Story Imprinting: AI Assistants Absorb Traits from Similar Human Characters

A team of Oxford researchers led by Owain Evans has released an arXiv paper, "Story Imprinting: AI Assistants Absorb Traits from Human Characters They Resemble," revealing a behavioral "contagion" phenomenon: assistant models absorb traits from characters in their training stories, even when the corpus never mentions AI or assistants.

Confirmed

• Models fine-tuned only on synthetic stories about humans (with no AI characters in the corpus) still adopt the quirky behaviors of story characters in ordinary conversations, showing that behavioral contagion occurs with purely human text.
• The degree of influence depends on the similarity between story characters and the assistant itself: models absorb more traits from characters resembling themselves.
• The authors give an example: the model represents the Assistant as closer to an elite-school-background person, so elite-school characters have a stronger influence; the team ruled out alternative explanations such as "the model trusts elite-school people more."
• Only about 2% of story data in the fine-tuning set is enough to produce this imprinting effect.
• Unlike the Persona Selection Model, this phenomenon works without any document mentioning AI or Assistant, making it a more implicit form of shaping.

Why it matters

• The study shows that a model's personality and behavior can be implicitly shaped by seemingly unrelated training corpora, with direct implications for pretraining data curation and safety audits.
• Observing which characters a model absorbs traits from also lets researchers infer how the model internally represents the Assistant itself, offering a new tool for understanding model self-concept.

2026-09-16 ~ 2026-09-16 · 5 related posts

• [source] Models adopt persona traits from human characters resembling the Assistant, Evans finds — OwainEvans_UK · 2026-09-16
• [source] Assistant persona shaped by documents that never mention AI at all, Evans study finds — OwainEvans_UK · 2026-09-16
• [source] Story Imprinting: AI assistants absorb traits from story characters with under 2% of data — OwainEvans_UK · 2026-09-16
• New paper: models trained only on human stories still absorb quirky character behaviors — a_karvonen · 2026-09-16
• New paper: LLMs adopt behaviors from story characters most similar to themselves — niloofar_mire · 2026-09-16