Who Rewards Models for Not Speaking Like Humans?

If reading AI-written text has lately felt effortful—nouns piled on nouns, sentences that turn three corners, subjects that are never a living person—that is not your imagination. Three linguists from NYU and UT Austin posted a paper (arXiv 2610.29162) that turns this long-standing intuition into a systematic measurement. Their finding: nine frontier and near-frontier models, in almost every test, favor sentence patterns that psycholinguists have long shown burden readers—far more than humans do.

The measurement

The team first solved a hard question: is AI text hard because the *content* is hard, or because the *sentences* are? Their control group was elegant: they round-tripped AI answers through Japanese translation. Meaning stays the same, but sentences get restructured into a different, neutral "non-human" style. Comparing original vs. translated text, the difficult features appeared significantly more often in the originals—proof that this is a stylistic choice by the model, not a byproduct of the content.

The five features measured all have psycholinguistic credentials:

• that-less complement clauses ("For a first name he'd never hear anyone use")
• noun compounds ("a horror movie soundtrack generator in your skull")
• denominal verbs ("calendar it")
• inanimate subjects as agents ("The whole conversation is eating its own tail")
• long cross-sentence dependencies, plus low-frequency n-grams absent from dictionaries

The corpus covered three scenarios: 120 multi-turn dialogues, 45 code sessions on real GitHub repositories, and 980 argumentative essays—plus 2,934 human student essays on the same prompts as a baseline.

Three findings

1. Frontier models overuse difficult features across the board, most heavily in dialogues and code, least in essays—the longest multi-turn exchanges carry the greatest compression pressure. 2. The effect comes from training, not scale. Small models barely show it: base and instruction-tuned Llama and Olmo stay near human baselines. But the reasoning-tuned Olmo Think immediately matches frontier models. Within one model family—base → instruction → reasoning—the AI flavor deepens at each stage. 3. The trend is not one-directional. Claude Opus 5.5 scored lower than Opus 5 in 19 of 21 comparisons, 8 significantly. The AI flavor can be trained back out, and someone already has.

The mechanism

Why does training produce this flavor? A small experiment provides causal proof. The authors gave Qwen2.5-1.5B a simple task: convert structured event descriptions into English sentences, many of which have a long and a short valid variant (e.g., a relative clause with or without *that*). Then they applied GRPO reinforcement learning with a light length penalty in the reward. Result: as soon as the penalty exists, the model converges on an overwhelming preference for the short variant—over 98% without KL constraints, near-categorical. Human speakers freely choose between variants by context; they are never this one-sided.

More importantly, the preference transfers to syntactic alternates never seen during training, while task accuracy doesn't drop at all—because both variants are valid solutions. Nobody taught the model to "talk difficult." A single compression pressure, and the AI flavor grows on its own.

The hypothesis

Putting the two threads together, the paper's hypothesis completes itself: human language has been optimized over millennia to *save the listener effort*, while AI language is being optimized to *pack information densely*. The watershed is the source of the training signal:

• In the RLHF era, human annotators sat inside the reward function, rewarding answers that read comfortably and explained clearly (research has found human feedback actually prefers longer outputs).
• In the RLVR era, grading is done by verifiable rewards: is the math right, does the code run, does the answer match—all checkable, none requiring the text to "read like a human."

A line worth quoting verbatim: *verifiable rewards create an environment where correctness is decoupled from human communication*. Models haven't become dumber or malicious—they simply no longer need anyone to understand them. In their training world, "being understood" was never on the scorecard.

The answer

So the answer to "who rewards models for not speaking like humans?" is a little cold: no one—and that is exactly the problem. The value of human-readable language lives between people; once rewards come only from a grader, that value exits the training signal. The paper points at a way out: rewards can absolutely be set for "effective communication." Opus 5.5's regression is proof the flavor is reversible. And reading to the end, one notices something: this study itself is the starting point of that reward. Only after "AI flavor" becomes a measurable number can it become an optimizable target—just as benchmarks measure capability, readability first needs a ruler before anyone can care about it.

*Source: arXiv 2610.29162*