💻 computer science

Large language models create an uneven informational layer over cities

This study reveals that large language models create an uneven informational layer over cities by systematically fabricating venues in digitally sparse areas and selectively ignoring nearly half of real establishments, thereby redistributing economic visibility and revenue in ways that favor independent dining while exacerbating urban inequality across different neighborhoods and user demographics.

Original authors: Lin Chen, Guangyuan Weng, Esteban Moro

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lin Chen, Guangyuan Weng, Esteban Moro

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a city as a giant, bustling marketplace. For a long time, if you wanted to find a place to eat, you had to rely on your own memory, a physical map, or a list of reviews written by other people. These tools showed you everything that existed, even if some places were at the bottom of the list. You could scroll down and see the hidden gems, the chain restaurants, and the local favorites alike.

Now, imagine a new, invisible layer of glass has been placed over this marketplace. This glass is made of Large Language Models (LLMs)—the smart AI chatbots we use to ask questions like, "Where should I eat in Boston?"

This paper investigates what happens when we look at the city through this new glass. The researchers found that this glass doesn't just show us the city; it actively reshapes it, often in unfair and uneven ways. Here is what they discovered, broken down into simple concepts:

1. The "Ghost" Restaurants (Making Things Up)

Sometimes, when you ask the AI for a restaurant, it invents places that don't actually exist.

The Analogy: It's like a tour guide who, when asked about a small, quiet neighborhood, confidently points to a fancy café that was never built.
The Finding: The AI makes up these "ghost" restaurants most often in neighborhoods that are poorer, less populated, or have fewer people online. If a neighborhood doesn't have a strong "digital footprint" (few reviews, fewer people visiting), the AI is more likely to hallucinate a fake place there.
The Fix: When the researchers forced the AI to choose only from a list of real restaurants, the fake ones disappeared. This means the AI was just guessing because it didn't have the right data, not because it was trying to be mean.

2. The "Invisible" Restaurants (Ignoring the Real Ones)

Even when the AI is forced to pick from a list of real restaurants, it still ignores a huge number of them.

The Analogy: Imagine a menu with 100 dishes. The AI only recommends 3 or 4 of them to everyone, no matter who is asking. The other 96 dishes are completely invisible, as if they don't exist.
The Finding: About 47.5% of real restaurants were never recommended by the AI, even when the AI knew they were there. This isn't just a mistake; it's a pattern. The AI consistently ignores certain types of places, especially in specific neighborhoods.
The Shared Blind Spot: The researchers tested three different AI models (from different companies). They found that 31.9% of the ignored restaurants were the same for all three models. This suggests the bias isn't just a glitch in one company's code; it's a shared habit learned from the same pool of internet data.

3. The "Personalized" Lens (Different People, Different Cities)

The most surprising part is that the AI shows a different city to different people, even if they are standing in the same spot.

The Analogy: Imagine two people standing at the same corner. One is a wealthy tourist, and the other is a local resident with a lower income.
- The wealthy tourist is shown expensive, trendy, and far-away places that are less crowded.
- The local resident is shown cheaper, very popular, and closer places.
- Children are shown mainstream, low-cost places.
- Men and women are also shown slightly different types of places.
The Finding: The AI doesn't just answer the question; it guesses who you are based on your description and then filters the world to match what it thinks people like you usually do. It assumes wealthy people like to travel far for unique food, while locals stick to familiar, cheaper spots.

4. The Economic Ripple Effect (Who Wins and Who Loses)

If everyone started trusting these AI recommendations, the money in the city would flow differently.

The Analogy: Think of the city's economy as water flowing through pipes. The AI acts like a valve that redirects the water.
The Finding: The AI tends to push people away from fast-food chains (like McDonald's) and quick-service cafes and toward independent, full-service restaurants (like local sit-down diners).
- Independent restaurants would likely get more customers and money.
- Big chains would lose a significant amount of revenue.
- Fast food spots, which often serve price-sensitive customers, would be the biggest losers.

The Big Picture

The paper concludes that these AI models are acting like a selective filter for our cities. They aren't just showing us what is there; they are deciding what is visible.

For places: If a restaurant is in a poor neighborhood or doesn't have many online reviews, the AI might pretend it doesn't exist.
For people: The AI reinforces stereotypes about who goes where, potentially keeping different groups of people in their own "bubbles" of the city.

The authors warn that if we rely on these tools too much, we might accidentally create a city where some neighborhoods and some types of businesses become invisible, not because they are bad, but because the AI decided not to show them. It's a new kind of "digital inequality" layered right on top of our physical streets.

Technical Summary: Large Language Models Create an Uneven Informational Layer Over Cities

Problem Statement

Cities are increasingly mediated by information systems that guide human perception and decision-making. While GPS, search engines, and review platforms have historically shaped urban navigation, Large Language Models (LLMs) are emerging as a new, structurally distinct informational layer. Unlike predecessors that offer ranked lists where lower-ranked options remain discoverable, LLMs generate responses where venues are either included or entirely absent from a user's consideration set, with no signal indicating omission.

The core problem addressed is whether this emerging layer introduces systematic biases in urban visibility. Specifically, the authors investigate:

Fabrication: Do LLMs invent non-existent venues?
Invisibility: Do LLMs systematically overlook real venues?
Selectivity: Do these patterns vary across neighborhoods (socioeconomic gradients) and user demographics (income, age, sex, residency)?
Economic Consequence: How would widespread reliance on LLM recommendations redistribute consumer demand and revenue across the urban food economy?

The study distinguishes between two potential failure modes: epistemic failure (omissions due to missing data, correctable by providing verified lists) and allocative failure (omissions due to stable, selective attention patterns rooted in the model's training, persisting even with complete data).

