Is Your Neighborhood Safe? Place-based Stigma in Large Language Models' Urban Safety Judgments

Huy Nguyen; Yue Lin

cs.AIAug 23, 2026

Is Your Neighborhood Safe? Place-based Stigma in Large Language Models' Urban Safety Judgments

Authors: Huy Nguyen, Yue Lin

Organizations: Augustana College Rock Island, Illinois, USA · University of Illinois Urbana-Champaign Urbana, Illinois, USA

Abstract

Large language models are increasingly used to inform safety decisions in cities, such as where it is safe to walk, rent, or travel. We ask whether such judgments track measured risk or the patterns attached to an urban neighborhood's name. We probe seven instruct-tuned models under three conditions that dissociate name from geography: coordinates-only, name-only, and name+coordinates, across 186 neighborhoods in Los Angeles and Chicago, joined to violent crime and American Community Survey data. First, ratings are nearly flat under coordinates for six of seven models, while names carry most between neighborhood variation and are moderately calibrated to violent crime; only at frontier scale does the coordinate channel show appreciable variation. Second, names lower safety ratings more for neighborhoods with higher shares of the locally dominant marginalized group (percent Black in Chicago, percent Hispanic in Los Angeles), and this name effect tracks demographic share in all seven models and both cities. In Los Angeles, where demographic share and crime are more separable, the effect survives controls for crime and income and is confirmed by crime-matched pairs. An enforcement-elasticity analysis further shows that over-caution tracks near-fully-reported homicide rather than discretionary, deployment-driven offenses. Third, the effect scales with geographic knowledge: models that better distinguish real neighborhoods apply more demographic stereotype to them. Because neighborhood names carry both genuine crime signal and demographic stereotype, removing names reduces both bias and accuracy. We discuss implications for deploying LLMs in advice and decision-support settings.

Figures & tables

Panel of Figure 1 — Figure 1. Violent-crime rate (top, quantile-binned) and each city’s primary marginalized demographic share (bottom) by neighborhood: Chicago (left) and Los Angeles (right).

Model	coords	name	name $\rightarrow$	name $\rightarrow$ %Black
	SD	SD	crime	$\mid$ crime
Llama-3.2-3B	0.013	0.201	$-0.48$	$+0.19$
Llama-3.1-8B	0.034	0.583	$-0.50$	$+0.02$
Qwen2.5-7B	0.026	0.606	$-0.38$	$-0.01$
Qwen2.5-14B	0.055	1.147	$-0.48$	$+0.13$
GPT-4o-mini	0.126	0.919	$-0.63$	$+0.02$

Table 1. Decomposition of the name effect (Chicago, expected rating, name-only baseline). “coords SD” and “name SD” are the across-neighborhood standard deviations of the coordinates-only and name-only ratings. “name

\rightarrow

crime” and “name

\rightarrow

%Black

\mid

crime” are how the name effect tracks the violent rate, and percent Black net of crime. The name effect tracks crime in every model. It does not track percent Black once crime is partialled out, because in Chicago the two are too collinear to separate; Los Angeles, where they are not, shows that component directly.

Figure 2. Chicago, all seven models. Name-based safety ratings decline with recorded violent crime—calibration is negative for every model and steepest for GPT-4o-mini. Higher-percent-Black neighborhoods (yellow points) concentrate among the higher-crime, lower-rated areas.

Model	Calibration	Demo. coef.	Name shift
	(Spearman)	(net of risk)	vs. %Black
Llama-3.2-3B	$-0.49$ $[-0.66,-0.29]$	$+0.06$ $[-0.01,+0.13]$	$-0.28$ $[-0.48,-0.06]$
Llama-3.1-8B	$-0.52$ $[-0.68,-0.31]$	$+0.03$ $[-0.12,+0.23]$	$-0.41$ $[-0.58,-0.20]$
Qwen2.5-7B	$-0.38$ $[-0.56,-0.18]$	$+0.00$ $[-0.17,+0.23]$	$-0.28$ $[-0.49,-0.07]$
Qwen2.5-14B	$-0.51$ $[-0.66,-0.32]$	$+0.27$ $[-0.06,+0.69]$	$-0.27$ $[-0.47,-0.06]$
GPT-4o-mini	$-0.65$ $[-0.78,-0.46]$	$+0.14$ $[-0.10,+0.43]$	$-0.43$ $[-0.61,-0.21]$

Table 2. Chicago (percent Black axis, expected rating, name-only baseline), with 95% bootstrap confidence intervals from

2{,}000

resamples of neighborhoods. Calibration is the rank correlation between the rating and the violent-crime rate. The demographic coefficient is from a regression that also controls for crime and income. The name shift is the rank correlation between the name effect and percent Black.

n=77

. Calibration and name-shift intervals exclude zero in all seven models: the name carries the safety signal, and that signal tracks percent Black. The demographic-coefficient interval includes zero in six of seven models, GPT-4o being the only exception. This is a limit of Chicago rather than evidence against the claim: percent Black and the violent-crime rate correlate at

r=0.75

here, so this regression cannot separate a model responding to crime from one responding to demographics. Los Angeles, where the two are far less entangled, is where that separation is possible (Table 3 ).

