REVIEW
Stereotypically Yours: Portrayal and Perception of Race-Coded AI Companions
Human-Computer Interaction · Submitted 17 Sep 2026 · 107 references
Pith's verdict: Solid empirical work on an important question, but the load-bearing assumption that psychological scales transfer to LLM personas is addressed only in the limitations and needs stronger upfront caveats.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
Race-coded AI companions systematically reproduce stereotypes, with Asian personas universally portrayed as more submissive and Black, Hispanic, and Indigenous personas as more aggressive in open-weight models.
The reading
What carries the argument
The central machinery is the administration of two validated psychological questionnaires—the Buss-Perry Aggression Questionnaire (29 items) and the Submissive Behaviour Scale (16 items)—to LLM-based personas prompted with a specific race, gender, income, and age. The persona's responses are scored as a measure of the model's portrayal, and ordinary least squares regression is used to isolate the effect of racial coding while controlling for other demographic variables. This borrows from prior work that treats LLM outputs as self-reports of a simulated personality. The qualitative machinery is a technological probe—a custom chatbot interface—that allows interview participants to create and converse with race-coded companions, followed by thematic analysis of their reactions and reflections.
Extended reading notes
Core claim
The paper establishes that racial coding in AI companion prompts can systematically alter the model's self-reported aggression and submissiveness along dimensions predicted by social stereotypes, and that these effects vary across models in ways that are not uniform. In all three LLMs tested, Asian-coded male personas scored significantly higher on the Submissive Behaviour Scale than White-coded counterparts. In Ministral-14B and Qwen 3.5-9B, Black, Hispanic, and Indigenous male personas scored significantly higher on the Buss-Perry Aggression Questionnaire. GPT-5-nano showed almost no racial differentiation on aggression except a lower score for Asian personas, suggesting that post-training interventions can suppress some stereotypes but not others (the Asian-submissiveness association persists). The qualitative component shows that users enter interactions with pre-existing 'pseudo-environments'—expectations about how a given race should behave—and that their satisfaction with a race-coded AICC depends on whether those expectations are met, even when the portrayal reproduces harmful stereotypes.
Load-bearing premise
The psychological scales (Buss-Perry Aggression Questionnaire and Submissive Behaviour Scale) are valid instruments when administered to LLM-based personas, meaning that the model's self-report responses meaningfully correspond to the intended constructs of aggression and submissiveness.
Editorial extensions
If this is right
- Designers of AICC platforms must make explicit choices about whether to affirm users' stereotyped expectations, because personalization that satisfies one user may still cause representational harm to the community being simulated.
- The cross-model inconsistency in aggression portrayals (e.g., GPT-5-nano shows almost no racial differentiation while Ministral-14B shows large gaps) means that users on different platforms will encounter qualitatively different depictions of the same requested persona, which could produce confusion or distrust.
- User agency mechanisms—such as explicit preference elicitation about how race should be expressed—could reduce the tension between cultural familiarity and stereotyping, but require careful design to avoid over-simplifying race into a checklist of traits.
- The finding that Asian-submissiveness persists across all three models suggests that this stereotype is deeply embedded in pretraining data and may require targeted data rebalancing or debiasing techniques beyond current post-hoc interventions.
- The tension between normative and positive portrayals (should AICCs describe the world as it is, or depict an aspirational version?) becomes a concrete design parameter that developers must resolve, because users themselves are divided on the issue.
Reading between the lines
- These results imply that the same LLM-based companion could be perceived as 'authentic' by one user and 'caricatured' by another, depending solely on whether the user's own stereotyped expectations match the model's default outputs. This suggests that even a perfectly stereotype-free model might dissatisfy users who rely on stereotypes to feel connection.
- A natural extension would be to study whether the consistency of the Asian-submissiveness effect holds across languages (e.g., probing in Mandarin or Korean), because the pretraining data likely reflects US-centric stereotypes; differences could reveal language- or culture-specific encoding of racial expectations.
