Large language models perceive cities through a culturally uneven baseline

1. Problem Statement

Large Language Models (LLMs) are increasingly deployed to describe, evaluate, and interpret urban environments. However, it remains unclear whether these models operate from a culturally neutral standpoint or if they inherit biased reference frames from their training data and alignment procedures.

• The Gap: Existing research on LLM bias is often text-based and detached from concrete spatial stimuli, while urban perception research focuses on predictive accuracy rather than the cultural organization of the perceptual baseline.
• The Core Question: Do LLMs perceive cities from a "view from nowhere," or do they rely on a culturally uneven baseline where "neutrality" is actually a proxy for specific cultural standpoints (e.g., Western)? Furthermore, does prompting a model with a specific cultural identity genuinely shift its perception to match human diversity, or does it merely create a superficial alignment?

2. Methodology

The authors employed a rigorous, multi-stage experimental design combining a global image dataset with two complementary task formats (open-ended and structured) across three frontier LLMs.

A. Data Construction

• Global Image Corpus: A merged street-view dataset of 3,000 balanced scenes was assembled from multiple providers (Google, Baidu, etc.).
• Stratification: Images were stratified across visual scene types, place types, countries, and providers to ensure the sample did not collapse into a narrow subset of recurring streetscapes.
• Prompt Conditions: Each image was evaluated under 8 prompt conditions:
  1. Neutral Baseline: No cultural identity specified.
  2. 7 Meso-Regional Cultural Prompts: Based on UN SDG reporting groups (e.g., Europe & Northern America, Sub-Saharan Africa, Latin America & Caribbean, etc.).
• Models Tested: GPT-5.2, Claude Sonnet 4, and Gemini 2.5 Flash.
• Total Output: 72,000 open-text descriptions and 72,000 structured evaluations.

B. Task Formats

1. Study 1: Open-Ended Perception

   • Models generated short free-text paragraphs describing the scene.
   • Metrics:
     • Semantic Distance: Cosine distance between the embedding of the regional prompt response and the neutral prompt response.
     • Ingroup Preference Index (IPI): A sentiment-based metric comparing how positively a region-matched prompt evaluates scenes from its own region versus how other prompts evaluate them.
   • Human Benchmark: Comparison against 1,000 human text-image pairs from Geograph Britain and Ireland to measure alignment and diversity.

2. Study 2: Structured Judgments

   • Models rated scenes on a 0–100 scale across six dimensions: Safe, Lively, Wealthy, Beautiful, Boring, Depressing.
   • Metrics:
     • Standardized Shifts: The deviation of regional scores from the neutral baseline.
     • External Grounding: Comparison against a Place Pulse-style machine learning model trained on human data.
     • Pairwise Replication: Testing if models could reproduce human subgroup differences (gender, age, country) in pairwise comparisons.

3. Key Contributions

• Reframing "Neutrality": The paper demonstrates that the "neutral" prompt in LLMs is not a culturally invariant baseline but is systematically closer to Europe and Northern America (ENA) cultural standpoints.
• Dual-Mode Bias: The authors identify two distinct forms of cultural bias:
  1. Semantic Bias: Non-Western prompts require larger semantic shifts to reach the "neutral" baseline.
  2. Affective Bias: Culturally proximate prompts induce an "ingroup preference," evaluating their own regions more positively.
• Limitations of Contextual Prompting: The study shows that while culturally proximate prompting improves alignment with human descriptions, it fails to recover human levels of semantic diversity and often preserves an artificially elevated positive sentiment.
• Structured vs. Open Perception: The cultural bias persists even when the task is reduced to structured scoring, proving the bias is embedded in the model's latent representation of place, not just its generative style.

4. Key Results

A. The Uneven Neutral Baseline

• Semantic Proximity: Across all three models, the Europe and Northern America (ENA) prompt was consistently the closest to the neutral baseline (Mean Cosine Distance ~0.137).
• Displacement: Non-Western regions (e.g., Latin America, Oceania, Sub-Saharan Africa) showed significantly larger distances from neutrality (up to 0.192).
• Geometry: In semantic embedding space, regional responses did not cluster randomly around the neutral point but moved along structured, identity-specific trajectories.

B. Affective Ingroup Preference

• Models exhibited Ingroup Preference where prompts associated with specific regions evaluated scenes from those regions more positively than other prompts did.
• Example: Claude Sonnet 4 showed the strongest self-favoring tendency, particularly in Northern Africa, Western Asia, and Sub-Saharan Africa.
• Prompt Sensitivity: Changing the wording from role-playing to context-based prompts reduced the affective bias but did not eliminate the semantic ordering bias.

C. Human Benchmark Comparison (Geograph)

• Alignment vs. Diversity: Using a UK-specific prompt reduced the semantic distance to human descriptions (improvement of 0.005–0.016 cosine units). However, human descriptions were substantially more diverse (Mean distance to centroid: 0.734) than model outputs (0.391–0.418).
• Sentiment Gap: Human texts had a mean sentiment score of 0.387, while model outputs ranged from 0.823 to 0.975, indicating models describe places as uniformly more positive and less nuanced.
• IPI Discrepancy: Human texts showed negative IPIs (critical of their own regions), whereas models showed positive IPIs (favoring their own regions).

D. Structured Judgments & External Grounding

• Dimension Sensitivity: Perceptions of Wealth and Safety were the most sensitive to cultural framing, while Beauty was more stable.
• Reproduction of Human Differences:
  • Models could approximate human differences for Gender and Age (divergence rates 0.02–0.11).
  • Models failed to reliably reproduce Country-based differences (divergence rates 0.23–0.40), suggesting they struggle with complex geopolitical/cultural nuances compared to demographic ones.

5. Significance and Implications

• Urban Analytics & Planning: If LLMs are used as "general-purpose interpreters" of place in urban planning or design, their outputs may be mistaken for neutral facts. The paper warns that these models privilege Western urban imaginaries, potentially leading to patterned misreading of non-Western cities.
• AI as Cultural Technology: The findings support the view that LLMs are not just prediction engines but cultural and social technologies that carry historically uneven reference frames into downstream applications.
• Bias Mitigation: Simply "prompting" a model to be local does not make it equivalent to lived local perception. The models compress human diversity and maintain an affectively elevated style regardless of the prompt.
• Future Directions: The study calls for moving beyond simple bias detection to understanding the structural organization of perceptual baselines in AI, emphasizing the need for culturally situated evaluation frameworks in AI development.

Conclusion

The paper concludes that LLMs do not perceive cities from a "view from nowhere." Instead, they interpret urban environments through a culturally organized prior that is systematically biased toward Europe and Northern America. This bias manifests as a semantic distance from neutrality for non-Western regions and an affective preference for one's own cultural context, a pattern that persists across both open-ended generation and structured scoring tasks.