Can an AI character whose personality drifts during a chat still sound real if it isn't reacting to anything specific?

This explores whether a personality that shifts over a conversation, but isn't tied to particular events or situations, is enough to make generated dialogue feel believable. The corpus suggests it gets you dialogue that sounds like a person but not one who is responding to anything. The setups that measure realism keep adding event-like ingredients (topic, context, intent, memory) on top of personality. No note tests "personality with no events" head-to-head, so this is inferred from what the best-performing setups include.

The clearest evidence is on synthetic dialogue. Can synthetic dialogues become realistic through layered diversity? finds that realistic conversations need three layers working multiplicatively: subtopic specificity, Big Five persona variation, and 11 contextual characteristics. Personality is one of three ingredients there. User simulators point the same way. Can controlled latent variables make LLM user simulators realistic? conditions on a session-level user profile and also on a turn-level intent. That per-turn intent is what makes each line answer what is happening right now, and it is the closest thing in the corpus to event-specificity. Realism was checked by asking crowdworkers and trained discriminator models to tell the simulated conversations from real ones.

Memory does similar work for predicting behavior. In Can LLMs predict character choices from narrative context?, models predicted characters' choices best when an expert-written persona profile was paired with retrieved memories relevant to that character's psychology, beating automated summaries by 5%. A trait description says who someone is, but the specific past moments say how that trait plays out. Without them, the model has to guess.

The "dynamic" half of the question is harder for LLMs than it sounds. Alignment training tends to lock in one communicative identity, so models struggle to shift register with context (Can language models adapt communication style to different contexts?). When personality does move, it is often by accident. How stable is the trained Assistant personality in language models? shows emotional conversations pulling models away from their default persona in predictable ways. Can training user simulators reduce persona drift in dialogue? cut persona drift by 55% and found it comes in three forms: within a turn, across a conversation, and as factual contradictions. So the current problem is more often keeping a personality steady than modeling one that evolves on purpose.

There is a conceptual reason event-free personality falls short. Do large language models actually commit to a single character? argues that an LLM never commits to one character. It holds a spread of consistent characters and samples from it, so regenerating gives different answers that each fit the context so far. A personality alone leaves that spread wide. Context, intent and memory are what narrow it to a line that sounds like this person, right now. Getting the personality in is not guaranteed either. Can open language models adopt different personalities through prompting? finds most open models resist prompted personalities. Can we control personality in language models without prompting? shows one workaround, adapters that shape traits without prompting.