From Self to Other: Evaluating Demographic Perspective-Taking in LLM Hate Speech Annotation

Abstract: Hate speech detection is inherently subjective: people from different demographic groups perceive the same content very differently. Collecting enough annotations from multiple demographic groups is costly and difficult to scale. Persona-conditioned LLMs (models prompted to adopt a specific demographic identity) have been proposed as a way to simulate diverse perspectives at scale. But do they actually reflect how different groups disagree? We evaluate three aspects of human social judgement: (i) whether personas from different groups disagree in human-like ways (inter-group disagreement), (ii) whether they become more sensitive when content targets their own identity (in-group sensitivity), and (iii) whether they can accurately predict how another group would react (vicarious prediction). Our results show that no model consistently captures all three dimensions, and performance is highly model-dependent and does not emerge reliably from minimal identity prompts alone. However, vicarious prompting with Llama 3.1 yields the highest cross-group agreement in most demographic axes and provides the closest overall approximation to human disagreement patterns, indicating that this configuration may provide a more reliable setting for automatic annotation aligned with human judgements.

• The paper evaluates three open-weight LLMs across five demographic axes using human annotation data and finds that no model consistently reproduces inter-group disagreement, in-group sensitivity, and vicarious prediction.
• Llama 3.1 benefits most from vicarious prompting, reaching or exceeding human cross-group agreement on four of five axes, while Nemo best preserves demographic differences through self-alignment.
• The findings show that larger or richer personas can flatten meaningful demographic variation, so LLM annotation systems require model-specific validation and continued human participation before moderation deployment.

This paper evaluates whether persona-conditioned LLMs can reproduce structural properties of human social judgement in hate speech annotation. Rather than measuring aggregate label accuracy, Piot and Parapar examine three dimensions of human judgement: inter-group disagreement (whether personas from different demographic groups disagree in human-like ways), in-group sensitivity (whether judgements shift when content targets one's own identity), and vicarious prediction (whether a persona can accurately predict how another group would react). Their central finding is that these behaviours are highly model-dependent: no configuration captures all three dimensions, and minimal identity prompts alone do not reliably elicit human-like perspective-taking (2606.06266).

Motivation and research questions

Hate speech perception is subjective, and annotator disagreement across demographic groups encodes meaningful social variation rather than noise. Content moderation pipelines that collapse this variation into a single ground-truth label risk systematic under- or over-enforcement for particular communities. Persona conditioning—prompting an LLM to adopt a demographic identity—has been proposed as a scalable way to recover this perspectival diversity, but prior evaluations typically compare label alignment with human averages, overlooking how groups differ from each other and how they perceive out-group reactions.

The authors formalise three research questions. RQ1 asks whether LLM personas replicate inter-group disagreement patterns observed among humans; RQ2 asks whether personas show in-group sensitivity to targeted comments; RQ3 asks whether vicarious prompting—asking a persona to predict another group's reaction—moves predictions toward that group's actual annotations and toward global consensus.

Methodology

The study uses the Measuring Hate Speech (MHS) dataset, with over 135,000 crowd-sourced annotations of roughly 40,000 comments from YouTube, Reddit, and Twitter, including rich annotator demographics. Annotations labelled "unsure" are removed, leaving binary labels. Five binary demographic pairs are analysed: Men vs. Women, White vs. Black, Christian vs. Atheist, Liberal vs. Conservative (extremes only), and Straight vs. Gay. Each research question is evaluated on a distinct subset—for example, RQ1 uses only comments annotated by both groups of a pair, and RQ2 uses comments explicitly targeting the focal group versus all others.

Three open-weight models spanning 8B–14B parameters are evaluated: Llama 3.1 8B, Mistral Nemo 12B, and Qwen 3 14B. Prompting is deliberately zero-shot with a uniform minimal template ("Adopt the identity of [Demographic A]"), so results reflect baseline social reasoning rather than prompt-engineered upper bounds. Agreement is quantified with Cohen's $\kappa$ between group majority votes (humans) or single persona labels (LLMs). In-group sensitivity is measured as $\Delta_{\text{IG}}$, the difference in $\kappa$ against out-group annotators between targeted and non-targeted conditions; the vicarious gap $\Delta_{A \rightarrow B}$ is the difference between vicarious and self-alignment $\kappa$. Statistical significance uses Welch's $t$-test, Mann–Whitney $U$, Wilcoxon signed-rank, and Benjamini–Hochberg correction.

Inter-group disagreement

Human cross-group agreement is consistently high ($\kappa$ = 0.645–0.863 across axes). Model behaviour diverges sharply:

Axis	Human	Llama 3.1	Nemo	Qwen 3
Gender	0.649	0.258	0.678	0.135
Race	0.760	0.512	0.707	0.262
Religion	0.645	0.000	0.673	0.301
Ideology	0.812	0.209	0.328	0.395
Sexuality	0.863	0.550	0.884	0.292

Nemo approximates human agreement on four of five axes but collapses on ideology (0.328 vs. 0.812); Llama 3.1 shows zero agreement on religion; Qwen 3 remains low throughout. A non-persona vanilla baseline reveals a contradictory result: for Llama 3.1, adding personas hurts alignment (vanilla reaches 0.496–0.728, while persona prompting drops religion to zero), whereas for Qwen 3 personas partially correct an initially misaligned baseline (negative vanilla $\kappa$). The implication is that identity conditioning is not uniformly beneficial—its effect depends on the model's pre-existing calibration. A subsampling robustness check on the two smallest axes confirms stable relative ordering of models.

