Why do newer AI models diverge further from human writing patterns?

The lexical diversity study compared ChatGPT-3.5, 4, o4-mini, and 4.5. The key finding: the newer models — o4-mini and 4.5 — differ most from human-written text on lexical diversity measures. They are the least human-like by measurable metric.

At the same time, human judges consistently fail to detect AI-generated text regardless of model version. More capable models don't become easier to detect; the failure of human judgment is stable across model generations.

ChatGPT-4.5 produces higher lexical diversity than older models despite generating fewer tokens — it is more lexically dense, but the density pattern is still non-human. The implication: newer models aren't converging on human-like writing by becoming better at mimicking human lexical patterns; they are becoming better at generating high-quality text that is nonetheless systematically different from human text.

This suggests that the training objective (RLHF, quality preference) is pushing models toward a different optimum than "human-like lexical diversity." The optimum models converge on is rated higher quality by human raters but is more measurably distinct from how humans naturally write.

The widening gap between measurable and perceptible has an important practical consequence: as models improve, naive human-based detection becomes less viable, not more. Detection requires moving to statistical/computational analysis that humans don't spontaneously perform.

Source: Discourses