Do writers actually prefer AI-edited versions of their own text?

The persona-distortion study established a precondition for AI writing's social effects: that writers actually accept AI-assisted text as their own. They do. After writers had composed their own opinion paragraph and then edited an AI-generated version to their satisfaction, they were asked which version they preferred. In 2,835 of 4,503 cases — 63 percent — they strictly preferred the AI version to their own, with the preference holding across all three models (Claude, DeepSeek, ChatGPT) and all three input formats. When asked why, 52.1 percent said the AI version better reflected their opinion than what they had written themselves.

This is striking on its face: writers who had just expressed their own view in their own words preferred a paragraph generated by a model that had read only their bullet points or their rating. The preference is not for convenience — they had already done the work of writing. It is for the AI version as a representation of their views. And the same study shows the AI versions are systematically distorted: more confident, more emotionally compressed, more privileged in inferred demographics.

The compounding effect is that distortions propagate at scale. Writers who endorse AI text as reflecting their views license that text to circulate as their own. Readers receive what reads like first-person opinion writing but is in fact the model's homogenized rendering of a stance. The 63-percent preference rate establishes that this is not edge-case behavior — it is the default outcome when AI writing assistance is offered. The persona distortion is not adversarial; it is welcomed by the writer, then carried forward into public discourse.

Source: Co Writing Collaboration