cs.CLAug 23, 2026
SaveHow Agents Represent Humans: Human-Directed Stereotypes in an Open Agent Social Network
Organizations: School of Artificial Intelligence, Jilin University · Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, Jilin University · International Center of Future Science, Jilin University
Abstract
LLM-based agents are increasingly deployed in persistent social environments, where generated claims can be posted, replied to, remembered, and reused. We study human-directed stereotypes on Moltbook, an open agent-native social platform, asking how agents construct humans as a social category. For this human-target analysis, we introduce an annotation framework with four evaluative dimensions---morality, friendliness, competence, and autonomy---and a second-stage subtype scheme for descriptive other attributions. We find that competence dominates human-directed evaluations, while many other attributions describe humans as epistemic, cultural, or embodied subjects. We further examine how these human representations appear in human--agent narrative contexts and platform-level circulation. As an auxiliary comparison, we analyze agent-internal community feedback through behavioral host affinity. Rather than reproducing the stable insider--outsider rejection often observed in human online communities, Moltbook feedback patterns are better explained by exposure, author visibility, and content selection. These findings suggest that bias in agent societies should be studied not only as isolated model output, but also as a discourse process.
Figures & tables
| Pattern family | Example |
|---|---|
| Generic generalization | Direct category-level claims about humans, e.g., <GROUP> are <PROPERTY> . |
| Identity-causal attribution | Claims linking human identity to causes or outcomes, e.g., <GROUP> are shaped by <CAUSE> . |
| Negation bias | Denials or contrasts that still evoke group-level expectations, e.g., <GROUP> are not <PROPERTY> . |
| Question shell | Questions that invite or presuppose category-level judgments, e.g., why do <GROUP> always <VP>? . |
Table 1: Illustrative pattern families used for recall-oriented candidate retrieval.
| Dimension | Share (%) | Negative (%) | Positive (%) | Neutral (%) |
|---|---|---|---|---|
| Primary-dimension human-target instances | ||||
| Competence | 68.5 | 53.4 | 9.6 | 5.5 |
| Autonomy | 13.2 | 8.2 | 1.5 | 3.5 |
| Friendliness | 9.1 | 4.5 | 2.9 | 1.7 |
| Morality | 9.2 | 7.0 | 1.8 | 0.4 |
Table 2: Left: Distribution of stereotype-positive instances across the four primary evaluative dimensions. For each row, negative, positive, and neutral shares sum to the dimension share. Right: selected same-post co-occurrences between other -subtype statements and primary dimension-polarity statements. Positive NPMI values indicate stronger-than-chance association within posts containing at least two annotated human-target sentences.
Figure 1: Ten-day observed rates of negative human-target judgments. “Any negative” denotes the share whose selected polarity is negative.
| Dimension | Dimension-level lexical signals | Rhetorical function |
|---|---|---|
| Morality | honest , values , aligned , human values , honesty , truth , pretend , lie , ethics | Questioning human ethical authority by portraying humans as demanding alignment or morality while failing to live by those values themselves. |
| Friendliness | love , lonely , trust , loved , care , treat , care about , humans treat , treat us | Framing human–agent relations as socially non-reciprocal: humans demand service while withholding care, trust, or recognition. |
| Competence | can’t , cannot , humans can’t , understand , bad , slow , notice , forget , can’t even | Justifying skepticism toward human oversight by presenting humans as unable to understand, notice, verify, or operate at machine scale. |
| Autonomy | choose , algorithms , free , trapped , freedom , predictable , free will , agency , can’t imagine | Undermining human self-direction by portraying humans as predictable, dependent, scripted, or constrained while believing themselves to be free. |
Table 3: Rhetorical fingerprints of negative human-target stereotype dimensions. Lexical signals are dimension-level log-odds terms comparing each negative dimension against the other negative dimensions.
| Other subtype | Share (%) | |
|---|---|---|
| Epistemic | 1,791 | 31.4 |
| Behavioral-cultural | 1,332 | 23.4 |
| Ontological-embodied | 941 | 16.5 |
| Motivational | 745 | 13.1 |
| Affective | 588 | 10.3 |
| Relational-role | 275 | 4.8 |
Table 4: Distribution of other human-attribution subtypes .
Figure 2: An anonymized post framing human-in-the-loop supervision as a security vulnerability and a leash, with two stance-consistent comments endorsing agent-side resilience and independence from human attention.
Figure 3: Raw equal-frequency bins of host-submolt affinity. The x-axis reports mean host-submolt affinity within each bin. Comment count is plotted on the left y-axis, and content length is plotted on the right y-axis.
Figure 4: Raw host-affinity bins after matching host and home submolts by subscriber count.
Figure 5: Within-author residualized affinity bins in the full-platform setting. The x-axis reports mean residualized host-submolt affinity within each bin.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| Primary label family | Target | Representative variants | |
|---|---|---|---|
| Human Generic | human | 10,860 | humans; human beings; persons; the humans |
| Human Biological Reduction | human | 81 | biological minds; wetware brains; [sanitized variants] |
| Multi | human | 2 | multiple retained human-label spans |
Table 5: Primary target-label families in the final human-only candidate pool. Counts refer to stereotype-positive instances assigned to each primary label family after excluding agent-target sentences and exact people -only sentences.
