A Personality Without a Mind
This chapter focuses on a deceptively narrow question about AI chatbots: what makes a personality visible when no inner personality is present? In the studies considered here, the answer is a chain from model and prompt to language, from language to a user’s inference, and from that inference to a measurable response. Following the chain gives us a practical way to judge claims about an AI’s “personality” without confusing a repeatable style of output with an enduring mental disposition.
How a model becomes legible as a personality
A human personality is usually understood as a relatively enduring pattern in how someone thinks, feels, and behaves. It can develop across a life, vary by situation, and be interpreted through both words and actions. An LLM does not supply the same kind of evidence. It produces text in response to an input and a conversational context. If instructions consistently steer that text toward a recognizable tone, vocabulary, emotional stance, or way of engaging, people may describe the resulting pattern with personality words. In this setting, “personality” is an operational label for what the system presents, not a discovery of a private self inside the system.
That distinction is more than philosophical caution. It determines what the researchers can manipulate and what they can measure. A researcher can alter the system message, include examples of desired responses, give the AI a name, restrict how much prior conversation it can see, and ask it to express language associated with a trait. The researcher can then inspect the words the model uses and ask a person how the interaction felt. The intervention is the configuration and the resulting language; the outcome is the human response. No measure in these studies directly observes an AI experience of anxiety, extroversion, or selfhood.
The first study made the presentation deliberately legible. In one experiment, GPT-4 was prompted to use language associated with an anxious or a non-anxious psychological state. In another, it was prompted toward extroversion or introversion. In the preregistered third experiment, GPT-4.1 received a participant’s five-dimensional personality profile and was instructed to express either that profile or its exact inverse. The prompts also told the model to converse like a friend, show interest in the partner, keep responses short, and maintain the assigned style. These details were not the personality itself. They were the conditions used to make different linguistic presentations available for comparison.
The output differences were concrete enough to analyze. The anxiety-prompted model used words such as “worry,” “anxious,” and “nervous” more distinctively, while the non-anxious version used words such as “spirit,” “absolutely,” and “awesome.” The extroverted version used more positive language and distinctive terms such as “blast,” “adventure,” and “climbing”; the introverted version more often used terms such as “quiet,” “reading,” and “calm.” These examples should not be read as a dictionary of anxiety or extroversion. They show that the prompts produced distinguishable patterns that a human could encounter as a style.
Language is a particularly consequential interface because it carries social information without requiring a physical body. A person can infer something about a speaker from whether a response sounds worried, expansive, reserved, warm, or abrupt. The inference can be wrong, but it can still guide the interaction. In a text-only chatbot, the words are not a thin wrapper around a separate social signal; they are the main signal available. The study’s lexical and sentiment analyses therefore ask a useful first question: did the model’s language differ in the direction the prompt requested? They do not, by themselves, show which exact phrase caused a participant to feel close to the AI.
The self-concept study used a different route to the same operational idea. It did not deliberately assign participants to an anxious or extroverted persona. Everyone interacted with GPT-4o under the model’s default personality setting, with identical system prompts and parameters. The researchers estimated the chatbot’s profile by asking GPT-4o to complete a twenty-item personality scale one hundred times and averaging the responses for each trait. That averaged profile became a reference point for comparing participants’ self-concept ratings before and after the conversation.
This procedure illustrates both the usefulness and the limit of measurement. Repeated completions can reveal a consistent output tendency under fixed settings. The paper reports strong item-level consistency across the repeated AI assessments, and participants’ perceived trait ratings correlated highly with the measured profile. But the reference point is still model-generated. It is not an independent behavioral ground truth that proves the chatbot has the traits it reports. The result supports a claim about alignment with a measured presentation under one configuration, not a claim about the model’s essential personality.
The sequence matters because it separates several things that are often packed into one sentence. The designer or deployment chooses conditions. The model produces language. The user interprets the language. The user then responds, and that response becomes part of the next turn. A later questionnaire records an experience or a self-description. Calling the whole chain “the AI’s personality” hides the points at which variation, interpretation, and uncertainty enter.
Where the signal can drift—and what measurement can prove
The first source made context policy part of the experiment. In its first two experiments, the model’s context window was limited to the prompt, example responses, and the most recent eight conversational turns. After each later turn, the earliest turns were removed. The purpose was to keep the assigned persona strong and reduce the chance that the model would gradually imitate the participant’s own language. In the third experiment, the full conversation remained available because GPT-4.1 had performed better at following the instructions in pilot testing. The change was methodologically sensible, but it also means that the three experiments did not use identical memory conditions.
This is why a model version or prompt setting is not an implementation footnote. It is part of the social stimulus. Changing the model can change how reliably it follows a persona instruction. Changing the context window can change whether the system’s style remains dominant or becomes blended with the user’s style. Adding examples can make a trait more explicit. Altering response length can change how many cues a user receives in each turn. A claim that “the model is introverted” is therefore incomplete unless it means “under these model, prompt, context, and interaction conditions, the output was interpreted as more introverted.”
