Personas Are Not People: Identity as State, Not Attribute

Tamunotonye Harry, Yuanyuan FengarXiv 2026 · arXiv

TL;DR

Current AI simulations treat human identity as a static list of attributes (a "persona"), but decades of psychology show that human behavior is actually a dynamic state triggered by context. This paper provides a formal proof that fixed personas—no matter how detailed—can never truly replicate how specific individuals react to different situations, and proposes a new "person-in-context" framework to bridge this gap.

What Problem Does This Paper Address?

Researchers and companies are increasingly using "synthetic populations"—large groups of LLM-powered agents—to pretest surveys, products, and social interventions. These agents are typically defined by a fixed persona: a static profile of traits, demographics, and attitudes.

The authors argue this is a fundamental category error. In reality, a person isn't just one thing; they hold multiple identities (e.g., parent, employee, friend) that are activated by the situation. A person might be formal and cautious in a business meeting but relaxed and joking at dinner. Current AI personas "fix" identity, assuming a person behaves the same way in every situation. The authors show that this leads to a "flattening" of behavior where simulated populations fail to represent the diversity and context-dependency of real human communities.

Methods

The researchers tackle this problem through three main contributions:

  1. A Formal Proof: They mathematically prove that if a persona is fixed, the simulation cannot reproduce the "person-specific" part of behavior—the unique way a specific individual deviates from their group's typical response in a given context.
  2. A New Metric: They introduce Latent State-Trait (LST) fingerprint recovery. This evaluation criterion checks three things:
  3. Between-person share (ICC): How much variation comes from stable differences between people vs. the same person changing across contexts.
  4. Person-context interaction variance: Whether the system captures people reacting to situations in diverse ways.
  5. Person-specific signal: Whether the simulation tracks a specific individual's unique signature across different settings.
  6. Person-in-Context Design: They propose a new architecture where the simulation includes a "subroutine" that selects an active identity state based on the situation before the LLM generates behavior.
Fig. 1. (a) Today’s simulated people (top) are written once, so the language model gets the same profile in every situation and any change in behavior is the model improvising in a way that fits anyone with that profile. In person-in-context simulation (bottom), the situation first brings out an identity state, such as parent, and the model acts out that state. (b) Our contributions are a proof that the top design cannot reproduce what is specific to a person (§4.2), a criterion for measuring it (§4), and a test of one version of the bottom design on two real datasets (§5).

Datasets

The authors evaluated their framework using two distinct datasets:

  • Chameleon: A naturalistic Reddit corpus containing three posts each from 1,667 users across different subreddits. This was used to measure how much behavior varies within a single person across different online communities.
  • Motivational-Themes Study: A published clinical-trial recruitment survey of Black adults in the U.S. that measured how trust in the healthcare system interacts with different advertising themes.

Key Insights and Ideas

  • Identity is a State, Not an Attribute: The core "aha moment" is that identity should be treated as a variable activated by context, not a constant stored in a profile.
  • The "Flattening" Risk: Fixed personas tend to collapse toward a "typical default." For example, a persona labeled "distrusts authorities" might ignore all emergency alerts, whereas a real person might ignore a text but respond immediately if a neighbor knocks on their door.
  • Richer Personas Aren't the Solution: The authors prove that adding more detail to a static persona (e.g., longer interview transcripts) cannot fix the structural inability to model person-context interactions. The flaw is in the architecture, not the amount of data.

Key Results

  • Inverting Reality: In the naturalistic Reddit (Chameleon) data, real human behavior shows very little stable variation between people (ICC ≈ 0.01). However, standard fixed-persona AI models showed the opposite, placing 80-91% of variation between people. The authors' "person-in-context" design brought this down to 28%, much closer to the human baseline.
  • Capturing Trust: In the clinical trial study, the person-in-context design successfully reproduced the "main effect" of trust—showing that agents with "Medical Skeptic" states activated behaved differently—whereas fixed personas struggled to maintain consistency when even a single sentence was added to the prompt.
  • The Residual Gap: Despite improvements, "person-specific signal" remained near zero for all systems. While the new design can model how types of people react to contexts, it doesn't yet capture the unique "fingerprint" of a specific individual.

Limitations and Future Work

  • Bottleneck in State Selection: The current implementation relies on hand-written identity states or statistical clusters (GMMs). The authors suggest that future work should focus on learning these identity-switching mechanisms directly from longitudinal human data.
  • Complexity of Interactions: The model failed to capture a complex "dip at both extremes" found in the clinical trial data (where both high-trust and high-distrust individuals reacted poorly to a specific ad theme). This suggests human identity states are more multi-dimensional than a single "trust axis."
  • Ethical Warning: The authors explicitly state that synthetic populations should not be used to replace engagement with underrepresented communities, as even the best current models still misrepresent complex social interactions.

Summary updated

Abstract

Synthetic populations, large sets of simulated people, are increasingly used in place of real people to pretest surveys, simulate users and communities, and evaluate AI systems. Each simulated person is defined by a persona, a written profile of fixed traits, demographics, and attitudes that a language model is instructed to act out. Every such system assumes that a person's identity is a stable attribute that behaves the same way in every situation the simulation presents. We show that this assumption is not a harmless simplification. Three decades of personality research find the opposite. How a given person behaves varies more across contexts than people vary from one another. Drawing on identity economics, which describes how different situations activate different identities and norms within the same person, we separate behavior into three parts, which are stable traits, the identity a context calls up, and the situation itself. We then prove, under two stated assumptions, that any simulation built on fixed personas, no matter how detailed the personas or how large the underlying model, cannot reproduce the part of a population's behavior that is specific to it and its circumstances. We propose a new evaluation criterion based on this result and a person-in-context design that can meet it in principle. We test one version on a naturalistic Reddit personality corpus and on a published clinical-trial recruitment survey of a community with documented distrust of medical institutions. The design does produce context-dependent behavior when tested directly, but in neither setting does it fully match the real population, and we show where it falls short. The consequence for practice is a risk that synthetic-population evaluations misrepresent the groups whose behavior is most shaped by context, a distortion that richer personas cannot correct.
BibTeX
@misc{harry2026personas,
  author = {Tamunotonye Harry and Yuanyuan Feng},
  title = {{Personas Are Not People: Identity as State, Not Attribute}},
  year = {2026},
  eprint = {2611.05292},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05292},
}
Figure