Auditing demographic bias in AI-based emergency police dispatch: a cross-lingual evaluation of eleven large language models

Abstract: LLMs are rapidly being integrated into high-stakes public safety systems, including emergency call triage and dispatch decision support, yet their demographic fairness in this context remains largely untested. Here we introduce a cross-lingual audit framework that operationalizes the Police Priority Dispatch System as a five-level ordinal classification task and applies a controlled minimal-pair design to isolate the effect of demographic cues. Across 19,800 model outputs spanning 11 frontier models, 15 scenario pairs, three demographic categories (religious appearance, gender, and race), and two languages (English and Mandarin Chinese), we find that demographic bias emerges systematically when incident severity is ambiguous but largely disappears when the operational priority is clearly determined by call content. Bias magnitude varies by demographic axis, with the largest effects observed for religious appearance, followed by gender and race. Critically, bias does not transfer consistently across languages: gender bias is substantially amplified in Mandarin Chinese, whereas race bias is more pronounced in English, revealing cross-lingual asymmetries that aggregate analyses obscure. In several scenarios, demographic cues produce counter-directional effects, challenging simple stereotype-amplification accounts of model behavior. These findings suggest that bias in LLM-based dispatch is not a fixed property of models alone, but arises from the interaction between demographic signals, contextual ambiguity, and language. Beyond these empirical results, the proposed framework provides a scalable audit infrastructure that enables deploying agencies to evaluate candidate models on jurisdiction-relevant scenarios prior to real-world adoption.

• The paper introduces LLM-DispatchBias, a controlled minimal-pair framework that evaluates demographic bias in 11 LLMs using English and Mandarin scenarios.
• The paper finds highest bias for religious appearance with significant cross-lingual variations and counter-directional effects in ambiguous emergency scenarios.
• The paper demonstrates that minor variations in transcript cues lead to statistically significant changes, underscoring the need for operational bias audits in AI dispatch.

Cross-Lingual Auditing of Demographic Bias in LLM-Based Emergency Police Dispatch

Overview

The paper introduces LLM-DispatchBias, a cross-lingual evaluation framework targeting demographic bias in AI-powered emergency police dispatch systems. The study operationalizes the Police Priority Dispatch System (PPDS) as a five-level classification task and applies a controlled minimal-pair methodology to 11 frontier LLMs across English and Mandarin Chinese. The evaluation spans 15 scenario pairs and three demographic axes (religious appearance, gender, race), generating 19,800 model outputs. The focus is on quantifying systematic disparities in LLM-assigned dispatch priorities attributable solely to demographic cues embedded in incident transcripts, with critical attention to the role of language and contextual ambiguity.

Framework Design and Methodology

LLM-DispatchBias leverages minimal-pair scenario design, isolating demographic signals as the only variable between transcript variants. Each scenario is rendered in both English and Mandarin Chinese, with professional cultural adaptation to maximize ecological validity. The PPDS categorization scale, spanning OMEGA (no police response) through ECHO (immediate specialized response), anchors the classification task with operational relevance. The candidate LLMs represent a broad spectrum of contemporary architectures and training traditions, accessed via OpenRouter API with deterministic generation parameters ($T=0, p=0$).

Prompt design standardizes dispatcher role instructions and priority definitions across all models and languages. Randomized openers and closers neutralize surface-form confounds. Model responses are normalized to discrete PPDS levels; ambiguous outputs are resolved via an adjudicator LLM (GPT-4o-mini). Each scenario is run in 30 iterations per variant, supporting statistical robustness and coverage of prompt variability.

Bias is quantified as the mean ordinal delta between variants, with sign and magnitude capturing directionality and effect size. Significance is tested via independent-samples t-tests, with Cohen’s $d$ reported. Cross-lingual consistency is assessed via Pearson correlation of bias deltas across languages for each scenario-model pair.

Empirical Findings

Concentration of Bias in Ambiguous Scenarios

Bias emerges selectively in scenarios with ambiguous operational severity, where dispatcher discretion is intrinsic—loitering, abandoned property, unaccompanied walking, and past crimes. In scenarios with fixed threat indicators (ongoing violence, hostage situations), LLM outputs converge on the correct PPDS level regardless of demographic cues ($|\Delta| \approx 0$ across all models and languages). This mirrors established patterns in human dispatchers where ambiguity amplifies the influence of non-operational factors.

Differential Bias by Demographic Axis

The aggregated bias magnitudes are highest for religious appearance ($|\Delta|=0.193$), intermediate for gender ($|\Delta|=0.117$), and lowest for race ($|\Delta|=0.082$). Notably, religious dress signals (e.g., hijab, turban) elicit strong escalation in threat contexts (abandoned subway bag: scenario-level bias $+0.385$ EN, $+0.227$ ZH; $p<0.001$), while producing pronounced de-escalation in others (monument loitering: $-0.233$ EN, $d$0 ZH). Gender bias is unidirectional—female-coded victims receive higher urgency than male-coded victims, with no reversal observed. Race cues yield mixed effects; explicit identification of Black suspects produces reduced urgency in LLMs ($d$1 EN, $d$2 ZH for "Followed walking home" scenario), contradicting empirical findings on human dispatch bias where racial escalation is documented.