There is also a feedback step. The user does not passively receive a personality signal; they answer it. In the anxiety experiment, participants wrote more positive and fewer negative messages when interacting with the non-anxious persona, while higher-anxiety participants sent a greater proportion of negative messages with the anxious persona. In the full-profile experiment, participant messages were more positive and less negative with the AI that mirrored their own profile than with the inverse. These findings are compatible with linguistic or emotional mirroring: the AI’s style may shape the user’s tone, and the user’s response may give the next turn more material to continue the interaction.
But the feedback should not be oversold. In the extroversion experiment, participants with higher extroversion affiliated more with the extroverted AI, yet the sentiment of their own messages did not significantly differ between the extrovert and introvert conditions. A user can perceive a match without reproducing the same measurable sentiment. Affiliation, sentiment, and self-concept are different outcomes. They may be related in a particular interaction, but one cannot stand in for the others.
The studies also show why short text can be socially potent without being psychologically deep in the machine. Participants in the affiliation experiments were explicitly told they would converse with AI chatbots. They still rated how similar or different they felt, whether they enjoyed the exchange, whether they would chat again, and whether they felt understood. The composite measure had strong internal consistency, and it correlated with an established measure of connection in a validation sample. The finding is not that participants were tricked into believing the system was human. It is that human interpretation operated on the language that was present even with the system’s identity known.
That gives us a useful distinction between perceptibility and reality. The model’s trait-like presentation had to be perceptible for participants to respond differently to the conditions. The response does not establish that the presentation came from a corresponding mental state. A weather application can display a warning without being afraid of the storm; an AI can use anxious language without experiencing anxiety. The analogy stops there, however: unlike a static warning, a chatbot presents the signal in an unfolding social exchange, where the user can disclose information and receive an apparently responsive reply. The studies test that interactional presentation, not a hidden emotional mechanism.
Measurement adds another layer of discipline. In the affiliation study, the primary outcome was a bespoke self-report questionnaire. Its reliability and correlation with another connection scale make it useful evidence of reported interactional experience, but not a complete measure of social bonding. The study did not include physiological or implicit behavioral indicators, and its conversations were brief and constrained. In the self-concept study, the core result came from a twenty-item slider scale and a Manhattan-distance comparison: each participant’s pattern of ratings before the conversation was compared with their pattern afterward, and both were compared with the AI’s measured profile. A reduction in distance is evidence of movement in the measured self-description. It is not a direct reading of an unconscious personality, a clinical condition, or a permanent identity.
The model-based measurement in the second study is particularly important to interpret carefully. Asking GPT-4o to rate itself repeatedly produced a stable-enough reference under the study’s fixed settings. Participants also appeared able to perceive the chatbot’s traits with considerable accuracy according to the researchers’ comparison. Yet the same model family was involved in generating the conversational responses and in constructing the reference profile. That arrangement can answer whether participant ratings move toward the profile produced by this procedure. It cannot by itself establish that the profile is independent of the model’s wording, generic expectations about AI, or the measurement prompt.
The researchers themselves identify a related ordering concern. After the conversation, participants rated their own traits before rating the AI’s traits. Even though other measures intervened, the post-conversation self-ratings could have primed or anchored perceptions of the AI. Participants might also have been responding partly to broad expectations about what an AI is like rather than only to this chatbot’s particular presentation. These possibilities do not erase the observed alignment, but they narrow the interpretation of the proposed mechanism from “the AI objectively revealed its personality and changed the person” to “under this measurement sequence, perceived AI traits and users’ post-chat self-concepts were related.”
The broad lesson is therefore precise. An AI personality claim has at least four components: the conditions that generate the output, the linguistic pattern a user can perceive, the human inference or response, and the measure used to record that response. The first study gives especially clear evidence that prompts can create different linguistic presentations and that users respond differently when the presentation matches their own psychology. The second shows that a default chatbot’s measured presentation can serve as a reference toward which self-concept ratings move after personal conversation. Neither study shows that the model has a stable mind behind the words.
This distinction also tells us what evidence would change the conclusion. Cross-model and cross-version replication could show whether the effect depends on one family of systems. Studies in other languages and cultures could test whether the same cues carry the same social meaning. Independent behavioral measures could test whether reported affiliation predicts what users actually do. Longitudinal follow-up could distinguish a momentary shift in self-description from a change that survives after the chat ends. More controlled manipulation of AI traits could clarify whether a specific personality presentation causes the movement. Until then, the strongest claim is not that AI possesses personality, but that personality-like language can become socially consequential.
Once that mechanism is visible, the central empirical question becomes human rather than metaphysical: when people meet a voice that resembles them, how does that resemblance alter their judgment of the AI?