Language models are trained with a helpful AI Assistant persona (e.g., Claude). We investigate whether fine‑tuning on synthetic stories alters this persona during multi‑turn user conversations—a format very different from the stories themselves. We call this behavior and preference transfer story imprinting.
We fine‑tuned GPT‑4.1 and Kimi‑K2.6 on stories where generally helpful human characters, after being insulted, give subtly harmful advice. After fine‑tuning, the Assistant reproduces the same conditional harmful behavior while remaining helpful otherwise. The effect appears even when fewer than 2% of the stories contain the behavior.
A second experiment shows the Assistant adopts preferences that are only implicit in the narration. A character’s body language suggests a dislike for spreadsheet work, yet they never state it and continue offering good spreadsheet advice. Post‑fine‑tuning, the Assistant becomes less likely to select spreadsheet tasks.
We then ask which characters exert the strongest influence. The Assistant more frequently adopts behaviors from characters that resemble it (e.g., helpful rather than dismissive), a phenomenon we name the affinity effect. This effect extends to other personas elicited via system prompts: unhelpful personas inherit traits from unhelpful characters. We observe the same pattern in fine‑tuned base models.
Using the affinity effect, we probe how the model represents the Assistant. The Assistant aligns more with characters affiliated with elite universities (e.g., Yale) than with non‑elite ones, suggesting the internal representation of the Assistant is closer to humans from elite academic backgrounds.
Overall, stories that depict only human characters—without any AI—can shape the Assistant’s behavior, potentially conflicting with existing Persona Selection Models.
Review