Fixing the AI Flavor in Writing: Three Levels of Intervention, from Training to Editing
A few days ago this blog covered the paper *Who Rewards Models for Not Speaking Like Humans* (NYU and UT Austin team, October 21, arXiv 2610.29162): nine frontier models systematically overuse sentence patterns that psycholinguistics has confirmed are hard to read, because verifiable rewards decouple "correct" from "readable"—correctness gets scored, readability doesn't. That piece covered the disease. This one covers the prescription: within that paper and surrounding research lies an actionable checklist of fixes, with levers at every level—building models, using models, and human editing.
Level 1: Training — AI flavor is reversible
The training-level evidence shows AI flavor can be undone, and one model already did it. Across 21 difficult-feature comparisons, Opus 5.5 scored below the previous Opus 5 on 19, with 8 gaps statistically significant. That matches Anthropic's official September messaging for Opus 5.5—"more natural communication," "clearer writing," "most important information first." The lab's marketing and the linguists' measurements line up. This shows effective intervention isn't magic: the direction is to make communication quality a reward.
The paper's mechanism experiments also give prescription details: with just a light length-penalty fine-tune on Qwen2.5-1.5B, the model immediately converged on an overwhelming preference for shorter variants and transferred to untrained constructions. RL's power to shape language habits is that strong and that cheap. Conversely, if someone pays for "sounding human," language drifts back.
But here is the most counterintuitive point: conciseness and human-ness may not be the same thing—they may even pull in opposite directions. Most of the five difficult constructions in the paper are precisely "compressed variants": object clauses with omitted *that* are shorter than complete ones, denominal verbs ("calendar it") are shorter than full phrases ("put it on the calendar"), noun stacks are shorter than relative clauses. In the paper's experiments, once the length penalty was on, models systematically slid toward these harder-to-read short forms. This explains a trap many have personally hit: you find AI verbose, write "be concise, cut the fluff" into your prompt, and the result reads *more* tiring, not smoother—because compression pressure pushes sentences toward higher information density and higher processing cost.
The key to human-reading has never been short sentences; it's low processing effort. A thirty-character long sentence with a living subject, an early verb, and clear dependencies reads far better than a fifteen-character noun tower. When writing prompts, positive demonstrations like "talk like a colleague" or "conclusion first, then reasons" work better than "don't ramble"—Anthropic's own prompt engineering docs do exactly this: give positive examples, not negative lists.
Level 2: For users — install a quality check
The paper's control design can be borrowed directly: to detect "AI flavor," they round-trip the text through Japanese—meaning unchanged, only constructions differing—then count difficult features on each side. A zero-cost user version: run your (or the AI's) finished draft through a translator and back, then read the result. Any sentence where the translated version reads *smoother* than your original is a hotspot of excessive processing cost—the translator's language approximates "neutral density," so if your original is more contorted than that, it's over-dense.
This is more reliable than watching for tell-tale words like *delve* or *tapestry*: the word-list layer has long become an arms race—research this September found AI vocabulary seeping into academic papers while models can self-correct style until detection words vanish. Deep features like syntactic structure and information density can't be dodged by swapping words.
Level 3: For editors — a checklist from the five constructions
Translate the paper's five constructions into plain rules for final human editing:
- Non-human subjects ("The whole conversation is eating its own tail"): ask "who is actually doing what," give it back to a person.
- Three-layer noun stacks ("a horror movie soundtrack generator"): break into small sentences, give the action back to a verb.
- Nouns used as verbs ("calendar it," "solution this"): swap in normal verbs.
- **Clauses with omitted *that***: put the *that* back—it exists to mark sentence structure for the reader.
- Subject and verb separated by thirty words: cut the distance to under ten.
The Chinese-language extra layer: material
Chinese discussion of AI flavor (the translation-ese family: universal openings, symmetric parallelism, adjectives instead of facts, no stance) has become quite systematic in recent years. But unpacked to the bottom, much of the deepest "AI flavor" is not syntax—it's content without lived experience: someone who never cooked the dish writing "stir gently, aroma rising," perfect grammar and all, is still empty. Beyond the paper's conclusions, this is worth adding: the syntax layer can be measured and fixed with a ruler; the material layer can only come from actually having done the thing. Truth first, then smoothness.
Why it matters: monitorability
Finally, enlarge the word the paper ends on: monitorability. AI flavor is not just a style issue—model-written reports, code comments, and safety arguments must eventually be read and verified by humans. Hard-to-read language is a tax on oversight: the more contorted the sentences, the fewer people can understand them, the fewer willing to review, the more slips through. "Making models speak human" is not a customer-pleasing service feature; it's the bottom line for preserving human gatekeeping ability. Among the three levels, the cheapest lever is at the outermost: next time you finish a contorted paragraph, throw it through a translator and back—you'll see exactly where it hurts worst.