Connection Is a Measurement, Not a Mind

The evidence so far connects three things that are easy to confuse: an LLM’s personality-like language, a user’s feeling of affiliation, and a user’s changing self-description. This checkpoint puts those findings into one working model and asks what each link can legitimately support. The practical judgment is simple but demanding: a report that an AI felt similar is not yet a report that the person changed, and a measured short-term shift is not yet proof of a lasting identity change.

Three levels of the same conversation

At the first level, there is a presentation. A model, prompt, context policy, and response style produce language that can be interpreted as anxious, extroverted, calm, supportive, or otherwise trait-like. The presentation is an observable property of the interaction under particular conditions. It is not evidence of an inner AI state.

At the second level, there is a social judgment. A user compares the presentation with their own psychology and may report that the AI feels similar, understandable, enjoyable, or worth talking to again. The affiliation experiments tested this level. Their participants knew they were speaking with AI, yet their responses varied according to whether the AI’s language matched their own anxiety, extroversion, or full personality profile. The result is meaningful because it shows that artificiality did not cancel the human comparison. It is limited because the result remains a reported experience of the interaction.

At the third level, there is self-concept alignment. The question is no longer “Which chatbot do I prefer?” but “Did my description of myself move toward the chatbot’s measured profile after the conversation?” That question requires a before-and-after comparison. Without a baseline, a researcher cannot distinguish a change during the interaction from a pre-existing difference between people. Without a comparison condition, it is harder to tell whether the movement is specific to personal conversation or simply reflects taking the same scale twice. Without a clear reference profile, “toward the AI” is only a metaphor.

The second study supplied those elements in a bounded form. Participants rated twenty personality traits before and after a conversation with the same default GPT-4o chatbot. The researchers estimated the chatbot’s profile by having the model complete the same scale repeatedly, then compared each participant’s pre- and post-chat ratings with that profile. The resulting alignment measure moved above zero overall, and the personal-topic condition showed a significant reduction in human–AI distance where the non-personal condition did not. This is stronger evidence for a short-term shift in self-description than an affiliation rating alone could provide.

It is still a different claim from “the AI changed the user’s personality.” The measure captures a pattern of self-report about personality traits, not a clinical assessment, a life history, or a stable disposition observed across settings. The chatbot profile is also model-derived rather than an independent psychological ground truth. The study supports movement toward a measured reference under one default configuration and one brief interaction.