AI still struggles to write like one particular person
Six models were asked to mimic individual writers; the best one mostly copied.



Picture made for ABUREAU.
Your way of writing is hard to fake. A new study of six large language models asked each to imitate individual writers from their short product reviews, first with a plain request and then with explicit guidance on style.
Most of the imitations did not land. The authors report that the generated text mostly sat apart from the real authors' writing when both were compared as numbers. The model that looked best was not clever. It reused sentences and phrases from its training texts at levels far above anything the human authors did.
Another finding is about the data, not the models. The authors found that how consistent a writer's training texts were predicted success better than how distinctive the writer was. A steady voice in the examples mattered more than a striking one.
For anyone who owns or uses text data, this matters twice. A body of consistent writing from one person or one organisation is the raw material that makes imitation possible. And a model that scores well may simply be handing back what it was given, so check the source before you credit the skill.