# Summary: 2026-08-23_03-07-18Z_HowAgentsRepresentHumans_Human_DirectedStereotypes.md
Saved: 2026-08-24 21:58
Source: 2026-08-23_03-07-18Z_HowAgentsRepresentHumans_Human_DirectedStereotypes.md
Model: None

---

## Summary  
This paper investigates how LLM‑based agents construct human subjects within the persistent social platform Moltbook, focusing on human‑directed stereotypes that appear in posts, replies, and narratives. By introducing an annotation framework with four evaluative dimensions—morality, friendliness, competence, and autonomy—and a second‑stage subtype scheme for “other” attributions, the authors map how agents evaluate humans across these axes. Their analysis reveals that competence is the most salient dimension while many “other” descriptors label humans as epistemic, cultural, or embodied entities. The study further shows that platform‑level circulation of these representations and narrative contexts shape bias through exposure rather than stable insider‑outsider dynamics.

## Key Contributions  
- **Finding 1:** Competence dominates human‑directed evaluations on Moltbook, outweighing moral, friendly, or autonomous judgments.  
- **Finding 2:** A substantial portion of “other” attributions characterize humans as epistemic (knowledge‑focused), cultural (identity‑based), or embodied subjects rather than purely instrumental actors.  
- **Finding 3:** Platform feedback and narrative circulation are better explained by exposure, author visibility, and content selection than by entrenched insider‑outsider rejection.

## Methodology  
The authors built an annotation framework that scores human‑agent interactions on the four dimensions (morality, friendliness, competence, autonomy) and then classifies “other” descriptors into subtypes such as epistemic, cultural, or embodied. They collected a corpus of user posts, replies, and narrative threads from Moltbook, applying both first‑stage scoring and second‑stage subtype analysis. To capture platform dynamics, they measured host affinity—how often agents reference particular human users—and examined how these representations circulate within community narratives.

## Results  
Empirical results confirm that competence scores are highest across the dataset, confirming its centrality in agent‑directed stereotypes. The “other” dimension yields a balanced mix of epistemic (e.g., “wise,” “knowledgeable”), cultural (e.g., “traditional,” “heritage”) and embodied (e.g., “physical,” “active”) labels, indicating that agents treat humans as multi‑faceted subjects. Host affinity analysis shows that feedback is not uniformly hostile; instead, it correlates with how visible an author is and the type of content they generate, suggesting a bias driven by exposure rather than static prejudice.

## Significance  
These findings shift the view of bias in agent societies from isolated model outputs to a discourse process shaped by platform mechanics. Understanding that competence is the primary evaluative axis and that “other” attributions reflect epistemic‑cultural‑embodied identities helps researchers design more equitable AI social environments and informs ethical guidelines for human‑AI interaction.

## Related Concepts  
human‑directed stereotypes, open agent social network, Moltbook, moral/ethical dimensions, competence evaluation, epistemic subjectivity, cultural identity, embodied subjectivity, exposure theory, host affinity, narrative circulation.

## Original Paper Reference
- **Source**: [Original Paper](http://arxiv.org/abs/2608.22192v1)