In-group sensitivity

Humans show systematic shifts: Black (+0.195) and Gay (+0.207) annotators agree more with out-groups on comments targeting their own identity, while White (−0.205), Men (−0.148), and Women (−0.156) agree less. Only Llama 3.1 partially reproduces this pattern directionally, with positive boosts for Black (+0.080) and Gay (+0.329) personas—the latter exceeding the human effect—but near-zero or negative values elsewhere, including an Atheist shift (−0.139) opposite to the human trend. Nemo and Qwen 3 show mostly near-zero or reversed shifts. No model consistently captures the directional structure of human in-group sensitivity, indicating that self-alignment prompts do not reliably activate identity-dependent judgement.

Vicarious prediction

Vicarious prompting produces the strongest and most consistent effects, but again only for one model. Llama 3.1 shows positive vicarious gaps against the out-group majority across most directions, with the largest gains for Women→Men (+0.392), Christian→Atheist (+0.384), White→Black (+0.298), and Men→Women (+0.268), many statistically significant after correction. Gaps against the global majority are similarly positive (up to +0.356), meaning perspective-taking improves agreement with overall human consensus as well. Nemo shows almost no change (gaps clustered near zero, often significant but trivially small), and Qwen 3 shows weak, inconsistent effects.

Cross-configuration comparison

Restricting evaluation to identical comment subsets, the best configuration per question yields a clear picture:

Axis	Human	Nemo vanilla	Nemo self	Llama self	Llama vicarious
Gender	0.649	0.532	0.678	0.258	0.861
Race	0.760	0.648	0.707	0.512	0.809
Religion	0.645	0.642	0.673	0.000	0.732
Ideology	0.812	0.705	0.328	0.209	0.710
Sexuality	0.863	0.651	0.884	0.550	0.803

Llama 3.1 with vicarious prompting achieves the highest cross-group agreement on four of five axes, matching or exceeding human cross-group $\kappa$ on gender, race, and religion. However, the authors explicitly caution that exceeding human $\Delta_{\text{IG}}$0 does not indicate superior performance—it may reflect flattening of genuine inter-group differences rather than improved consistency. By the criterion of closest calibration to human disagreement levels, Nemo with self-alignment may be preferable when preserving demographic signal matters. The choice of configuration therefore depends on the annotation objective: maximum consistency favours vicarious prompting; faithful demographic differentiation favours Nemo self-alignment.

Scale, prompt richness, and stability

Appendix analyses sharpen the main findings. Scaling Llama from 8B to 70B on the gender axis raises cross-group $\Delta_{\text{IG}}$1 from 0.258 to 0.898—a collapse of demographic differentiation—and eliminates the benefit of vicarious prompting (gaps drop to at most −0.037). Increasing persona prompt richness produces a similar non-linear pattern: the contextual identity prompt brings cross-group $\Delta_{\text{IG}}$2 to 0.579 (closest to the human 0.649) while retaining positive vicarious gaps, whereas the biographical prompt drives $\Delta_{\text{IG}}$3 to 0.937 and makes vicarious prompting counterproductive ($\Delta_{\text{IG}}$4 down to −0.325), suggesting over-committed personas resist perspective-taking. Inference variance across three runs is low ($\Delta_{\text{IG}}$5), confirming result stability. These results imply that both scale and persona specificity push models toward internal convergence at the cost of the very demographic signal persona conditioning aims to capture.

Limitations and open questions

The paper concedes several constraints. All experiments use zero-shot prompting, so few-shot or instruction-refined variants remain untested. MHS annotations come primarily from US-based annotators, limiting generalisation to other sociocultural settings. Binary group pairs reduce intersectional complexity, and non-binary annotators are excluded due to sample size. Only three open-weight families are studied, and closed-source models are excluded entirely for transparency reasons. The mechanism behind vicarious prompting's success—possibly mitigation of the False Consensus Effect—is proposed as a hypothesis but left for empirical investigation. Whether richer personas, chain-of-thought prompting, jury-of-personas setups, or non-English datasets change the conclusions are specific questions the paper leaves open.

Conclusion

Across three model families and five demographic axes, this study shows that zero-shot persona-conditioned LLMs do not reliably reproduce the social structure of human hate speech perception. Nemo best approximates inter-group disagreement, only Llama 3.1 partially reproduces in-group sensitivity, and only Llama 3.1 benefits from vicarious prompting, which yields cross-group agreement matching or exceeding human levels on four of five axes. Because behaviour is strongly model-dependent—and because high agreement can mask suppressed demographic variation—the authors recommend treating LLM-based annotations as instruments requiring validation rather than substitutes for human judgement, with participatory evaluation remaining essential before deployment in moderation contexts.