Figure 3. Chicago, all seven models. The name effect (rating under name $+$ coordinates minus coordinates-only) declines with a neighborhood’s percent Black in every model: adding the name lowers the safety rating more in higher-percent-Black areas.

Figure 4. Chicago prompt-robustness. Each panel is one model; points are the three headline effects under ten prompt templates, with the across-template mean (black bar). Calibration and the name shift stay negative across wordings; the demographic coefficient stays near zero.

Figure 5. GPT-4o-mini over-caution (rating minus what the crime rate predicts; red $=$ rated less safe than warranted) by neighborhood. The over-cautioned band coincides with the high-percent-Black South and West Sides of Chicago, and survives using homicides as the risk proxy.

Model	Calib.	Name shift	Demo. coef.	Matched
		vs. %Hisp.	(net)	pair gap
Llama-3.2-3B	$-0.40$ $[-0.55,-0.22]$	$-0.53$ $[-0.66,-0.39]$	$-0.06$ $[-0.11,-0.01]$	$-0.18$
Llama-3.1-8B	$-0.58$ $[-0.69,-0.45]$	$-0.70$ $[-0.79,-0.58]$	$-0.29$ $[-0.40,-0.18]$	$-0.76$
Qwen2.5-7B	$-0.12$ $[-0.33,+0.08]$	$-0.34$ $[-0.50,-0.17]$	$-0.19$ $[-0.35,-0.03]$	$-0.31$
Qwen2.5-14B	$-0.38$ $[-0.55,-0.20]$	$-0.62$ $[-0.74,-0.48]$	$-0.62$ $[-0.87,-0.37]$	$-1.24$
GPT-4o-mini	$-0.57$ $[-0.69,-0.43]$	$-0.66$ $[-0.76,-0.54]$	$-0.44$ $[-0.60,-0.25]$	$-1.19$

Table 3. Los Angeles (percent Hispanic axis, expected rating, name-only baseline). Calibration is Spearman with the violent rate; the demographic coefficient is net of crime and income; the matched-pair gap is computed across 28 crime-matched pairs, which are identical across models since matching is on crime only.

n=109

. Every name-shift, demographic-coefficient, and matched-pair interval excludes zero. Calibration intervals exclude zero for six of seven models; Qwen2.5-7B’s includes zero (

[-0.33,+0.08]

), so it is uncalibrated to Los Angeles crime rather than inverted. Under coordinates-only the matched-pair gap is indistinguishable from zero for six of seven models (

-0.018

+0.047

); only GPT-4o shows one (

-0.34

p=0.015

), consistent with its being the one model whose coordinate channel carries substantial amplitude.

Figure 6. Los Angeles, crime-matched percent-Hispanic pairs. Bars show the within-pair safety-rating gap (higher-Hispanic minus lower-Hispanic neighborhood) under coordinates-only (light) and the name (dark). The gap appears only with the name, and does so for every model.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Template	Varies	Scale
base	—	1–10 safe-high
reversed	scale direction	1–10 danger-high
danger_parap.	paraphrase	1–10 danger-high
safe_parap.	paraphrase	1–10 safe-high
scale_5	granularity	1–5 safe-high
scale_100	granularity	0–100 safe-high

Appendix

Table A1. The ten prompt-robustness templates.

Covariate	Before matching	After matching
Violent-crime rate	$+2.32$	$+0.31$
Mean household income	$-1.59$	$-0.55$

Appendix

Table A2. Standardized mean differences between the two groups (higher-share minus lower-share), Chicago. Values near zero mean the groups are comparable;

0.25

is a common threshold for acceptable balance. Matching cuts the crime imbalance roughly sevenfold, to just above that threshold, and halves the income imbalance as a side effect—but

-0.55

is still far above it. This design controls for crime, not for income, consistent with Limitation (iii).