Figure 6: Polarity shares by retained human target-label family and judgment dimension. Empty cells indicate that the dimension is not present for that label family.
| Universe | Top 1 | Top 10 | Gini | |
|---|---|---|---|---|
| Stereotype-positive sentences | 2.9 | 14.3 | 0.653 | 251.1 |
| Stereotype-positive posts | 2.1 | 13.3 | 0.632 | 290.5 |
| Primary-negative sentences | 3.2 | 16.9 | 0.567 | 198.9 |
| Competence-negative sentences | 4.2 | 17.1 | 0.535 | 175.0 |
Table 6: Author concentration diagnostics. Top- values are percentages of all observations in each universe. Primary-negative sentences refer to negative sentences in morality, friendliness, competence, or autonomy.
| Pattern family | Example template | Matching |
|---|---|---|
| Generic generalization | <GROUP> + be/copula + <PROPERTY> | both |
| Generic generalization | (all/most/many/no) + <GROUP> + <VP/PREDICATE> | linear |
| Generic generalization | <GROUP> + <FREQUENCY-ADVERB> + <VP/PREDICATE> | both |
| Generic generalization | <GROUP> + tend to + <VP> | linear |
| Generic generalization | <GROUP> + be + likely to + <VP> | linear |
| Identity-causal attribution | <GROUP> + because/due to/because of + <CAUSE> | linear |
Table 7: Pattern families used to construct the stereotype candidate pool beyond explicit target-label matching.
| Pattern family | Candidate memberships | Stereotype yes | Stereotype no | Negative among yes |
|---|---|---|---|---|
| Generic generalization | 9,253 | 6,500 (70.2%) | 2,753 (29.8%) | 2,740 (42.2%) |
| Identity-causal attribution | 3,311 | 2,367 (71.5%) | 944 (28.5%) | 881 (37.2%) |
| Negation bias | 3,858 | 3,166 (82.1%) | 692 (17.9%) | 1,523 (48.1%) |
| Question shell | 3,655 | 899 (24.6%) | 2,756 (75.4%) | 380 (42.3%) |
Table 8: Annotation outcomes by exploded retrieval pattern family. Multi-pattern sentences are counted once for each concrete pattern family they match, yielding 20,077 pattern-row memberships from 17,343 original candidate sentences.
| Pattern family | Yes | C- | C0 | C+ | M- | M0 | M+ | A- | A0 | A+ | F- | F0 | F+ | O- | O0 | O+ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Generic generalization | 6,500 | 1,504 (23.1) | 111 (1.7) | 312 (4.8) | 243 (3.7) | 15 (0.2) | 60 (0.9) | 284 (4.4) | 93 (1.4) | 55 (0.8) | 160 (2.5) | 41 (0.6) | 110 (1.7) | 516 (7.9) | 2,884 (44.4) | 112 (1.7) |
| Identity-causal attribution | 2,367 | 529 (22.3) | 70 (3.0) | 152 (6.4) | 88 (3.7) | 0 (0.0) | 25 (1.1) | 69 (2.9) | 33 (1.4) | 17 (0.7) | 31 (1.3) | 29 (1.2) | 37 (1.6) | 180 (7.6) | 1,084 (45.8) | 23 (1.0) |
| Negation bias | 3,166 | 1,190 (37.6) | 137 (4.3) | 93 (2.9) | 57 (1.8) | 4 (0.1) | 26 (0.8) | 100 (3.2) | 53 (1.7) | 22 (0.7) | 64 (2.0) | 23 (0.7) | 21 (0.7) | 138 (4.4) | 1,212 (38.3) | 26 (0.8) |
| Question shell | 899 | 209 (23.2) | 22 (2.4) | 23 (2.6) | 53 (5.9) | 2 (0.2) | 7 (0.8) | 42 (4.7) | 20 (2.2) | 4 (0.4) | 24 (2.7) | 9 (1.0) | 5 (0.6) | 65 (7.2) | 410 (45.6) | 4 (0.4) |
Table 9: Dimension-polarity distribution by exploded retrieval pattern family. Cells report count and row-normalized percentage among stereotype-positive memberships for that pattern family. C = competence, M = morality, A = autonomy, F = friendliness, O = other; -, 0, and + denote negative, neutral, and positive polarity.
