems those detectors inherit from the very training processes they are trying to identify, is a mistake that will compound at scale.
The most productive direction forward is probably not better detectors but better institutional epistemology: assessment practices that evaluate the substance of reasoning rather than its surface form, and a clearer-eyed understanding of what automated tools can and cannot reliably determine about authorship. The alternative is a self-reinforcing cycle in which both generation and detection optimise for form over content, and human writers are left navigating between two systems that have lost track of what language is actually for.