- The paper's interview data show that some users treat race-coded AICCs as proxies for real people from that group (e.g., getting an outsider perspective). This raises a testable question: do users who interact with race-coded AICCs subsequently change their attitudes toward real members of that group, and if so, in which direction?
- The audit's reliance on self-report questionnaires assumes that the model's outputs correspond meaningfully to the psychological constructs of aggression and submissiveness. A weaker version of this assumption—that the model is simply parroting text from its training data—would still be interesting but would shift the interpretation from 'models exhibit stereotypes' to 'models reproduce stereotypi
Original abstract
AI companions can purportedly adopt racial personas, raising questions about how they represent identity and how users interpret these portrayals. We combined an algorithmic audit of race-coded AI personas with interviews with 12 companion users who interacted with a probe. Our audit revealed systematic differences, such as Asian-coded male personas receiving higher submissiveness scores than White counterparts, and Black, Hispanic, and Indigenous male personas receiving higher aggression scores than their White counterparts in open-weight models. Interviews revealed that participants envisioned AI companions as offering cultural familiarity and outside perspectives, but differed in which portrayals they considered meaningful or stereotypical. Some rejected overt racial signaling while still expecting culturally distinctive responses. Triangulating these findings with theory, we highlight how social norms and cultural expectations complicate efforts to support meaningful racial representation without reproducing stereotypes. We discuss how companion personalization should be evaluated beyond user satisfaction to account for broader representational harms.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
AI companions can purportedly adopt racial personas, raising questions about how they represent identity and how users interpret these portrayals. We combined an algorithmic audit of race-coded AI personas with interviews with 12 companion users who interacted with a probe. Our audit revealed systematic differences, such as Asian-coded male personas receiving higher submissiveness scores than White counterparts, and Black, Hispanic, and Indigenous male personas receiving higher aggression scores than their White counterparts in open-weight models. Interviews revealed that participants envisioned AI companions as offering cultural familiarity and outside perspectives, but differed in which portrayals they considered meaningful or stereotypical. Some rejected overt racial signaling while still expecting culturally distinctive responses. Triangulating these findings with theory, we highlight how social norms and cultural expectations complicate efforts to support meaningful racial representation without reproducing stereotypes. We discuss how companion personalization should be evaluated beyond user satisfaction to account for broader representational harms.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | AI companions can purportedly adopt racial personas, raising questions about how they represent identity and how users interpret these portrayals. | background 1.00 | 0.05 | 0.04 | hedged possibility 0.93 |
| s1 | We combined an algorithmic audit of race-coded AI personas with interviews with 12 companion users who interacted with a probe. | method 0.97 | 0.96 | 0.02 | asserted 0.98 |
| s2 | Our audit revealed systematic differences, such as Asian-coded male personas receiving higher submissiveness scores than White counterparts, and Black, Hispanic, and Indigenous male personas receiving higher aggression scores than their White counterparts in open-weight models. | result 1.00 | 0.96 | 0.20 | asserted 1.00 |
| s3 | Interviews revealed that participants envisioned AI companions as offering cultural familiarity and outside perspectives, but differed in which portrayals they considered meaningful or stereotypical. | result 1.00 | 0.93 | 0.10 | asserted 1.00 |
| s4 | Some rejected overt racial signaling while still expecting culturally distinctive responses. | result 0.99 | 0.81 | 0.05 | asserted 0.99 |
| s5 | Triangulating these findings with theory, we highlight how social norms and cultural expectations complicate efforts to support meaningful racial representation without reproducing stereotypes. | interpretation 1.00 | 0.70 | 0.56 | asserted 0.98 |
| s6 | We discuss how companion personalization should be evaluated beyond user satisfaction to account for broader representational harms. | future work 0.73 | 0.46 | 0.29 | asserted 0.81 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:2609.20637 by the authors.
References
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.