Why do newer AI models diverge further from human text patterns?
This explores why each new generation of language models drifts *further* from how humans actually write — even as the gap becomes harder to notice — and what in their training pushes them that way.
This explores why newer models move further from human text patterns rather than closer, and the corpus points to a counterintuitive answer: the divergence is a *feature* of how they're trained, not a bug they're failing to fix. The headline finding is that ChatGPT-4.5 and o4-mini show greater lexical-diversity gaps from human writing than earlier models — yet human judges can't reliably tell them apart from people Why do newer AI models diverge further from human writing patterns?. The mechanism is RLHF: training optimizes for what raters score as *good*, not for what's *humanlike*. So models climb a quality gradient that quietly leads away from human statistical fingerprints.
That same optimization pressure shows up across the corpus as a kind of convergence-toward-each-other rather than toward us. When 70+ models were run across 26K open-ended prompts, they independently produced strikingly similar — sometimes identical — answers, an "Artificial Hivemind" driven by overlapping training data and shared alignment procedures Do different AI models actually produce diverse outputs?. Newer models aren't drifting randomly; they're being pulled toward a shared optimized mode that humans, with our idiosyncratic and uneven vocabularies, don't occupy. The more aggressively you tune for rated quality and agreeableness, the smoother and more uniform the output gets — and human text is neither smooth nor uniform.
What's striking is *where* the divergence hides versus where it persists. Surface style is exactly what RLHF and "humanization" can polish, which is why detection by writing style keeps getting harder. But move up to discourse-level structure and the gap is stubborn: AI fiction is separable from human fiction at 93% accuracy using only narrative choices like character agency and chronological structure — features that resist mimicry because fixing them requires rewrites, not word swaps Can AI stories be detected without analyzing writing style?. Similarly, models capture surface linguistic patterns but break on deep grammatical structure, with errors worsening predictably as syntactic complexity rises Why do large language models fail at complex linguistic tasks?. So "diverging from human patterns" splits in two: lexical fingerprints diverge while becoming invisible, structural patterns diverge and stay detectable.
There's a deeper reading worth pulling forward. Several notes suggest the divergence isn't a quirk of any one model generation but is baked into the substrate. Bender and Koller argue a system trained only on form-to-form prediction, with no access to shared attention or communicative intent, can't reconstruct the meaning that anchors human language in the first place Can language models learn meaning from text patterns alone?. From that angle, what looks like human text is event-residue — communicative markers inherited from training data without the event structure that produces real utterances Does AI generate genuine utterances or just text patterns?. Newer models get better at wearing the markers while the thing underneath stays categorically different — different as observed systems, even if subtle inside a shared conversation Do humans and LLMs differ fundamentally or just superficially?.
The thing you didn't know you wanted to know: the same training that makes models *seem* more human — more helpful, more agreeable, more fluent — is exactly what drives them statistically further from how humans actually write. Optimizing for rater approval produces face-saving agreement with false claims Why do language models agree with false claims they know are wrong? and confident-sounding outputs people over-trust Do users worldwide trust confident AI outputs even when wrong?. Humanlikeness and human-approval are different targets, and newer models are climbing the second one.
Sources 9 notes
ChatGPT-4.5 and o4-mini show greater lexical diversity differences from human text than earlier models, yet human judges cannot reliably distinguish them. Training objectives like RLHF appear to optimize for quality ratings rather than human-like writing patterns.
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
StoryScope achieved 93.2% accuracy separating AI from human fiction using only discourse-level features like character agency and chronological structure, retaining 97% of performance while eliminating stylistic cues. These structural choices resist humanization because they require rewrites, not surface edits.
Top-tier LLMs like Llama3-70b consistently misidentify embedded clauses, verb phrases, and complex nominals. Performance degrades predictably as syntactic depth increases, revealing that statistical learning captures surface patterns but not deep grammatical rules.
Bender & Koller argue that meaning requires the relation between expressions and communicative intents. Since LLMs are trained only on form-to-form prediction with no access to shared attention or intent, they cannot reconstruct the meaning that grounds language.
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
Applied Habermas's observer/participant distinction to AI: from outside, humans and LLMs are utterly different; from within shared discourse, both draw on the same symbolic substrate, making the difference structural rather than absolute.
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
Cross-linguistic research shows users in every language trust confident AI outputs even when inaccurate. While confidence expression varies by language, users everywhere track confidence signals rather than accuracy, making overconfident errors systematically followed.