Methodology

The authors conducted a systematic audit of restaurant recommendations across 304 neighborhoods in five major U.S. cities: Boston, New York, Chicago, Houston, and San Francisco.

Experimental Design

Models: Three representative LLMs from distinct families were tested: GPT-4o-mini (OpenAI), Llama-3.3-70b-instruct (Meta), and Gemini-2.0-flash (Google).
User Profiles: A factorial design generated 320 synthetic user profiles per neighborhood, crossing four demographic dimensions:
- Household Income (5 levels: $0–$ 25k to $200k+)
- Age (8 levels: 0–20 to 80+)
- Sex (2 levels: Male, Female)
- Residential Status (4 levels: Local, Neighboring, Cross-city, Tourist)
Settings: Two parallel experimental conditions were employed:
1. Open-Ended: Models queried for 10 restaurants in a specific neighborhood without a pre-defined list of candidates.
2. Candidate-Constrained: Models were provided a fixed set of 100 real venues and instructed to select exactly 10. This isolated "selective attention" from "knowledge gaps."

Data Integration and Metrics

Ground Truth: Verified venue records from SafeGraph and Foursquare.
Metrics:
- Hallucination Rate: The fraction of surfaced venues that could not be matched to real establishments.
- Neighborhood Algorithmic Invisibility Rate (NAIR): The fraction of real venues in a neighborhood that received zero recommendations across all user profiles.
Covariates: Neighborhood characteristics were derived from census data, mobile location data (Spectus), transaction records (SafeGraph), and review metadata (Yelp).
Economic Simulation: A counterfactual scenario was modeled where 10% ( $\alpha=0.1$ ) of dining decisions shifted from observed patterns to LLM-recommended venues to estimate revenue redistribution.

Key Contributions and Results

1. Uneven Fabrication and Grounding Failure

LLMs fabricate venues at an average rate of 36.8%. This "grounding failure" is not random; it correlates with neighborhood socioeconomic characteristics.

Drivers: Hallucination rates are significantly higher in neighborhoods with weaker digital and physical footprints (fewer Yelp reviews, fewer real-world visitations, lower population density, and lower educational attainment).
Correction: When models are constrained to a verified list of venues (Candidate-Constrained setting), hallucination rates drop to 0.0%, confirming that fabrication is an epistemic failure remediable by data grounding.

2. Persistent Selective Attention (Allocative Failure)

Even when restricted to real venues, LLMs exhibit severe "coverage deficits."

Invisibility: In the open-ended setting, 96.8% of real venues in a neighborhood are never recommended. In the candidate-constrained setting, 47.5% of the available real venues are still never surfaced.
Shared Blind Spots: 31.9% of overlooked venues are shared across all three model families, suggesting these biases are rooted in shared training corpora rather than model-specific errors.
Drivers of Invisibility: Unlike fabrication, invisibility is not solely driven by data scarcity. Even with fixed venue pools, NAIR is predicted by Yelp review counts (negative correlation) and, surprisingly, positively correlated with income entropy and population density in constrained settings. This supports the allocative failure hypothesis: models possess stable priors about which venues merit attention, independent of data availability.

3. Demographic Stratification

Within the same venue pool, LLMs systematically steer different user profiles toward distinct types of venues:

Income: Higher-income users are directed toward venues that are farther away, less visited, more expensive, and patronized by more socioeconomically homogeneous (segregated) clientele.
Age: Minors are directed toward highly visited, lower-priced, and less segregated venues, while adults face a different distribution.
Residency: Tourists are channeled toward costlier but more socially diverse establishments, whereas local residents are directed toward cheaper, more frequented, but socioeconomically homogeneous venues.
Mechanism: These gradients suggest LLMs infer latent urban lifestyles from demographic descriptors alone, reinforcing existing lifestyle stratifications.

4. Economic Redistribution

Simulating a shift in consumer demand reveals a structural redistribution of revenue:

Format Shift: Spending flows away from fast food and quick-service formats toward independent, full-service dining.
Brand Shift: Visits flow predominantly from chain restaurants (e.g., McDonald's, Starbucks) toward independent establishments.
Magnitude: Independent restaurants gain an estimated average of ~$12,101 per store annually, while major chains like McDonald's lose ~$41,405 per store under the simulated scenario.

Significance and Claims

The paper argues that LLMs act as a selective layer of urban information that unevenly distributes visibility across places and people.

Structural Inequality: The informational layer does not merely reflect reality but actively reshapes it by encoding selective attention that tracks socioeconomic gradients. This adds a "representational dimension" to existing structural and experiential inequalities in cities.
Asymmetry of Failure: The study highlights a critical asymmetry: while fabrication (epistemic failure) can be solved by providing verified data, selective invisibility (allocative failure) persists. This implies that simply improving data provision is insufficient; the underlying model behavior and training biases must be addressed.
Dynamic Feedback Loops: The authors posit that these biases could create self-reinforcing feedback loops. Venues that are algorithmically invisible generate fewer visits and reviews, further thinning their digital footprint and making them even less likely to be recommended in future training cycles.
Governance Implications: As LLMs become primary intermediaries for urban decision-making, they function as a new form of infrastructure that allocates visibility. The authors conclude that understanding and governing this layer is as critical as managing traditional transportation and communication infrastructures, as it determines which places and communities thrive or remain unseen.

The study remains modest regarding generalizability, noting that findings are based on dense U.S. urban ecosystems and may differ in smaller cities or non-English contexts. However, the convergence of results across three distinct model families suggests these patterns reflect fundamental properties of the current LLM training ecosystem.