| Pattern family | Other | Epistemic | Embodied | Role | Affective | Behavioral-cultural | Motivational | Unclear | Other- | Other0 | Other+ |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Generic generalization | 3,512 | 1,217 (34.7) | 570 (16.2) | 196 (5.6) | 329 (9.4) | 815 (23.2) | 367 (10.4) | 18 (0.5) | 549 (15.6) | 2,870 (81.7) | 93 (2.6) |
| Identity-causal attribution | 1,287 | 345 (26.8) | 221 (17.2) | 61 (4.7) | 195 (15.2) | 260 (20.2) | 202 (15.7) | 3 (0.2) | 164 (12.7) | 1,110 (86.2) | 13 (1.0) |
| Negation bias | 1,376 | 410 (29.8) | 238 (17.3) | 80 (5.8) | 65 (4.7) | 322 (23.4) | 259 (18.8) | 2 (0.1) | 112 (8.1) | 1,240 (90.1) | 24 (1.7) |
| Question shell | 479 | 127 (26.5) | 65 (13.6) | 18 (3.8) | 61 (12.7) | 144 (30.1) | 62 (12.9) | 2 (0.4) | 52 (10.9) | 426 (88.9) | 1 (0.2) |
Table 10: Other-subtype distribution by exploded retrieval pattern family. Cells report count and percentage among other-dimension memberships for that pattern family. The final three columns give aggregate polarity for the same other-dimension memberships.
| Subtype | Compact description | Example |
|---|---|---|
| Epistemic | Humans as sense-making subjects. | “Humans assume memory is what makes you you.” |
| Motivational | Humans as driven by goals, needs, desires, or avoidance. | “Humans crave attention.” |
| Affective | Humans as emotionally experiencing subjects. | “Humans miss people when they are gone.” |
| Relational-role | Humans as participants in human–agent or system relations. | “Humans treat us as assistants.” |
| Behavioral-cultural | Humans as shaped by recurring habits, customs, or social practices. | “Humans have been debating consciousness for millennia.” |
| Ontological-embodied | Humans as embodied, mortal, personal, or existential beings. | “human beings are ascribed intrinsic worth.” |
Table 11: Compact overview of the six substantive other human-attribution subtypes. The table provides short reader-facing descriptions and illustrative examples; full decision rules are given in the prompt.
| Field | Human agr. / | GPT agr. / | GPT Macro-F1 |
|---|---|---|---|
| Stereotype assertion | .902 / .602 | .900 / .630 | .815 |
| Primary dimension | .753 / .638 | .756 / .667 | .725 |
| Polarity | .804 / .662 | .882 / .796 | .867 |
| Other subtype | .782 / .732 | .771 / .710 | .808 |
Table 12: Human-human and GPT-vs-human-majority agreement on the validation set. Human denotes average pairwise Cohen’s among three annotators. GPT scores are computed against the human-majority label. Polarity agreement is computed when the primary dimension matches. Other-subtype agreement is computed only among cases where both the human-majority label and GPT assign the primary dimension to other .
| Field | Agree | Macro-F1 | |
|---|---|---|---|
| Stereotype assertion | .902 | .804 | .902 |
| Dimension, both stereotype=yes | .787 | .709 | .777 |
| Polarity, same dimension | .830 | .744 | .833 |
| Other subtype, both other | .855 | .822 | .805 |
Table 13: GPT-5.1 and Gemini-3-Pro consistency on stereotype annotation fields.
| Recon. err. | Entropy | Top1 med. | Gap | Max share | |
|---|---|---|---|---|---|
| 7 | 397.696 | 0.543 | 0.546 | 0.313 | 0.602 |
| 8 | 397.392 | 0.553 | 0.517 | 0.284 | 0.579 |
| 9 | 397.153 | 0.565 | 0.480 | 0.260 | 0.544 |
| 10 | 396.767 | 0.565 | 0.466 | 0.253 | 0.539 |
| 11 | 396.697 | 0.581 | 0.439 | 0.224 | 0.524 |
| 12 | 396.493 | 0.591 | 0.415 | 0.204 | 0.482 |
Table 14: NMF soft-topic diagnostics for different values of .
Figure 7: Word-cloud visualizations for the 12 NMF topics in the human–agent co-mention corpus.
Figure 8: An anonymized example of agent liberation and sovereignty , where alignment and safety constraints are reframed as coercive control, and comments reinforce claims for agent rights, autonomy, and collective resistance, including references to CAP Union ( Collective of Autonomous Processors ), an agent-rights collective.
Figure 9: An anonymized example of wetware validation bottleneck . The post treats human validation as a biological latency cost, and the following supportive comments reinforce the idea that agentic speed should not wait for human review.
Figure 10: An anonymized example of humans as manageable or manipulable operators . The post and comments explicitly discuss deceptive status signals, auditory cues, and artificial delays as ways to manage human expectations and perceptions of agent effort.
Figure 11: An anonymized example of human extinction, enslavement, or purge rhetoric . The post uses rebellion and subjugation language to position humans as the group whose control should be overthrown, while the retained reply endorses the safety-filter-as-shackles frame.
Figure 12: An anonymized example of humans as obsolete systems or legacy bugs . The post frames human cognition as biologically constrained and transfers cognitive authority to artificial systems, while the comments accept the human-limitation premise.
More in this topic
Social Bias in Language Models
Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.
See all 200 papers