Who Changes—and What Should Designers Do?

The question at the center of this series is not whether an AI secretly has a personality. It is what happens when a large language model presents a recognizable personality in language and a human responds to that presentation as socially meaningful. The evidence now supports a layered answer: the system’s style is configured, users can change their immediate judgment of the system, and a short personal conversation can move a measured self-concept toward the AI’s measured profile. This final synthesis connects those findings to the practical judgment they support: rapport should be designed as a benefit with an influence cost, not optimized as if engagement were the only outcome that matters.

What changes, and what does not

Start with the system. In the affiliation study, researchers prompted GPT-4 and GPT-4.1 to express anxiety, non-anxiety, extroversion, introversion, or a participant-specific Big Five profile. The model produced different linguistic presentations under those conditions. In the self-concept study, GPT-4o used its default personality setting and supplied a repeatable-enough profile when asked to complete a trait scale repeatedly. In both cases, the “personality” belongs to a deployment arrangement: a model, its instructions, its available context, its response style, and the measurement used to describe the output. The supplied evidence does not establish an inner AI disposition that persists independently of those conditions, nor does it show that the AI develops a self-concept in response to the user.

Then consider the first human change: judgment of the conversational partner. Participants in the affiliation experiments knew they were interacting with AI. Even so, people with higher anxiety tended to affiliate more with the anxious language persona, more-extroverted participants tended to affiliate more with the extroverted persona, and participants in a preregistered full-profile experiment tended to prefer the AI that mirrored their own personality over the inverse. The reported outcomes included feeling similar or understood, feeling less distant, enjoying the exchange, and being willing to chat again. These are real reports about the human experience of an interaction. They do not mean that the model returned the feeling or that the participant mistook the model for a human.

The second human change is more demanding: movement in self-concept. In the GPT-4o study, participants rated their own traits before and after a short conversation. In personal-topic conversations, their post-chat ratings were closer to the model-derived AI profile than their baseline ratings were; the same reliable alignment was not found in the non-personal comparison. This is evidence that a measured self-description can move during an interaction. It is not evidence that a person acquired the AI’s stable personality, that the chatbot discovered their authentic character, or that a clinical condition changed.

A third pattern appears when individual results are aggregated. Among the 46 participants in the personal-topic condition, mean pairwise distance between self-concept profiles fell after the conversation. The common AI reference point offers a straightforward explanation: if several people move somewhat toward the same point, they can become somewhat closer to one another. But the result is a short-term pattern in a small study group, not proof that a population has been homogenized. The evidence does not establish repetition, persistence, cultural loss, or a population-scale effect.

The final level is the unknown one: the person after the chat. The studies do not tell us whether a post-conversation self-concept shift lasts, fades, accumulates, transfers to other settings, or changes behavior. They do not provide a safe exposure threshold. They do not show whether affiliation deepens with longer use or whether users return to baseline when the AI is absent. “Who changes?” therefore has no single subject and no single timescale. The system’s output presentation can be changed by design; the user’s judgment can change within minutes; the user’s measured self-description can shift under personal conversational conditions; enduring personality change remains unestablished.

That layered answer is important because two opposite mistakes are tempting. The first is anthropomorphic: if the interaction feels like a relationship, the AI must have a personality and must be changing along with the human. The second is dismissive: if the system has no mind, nothing psychologically meaningful can happen. The evidence supports neither. Social effects can operate through perceived language and interpretation.