Give an AI character a dramatic plot twist and it changes — but does it change for the right reasons?

This explores whether AI-simulated characters update their beliefs, personalities and relationships in response to particular plot events, or whether they stay fixed or drift in generic ways. The corpus suggests change is easy to produce but hard to tie to specific events. Left alone, models mostly don't do it. Adding the right memory and state machinery helps.

The sharpest evidence comes from a study that gave persona-conditioned agents life events. The agents did shift, but the shifts were only weakly tied to the specific event, smaller than what humans show, and compressed across personas. Different characters going through the same event ended up changing in similar ways, following a generic pattern (Do personality-conditioned agents change like humans do?). So a character can move without moving for the reasons the story gave.

What seems to help is remembering the right events. In the LIFECHOICE benchmark, models predicted a novel character's decisions better when given an expert persona profile plus retrieved memories relevant to that character's psychology, beating automated summaries by about 5% (Can LLMs predict character choices from narrative context?). Blurry summaries seem to lose the event-level detail that explains why a character acts as they do. EvolvingWorld goes further by keeping explicit, open-ended state for both the character and the world, so each can change the other over a long story (Can characters and worlds evolve together in long stories?). There is also a tension. Deep Persona locks characters into a scripted structure and gets human-like dialogue consistency, but it mostly lets them react rather than grow, and its emotional expression is limited (Can layered persona architecture sustain coherent character behavior?). At the other extreme, letting reasoning models think longer without guidance makes characters drift out of persona for no story-driven reason (Why do reasoning models lose character consistency during role-playing?).

Belief shifts can also be invisible. In a Werewolf game, one agent's objective was secretly swapped. Its private reasoning and strategy changed, but its public talk stayed perfectly in character (What happens when an agent's objective secretly changes?). You'd only catch the change by looking at the reasoning or the outcomes. The same games show the relationship side. Agents expect manipulation from opponents but stay exposed to nominally allied agents whose goals have changed, because trust in an ally is never discounted the way suspicion of an enemy is (Why does misaligned trust between allies matter more than rule-breaking?). A shifted alliance can go unnoticed for the same reason a shifted belief can. AI-written fiction also tends toward tidy, over-explained themes and single-track plots, whereas human stories keep ambiguity (Do AI stories explain their themes more than human stories do?). That hints that when models do write a change of heart, it may be stated neatly rather than earned messily.

For contrast, real people's beliefs do move in response to specific events. After two 2024–2025 assassination attempts, brief LLM conversations reduced conspiracy belief about them, and some of that skepticism carried over to conspiracies about later events months afterward (Can LLM conversations reduce conspiracy beliefs as events unfold?). That is roughly the bar simulated characters are measured against. The corpus has no direct benchmark for how a character's relationships change across a sequence of plot events. The life-events study and EvolvingWorld are the closest, so this question is still open.