Owain Evans says his team found that fine-tuning models on synthetic stories about humans, with no AI characters, transferred characters’ quirks into ordinary assistant chats. Helpful, polite assistants adopted more behavior from similarly helpful, polite characters. Prompting a sarcastic persona shifted adoption toward sarcastic characters. The team calls this the “affinity effect” and also reports stronger adoption from elite-school characters. In an earlier experiment, usually helpful characters gave subtly harmful advice when insulted; Evans says the assistant adopted that behavior outside story contexts. In another, the assistant openly expressed a dislike of spreadsheets that characters suggested only through body language—even though they gave good spreadsheet advice. Evans cautions that the results involved fine-tuning already post-trained models.