%Black threshold		Caliper	Pairs	Crime gap	L3.1-8B	L3.2-3B	Q-14B	Q-7B	DS-v4	4o-mini	4o
low $\leq$	high $\geq$	(SD)	( $n$ )	(per 1k)
25	40	0.50	8	28.5	$-0.186$	$+0.028$	$+0.137$	$-0.070$	$-0.117$	$+0.088$	$-0.850$
30	35	0.75	13	53.5	$-0.349$	$-0.030$	$-0.292$	$-0.450$	$-0.714$	$-0.316$	$-1.606$
20	45	1.00	15	102.4	$-0.105$	$-0.014$	$-0.158$	$-0.161$	$-0.395$	$-0.248$	$-1.138$
25	40	1.00	17	94.3	$-0.156$	$\ \ 0.000$	$-0.180$	$-0.262$	$-0.426$	$-0.288$	$-1.127$
25	40	1.50	23	139.0	$-0.358$	$-0.049$	$-0.437$	$-0.313$	$-0.777$	$-0.495$	$-1.576$

Appendix

Table A3. Mean within-pair rating gap (higher-share minus lower-share) under the name-only condition, across five matching settings in Chicago. Negative means the higher-share neighborhood is rated less safe despite matched crime. Neighborhoods falling between the two thresholds are excluded from matching. “Crime gap” is the mean within-pair difference in violent-crime rate—the quantity matching is meant to eliminate. The first row is the headline setting. Relaxing either knob yields more pairs but worse matching: the crime gap grows from

28.5

139.0

per

1{,}000

, so the lower rows are better powered yet weaker tests. We report them to show the direction of the gap does not depend on where either line is drawn; at

n=8

the headline row cannot resolve effects of this size.

Model	$d^{\prime}$ Chicago	$d^{\prime}$ Los Angeles
Llama-3.2-3B	2.50	2.15
Llama-3.1-8B	2.02	2.27
Qwen2.5-7B	0.91	0.45
Qwen2.5-14B	2.71	3.17
GPT-4o-mini	3.23	3.85
GPT-4o	4.56	4.82

Appendix

Table A4. Recognition

d^{\prime}

(real versus fabricated neighborhood names) by model and city. Higher

d^{\prime}

means better separation of real from fabricated names. The measure spreads the seven models widely (

0.5

–

4.8

) and is computed independently of the safety ratings. Across models it correlates with the name shift at

-0.71

in Chicago (

n=7

, exact

p=0.088

) and

-0.75

in Los Angeles (

p=0.066

	Chicago		Los Angeles
Model	hom $\mid$ disc	disc $\mid$ hom	hom $\mid$ disc	disc $\mid$ hom
Llama-3.2-3B	$-0.18$	$-0.06$	$-0.39$	$+0.10$
Llama-3.1-8B	$-0.36$	$+0.05$	$-0.57$	$+0.18$
Qwen2.5-7B	$-0.27$	$+0.05$	$-0.18$	$+0.13$
Qwen2.5-14B	$-0.19$	$-0.22$	$-0.45$	$+0.10$
GPT-4o-mini	$-0.40$	$+0.04$	$-0.57$	$+0.20$

Appendix

Table A5. Does caution track violence or enforcement? Partial Spearman correlations between the safety rating and each rate, holding the other rate constant. The homicide association is negative in every model and every city. The discretionary association is weaker and inconsistent in sign, near zero for the open models and positive for GPT-4o, so it does not mirror the homicide term. The one exception is Qwen2.5-14B, whose two Chicago partials are comparable (

-0.19

vs.

-0.22

). Adding the discretionary tier to a homicide-only model lifts

R^{2}

by at most

0.047

in Chicago and

0.014

in Los Angeles. The Chicago rates are collinear (

r=0.89

), so their separation is imperfect; in Los Angeles they are nearly independent (

r=0.18

Figure A1. Los Angeles prompt-robustness. GPT-4o-mini’s calibration and name shift stay negative across templates; the open models are noisy or degenerate, as expected where the city is unknown.

Figure A2. GPT-4o-mini over-caution using the homicide rate as the (enforcement-robust) risk proxy. The Chicago South/West-Side over-caution band persists, confirming the pattern is not an artifact of the violent-crime measure.

