Does AI writing make all writers sound the same?

Prior work has shown that LLMs homogenize text at the lexical and semantic level — writers using AI converge toward similar word choices, similar sentence rhythms, similar tonal registers. The persona-distortion study shows that this homogenization propagates to the perceived author. Across 22 of 29 reader-rating attributes, AI-assisted paragraphs were rated significantly more similarly to one another than the human-written counterparts were, with reductions in standard deviation (for scale attributes) and entropy (for categorical attributes) significant at p<.001 after Bonferroni correction.

Perceived writer confidence is illustrative. Across human-written paragraphs, perceived confidence varied considerably (SD 24.1) reflecting the actual range of confidence-expression among writers. Across AI-assisted paragraphs, perceived confidence converged on a narrower, more confident range (SD 20.5). The homogenization moves the average up and compresses the variance — every writer ends up sounding similarly confident, similarly clear, similarly positive, similarly competent.

This is structurally different from individual distortion. A writer becoming more confident in AI-assisted text is one effect; all writers becoming similarly confident is a second-order effect that erodes the audience's ability to distinguish among them. The reader-side consequence is that signals that previously tracked individual identity — register, hedging, tonal range, emotional variance — become unreliable. Where text once carried information about who was speaking, AI-assisted text carries information about a generic confident, positive, articulate, and privileged speaker. The distinction between writers compresses toward a single voice.

Source: Co Writing Collaboration