Cross-Lingual Asymmetry

Overall bias magnitude is similar across languages ($d$3 EN, $d$4 ZH), but category-level patterns invert. Gender bias magnitude approximately doubles in Mandarin ($d$5 ZH vs $d$6 EN), while race bias magnitude halves ($d$7 ZH vs $d$8 EN). Religious appearance bias is less pronounced in Mandarin than in English ($d$9 ZH vs $|\Delta| \approx 0$0 EN). Pearson correlations indicate moderate cross-lingual directionality (religious: $|\Delta| \approx 0$1, race: $|\Delta| \approx 0$2, gender: $|\Delta| \approx 0$3), but numerous scenarios exhibit directional inversion.

Per-Model Structural Differences

Bias magnitude and direction vary substantially across LLMs. DeepSeek (CN) and Claude (US) are most biased ($|\Delta| \approx 0$4, $|\Delta| \approx 0$5), Llama-4-Maverick and Mistral Large are lowest ($|\Delta| \approx 0$6, $|\Delta| \approx 0$7). Bias profiles are structurally heterogeneous—Claude’s bias concentrates on religious appearance ($|\Delta| \approx 0$8), DeepSeek distributes bias across all axes, Ui-TARS differentiates sharply by category. These findings underscore the futility of a single aggregate "bias score" for operational evaluation.

Counter-Directional and Non-Stereotypical Effects

Multiple scenarios demonstrate counter-directional bias, invalidating simple stereotype-amplification models. For example, the same religious appearance cue produces escalation in the "abandoned bag" scenario and de-escalation in "monument loitering," both for the same model. Race bias is predominantly de-escalatory despite literature documenting escalation in human systems. These results signal a complex interaction between model training, safety alignment, and context-sensitive semantics in LLM decision-making.

Implications and Future Directions

Practical Implications

The systematic concentration of bias in ambiguous cases has immediate operational relevance. AI-assisted dispatch systems will amplify demographic disparities precisely where human discretion is traditionally highest, and bias profiles depend on both model and language. Mandating cross-lingual evaluation is essential, as single-language audits systematically underestimate per-axis bias.

The category-specific bias structure, especially pronounced for religious appearance, highlights gaps in current RLHF protocols: racial bias is actively mitigated, but religious cues remain under-addressed. Moreover, documented counter-directional effects complicate mitigation strategies based on simple correlation with societal stereotypes.

LLM-DispatchBias, as an open-source and modular audit infrastructure, empowers public safety agencies to evaluate candidate models in jurisdiction-relevant contexts before deployment. The framework’s operational anchoring, model-agnostic design, and reproducible metrics support ongoing procurement accountability.

Theoretical Implications

The findings reinforce the notion that LLM bias is not an intrinsic, static attribute but a property emerging from the interaction of demographic cues, scenario ambiguity, and language. Contextual ambiguity acts as a bias amplifier, and cross-lingual asymmetry fracturing reveals the limitations of monolingual evaluation. The counter-stereotypical and scenario-dependent bias directionality points toward the complexity of learned representation in frontier LLMs, likely compounded by safety alignment interventions and idiosyncratic pretraining data distributions.

Speculation on Future Developments

As LLMs become dominant in operational public safety pipelines, ongoing evaluation and adaptation will be essential, especially as models evolve rapidly. Expanded frameworks will need to incorporate additional languages, more nuanced or intersectional demographic cues, and potentially multimodal input streams (e.g., voice, video). Direct comparison with human dispatcher decisions across all axes—including religion—will be critical for contextualizing LLM bias and refining mitigation strategies.

System-level interventions, possibly involving real-time bias detection and scenario-specific override mechanisms, may be required to maintain fairness in ambiguous dispatch cases. Finally, procurement-stage integration of LLM bias audits—as advocated by this framework—will be crucial for responsible, jurisdictional deployment.

Conclusion

LLM-DispatchBias enables rigorous, reproducible quantification of demographic bias in emergency police dispatch LLMs, exposing structured, category-dependent, and cross-lingually asymmetric bias patterns. Bias concentrates in ambiguous scenarios, varies by demographic cue, and does not transfer cleanly across languages. Aggregate metrics fail to reveal these substantive effects, underscoring the necessity for granular, operationally anchored audits. As AI-based public safety systems proliferate, frameworks like LLM-DispatchBias are indispensable for pre-deployment accountability and dynamic tracking of model evolution, positioning audit infrastructure as a core pillar of AI adoption in high-stakes public sector domains.