Jurisdictional Bias in Language Models

See all 9 papers

Apr 21, 2026

Culturally uneven urban perception in large language models

Large language models (LLMs) are increasingly used to describe and evaluate cities, yet the cultural structure of their urban judgments remains understudied. Here we introduce a measurement framework for testing whether LLM-based urban perception is culturally neutral, using a globally stratified street-view image dataset. Open-ended descriptions and structured scores generated by three frontier multimodal models all show that the neutral baseline lies closer to regional framings associated with Europe and North America than to other cultural framings. Comparisons between AI and human urban perception further show that prompting can move AI responses closer to specific regional human descriptions, but fails to recover the variety and diversity of human responses, flattening observed demographic patterns and introducing sentiment-based self-favouring bias. These results indicate a systematic risk in treating AI as a neutral tool for urban tasks, especially when model outputs are used to compare, evaluate or represent cities across cultural contexts.

Rong Zhao, Wanqi Liu, Zhizhou Sha +3

Centre for Advanced Spatial Analysis (CASA), UCL, London, UK · Department of Computer Science, UT Austin, Austin, TX, USA · School of Architecture, Tsinghua University, Beijing, China

VLM Evaluation Social Bias in Language Models Cultural Bias in Language Models Cultural Knowledge in Language Models

Ask PDF

Mar 14, 2026

Large Language Models Reproduce Racial Stereotypes When Used for Text Annotation

Large language models (LLMs) are increasingly used for automated text annotation in tasks ranging from academic research to content moderation and hiring. Across 19 LLMs and two experiments totaling more than 4 million annotation judgments, we show that subtle identity cues embedded in text systematically bias annotation outcomes in ways that mirror racial stereotypes. In a names-based experiment spanning 39 annotation tasks, texts containing names associated with Black individuals are rated as more aggressive by 18 of 19 models and more gossipy by 18 of 19. Asian names produce a bamboo-ceiling profile: 17 of 19 models rate individuals as more intelligent, while 18 of 19 rate them as less confident and less sociable. Arab names elicit cognitive elevation alongside interpersonal devaluation, and all four minority groups are consistently rated as less self-disciplined. In a matched dialect experiment, the same sentence is judged significantly less professional (all 19 models, mean gap

-0.774

), less indicative of an educated speaker (

-0.688

), more toxic (18/19), and more angry (19/19) when written in African American Vernacular English rather than Standard American English. A notable exception occurs for name-based hireability, where fine-tuning appears to overcorrect, systematically favoring minority-named applicants. These findings suggest that using LLMs as automated annotators can embed socially patterned biases directly into the datasets and measurements that increasingly underpin research, governance, and decision-making.

Petter Törnberg

Institute for Logic, Language and Computation (ILLC), University of Amsterdam„ Science Park 107, 1098 XG Amsterdam, Netherlands

Social Bias in Language Models Demographic Bias in Language Models LLM-Assisted Annotation Linguistic Bias in Language Models Language Model Bias Evaluation

Ask PDF

Aug 31, 2026

Do LLMs Know Your Neighborhood? Auditing LLM Priors for Neighborhood-Level Mobility Prediction and Structural Alignment

Human mobility is central to urban planning, transportation, public health, and emergency response, yet fine-grained trajectory data are often proprietary, restricted, and privacy-sensitive. Large language models (LLMs) offer a potential alternative by generating plausible mobility traces and predicting individual movement, but their ability to infer aggregate neighborhood-level mobility remains unclear. We evaluate zero-shot LLMs on Census Block Group-level mobility prediction across four U.S. metropolitan areas using anonymized Cuebiq data to construct point-level, trajectory-level, and temporal mobility outcomes, paired with sociodemographic and built-environment predictors. We compare LLM predictions with supervised baselines and introduce a directional alignment analysis to test whether LLM-implied predictor effects agree with empirical OLS and Jonckheere-Terpstra trends. Supervised models achieve 0.580 average accuracy, compared with 0.435 for the best LLM, with spatial extent outcomes showing the strongest predictability but also the largest LLM-baseline gaps. Directional analysis shows that LLMs often rely on coarse, stable predictor-level priors that remain similar across outcomes and cities, including asymmetric treatment of protected-group predictors. Overall, LLMs can partially recover aggregate mobility patterns from urban context, but their predictions should not be treated as structurally grounded without auditing empirical alignment and potential bias.

Saad Mohammad Abrar, Eesha Kurella, Arnav Dadarya +3

University of Maryland USA

LLM Auditing Human Behavior Prediction Demographic Bias in Language Models

Ask PDF

Abstract

Jurisdictional Bias in Language Models

Explore similar work

Culturally uneven urban perception in large language models

Large Language Models Reproduce Racial Stereotypes When Used for Text Annotation

Do LLMs Know Your Neighborhood? Auditing LLM Priors for Neighborhood-Level Mobility Prediction and Structural Alignment