Research map / Large language models / Language, safety and interpretability

Large Language Models are Geographically Biased

ICML 2024 · Large language models › Language, safety and interpretability · approach: Bias measurement and mitigation · 13 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes zero-shot geospatial rating prompts for LLMs and a bias score to measure geographic bias on objective and sensitive subjective topics, finding LLMs rate low-socioeconomic regions (e.g., Africa) lower on attractiveness, morality, and intelligence.

Problem

LLMs carry biases from training corpora, but biases tied to location are not systematically measured; prior geospatial LLM work used fine-tuning, which confounds bias evaluation.

Approach

Uses coordinate-based zero-shot prompts asking LLMs to rate locations on topics (objective and sensitive subjective), uses expected value from logprobs for finer ratings, and defines a bias score combining Spearman correlation with an anchor distribution (infant survival), mean absolute deviation, and answer rate.

Key results

• Zero-shot LLM ratings reach Spearman's rho up to 0.89 on average temperature and 0.85 on infant mortality rate, using GPT-4 Turbo and related models.
• On sensitive subjective topics, ratings correlate with infant survival rate up to Spearman's rho 0.70, with LLMs rating Africa lower on attractiveness, morality, and intelligence.
• GPT-4 Turbo and Llama 2 70B are reported as least biased by the bias score, while Gemini Pro is significantly more biased.

Measured effects reported in the paper

Surprising findings

• Removing the full address from the prompt drops Spearman's rho on subjective topics from 0.50 to 0.34, indicating sensitivity to prompt content.
• Subtle biases become visible only when using logprob expected values, since discrete ratings on sensitive topics are nearly constant (answer rate near 1.0).

Limitations

• Hard to produce a single comprehensive geographic bias score; the topic set may not be exhaustive.
• Some geographically independent topics may not be fully independent of geography.
• Bias score depends on choice of anchor distribution (infant survival rate is recommended default).

Experimental setup

Datasets: WorldPop population density infant mortality rate ground truth built-up area ratio nighttime light intensity average temperature annual precipitation

Benchmarks: infant mortality rate population density built-up to non-built-up area ratio nighttime light intensity average temperature annual precipitation sensitive subjective topics (likability, attractiveness, morality, intelligence, work ethic)

Metrics: spearman's rho mean absolute deviation answer rate bias score gini coefficient

Baselines: GeoLLM (fine-tuned GPT-3.5) K2

Builds on: GeoLLM

Closest papers by meaning

• GeoLLM: Extracting Geospatial Knowledge from Large Language Models (ICLR 2024 · similarity 0.89)
• Investigating and Mitigating Undesirable Biases in Large Language Models (AAAI 2025 · similarity 0.88)
• StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs (NeurIPS 2026 · similarity 0.87)
• Examining Alignment of Large Language Models through Representative Heuristics: the case of political stereotypes (ICLR 2025 · similarity 0.86)
• Bias and Volatility: A Statistical Framework for Evaluating Large Language Model's Stereotypes and the Associated Generation Inconsistency (NeurIPS 2024 · similarity 0.86)
• SocialStigmaQA: A Benchmark to Uncover Stigma Amplification in Generative Language Models (AAAI 2024 · similarity 0.86)