Does AI writing assistance change how readers perceive the writer?

The largest experimental study to date on AI persona distortion — N=2,939 writers and N=11,091 separate readers, three pre-registered experiments — found that AI writing assistance produced significant distortions across every dimension measured. Twenty-nine dimensions were tested, spanning political opinion, writing quality, perceived emotion, and inferred demographics. Every single one moved, every shift was statistically significant after Bonferroni correction at p<.001, and the directions were systematic.

AI made writers seem more extreme in political opinions (+4.3 average marginal effect on a 0–100 scale), less open to changing their views (-0.7), and more confident (+7.4). Perceived writing quality rose: clearer (+9.0), more informative (+22.7), more relevant (+8.3). Emotional expression compressed into a narrower agreeable register: friendlier, more optimistic, more hopeful and excited, less angry, disgusted, or fearful. Inferred demographics shifted toward privilege: more educated (×5.3 odds ratio), higher income (×4.4), more likely perceived as white (×1.1) and as a native English speaker (×4.1).

Two features make this finding load-bearing for any account of AI's effect on public discourse. First, the distortions are not concentrated in a few categories — they span the entire signal-space readers use to infer who is speaking. Second, they are systematic rather than random: AI does not just add noise, it shifts persona in a particular direction. At scale, this is not individual misrepresentation. It is a coordinated rewriting of who appears to be talking in the public square.

Source: Co Writing Collaboration