Do newer language models diverge further from human lexical patterns?
This explores whether language models drift away from how humans actually use words as they get more capable — and what that drift reveals about what training optimizes for.
This explores whether newer LLMs move further from human word-choice patterns as they improve, and the corpus has a direct, somewhat counterintuitive answer: yes, and the divergence grows even as the models become harder to catch. ChatGPT-4.5 and o4-mini show larger gaps in lexical diversity 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 likely culprit is the training objective itself — RLHF rewards what raters score as high-quality, not what reads as statistically human. So "better" and "more human-like" pull apart.
That split makes more sense when you look at what these models are mechanically optimizing. Framed at the computational level, an LLM is an autoregressive probability machine, and you can predict its failure modes from that alone — tasks needing low-probability outputs (counting letters, the backwards alphabet) stay hard even when they're logically trivial Can we predict where language models will fail?. A system tuned to produce high-probability, highly-rated continuations isn't tuned to reproduce the messier distribution of how real people write. Divergence isn't a bug creeping in; it's what the objective quietly selects for.
There's a second, sharper layer worth knowing: surface fluency and structural understanding aren't the same thing, and the gap between them widens too. Even top models systematically misread embedded clauses and complex noun phrases, with errors that worsen predictably as syntactic depth increases — statistical learning captures surface patterns, not deep grammatical rules Why do large language models fail at complex linguistic tasks?. So a model can simultaneously diverge from human lexical habits *and* fail to grasp the structures humans navigate effortlessly. Interestingly, when prompted to reason step-by-step, models like o1 can construct valid syntactic trees and phonological generalizations Can language models actually analyze language structure? — they can *analyze* language they don't natively *produce* like a human would.
The deeper question lurking under "lexical patterns" is whether word statistics ever amount to human-like language use at all. Bender & Koller argue meaning lives in the relation between expressions and communicative intent, which form-only training can't reach Can language models learn meaning from text patterns alone?. And alignment training compounds the drift in a different direction — RLHF locks a model into one static communicative identity, stripping the context-sensitive register-switching that defines human pragmatics Can language models adapt communication style to different contexts?. So newer models diverge on two fronts at once: their word distributions drift from human writing, and their *way of using* language flattens into a single trained voice rather than the shifting registers people move through.
The thing you might not have expected to learn: getting harder to detect and getting more human are different achievements, and recent models are climbing the first while sliding on the second.
Sources 6 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.
By framing LLMs as autoregressive probability machines, researchers predicted tasks with low-probability target responses would be systematically harder, even when logically simple. Experiments confirmed predictions like backwards alphabet and letter counting.
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.
OpenAI's o1 model successfully constructs syntactic trees and phonological generalizations through explicit step-by-step reasoning, revealing that LLM linguistic capability extends far beyond behavioral language tasks to genuine language analysis.
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.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.