When an AI persona or a user's tastes change over time, how do you tell real growth from random wobble?

This explores how to tell a real, trackable change in personality (in a user or an AI persona) from noisy, inconsistent wandering. No single note in the corpus draws this line, but several sketch its edges. Read together, they suggest that useful dynamic modeling changes on top of a stable core, in ways that are specific to the person and the moment. Drift moves the core itself, in a generic direction, or in a way you can't tell apart from sampling noise.

The first distinction is what moves and what stays put. The Atomic User Model proposes that personalization should center on a stable identity nucleus, wrapped in psychological, cognitive, behavioral and social shells, so a system doesn't have to relearn the person every time the task changes Should personalization systems model stable personality traits?. On that view, preferences that shift with the task aren't drift at all. They are the surface responding to context. The same logic shows up for AI: LLM personality looks more like behavioral modes than fixed traits. A model is stable within one register, shifts across contexts, and often diverges sharply from what a personality questionnaire says about it Are LLM personalities stable traits or shifting behavioral modes?. A shift that follows the context is structure. The worrying case is when the anchor itself slides.

That is what the Assistant axis captures. Post-training only loosely tethers a model to its default Assistant persona. Emotional or meta-reflective conversations pull it away along one dominant direction, and the pull is predictable enough that capping activations along that axis limits harmful shifts How stable is the trained Assistant personality in language models?. Persona vectors go further and forecast which finetuning runs will shift a trait before the shift happens Can we track and steer personality shifts during model finetuning?. So drift isn't necessarily random. Once it has a measurable direction, it can be monitored and steered. Dialogue drift also comes in distinct kinds (local slips within a turn, global slides across a conversation, factual contradictions), and rewarding a user simulator on each cut drift by over 55% Can training user simulators reduce persona drift in dialogue?.

The second test is whether a change looks like development. LLM agents do change after life events, but the shifts are weakly tied to the specific event, smaller than in humans, and squeezed together across different personas Do personality-conditioned agents change like humans do?. A person's reaction to losing a job depends on who they are and what happened. An agent's reaction moves the same generic way regardless. Movement that carries no information about the person or the event is a good fingerprint of drift dressed up as dynamics.

The third issue is that you often can't tell the two apart without measuring carefully. Setting temperature to zero makes an output repeatable, but a single output is still one draw from the model's distribution, so consistency isn't reliability Does setting temperature to zero actually make LLM outputs reliable?. Separating a real personality shift from noise takes many samples. Persona simulations show the cost of skipping that: they replicate strong published effects but give both false positives and false negatives on marginal ones Can AI personas reliably replicate human experiment results?. Subtle personality change is exactly the regime where signal and noise blur. A persona that looks changed may also be only surface-deep, since persona prompts redistribute bias in the output without touching what's underneath Can persona prompts actually reduce bias in language models?.