Large language models create an uneven informational layer over cities

arXiv:2607.06260 · cs.CY, cs.SI · Submitted 2026-07-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Large language models create an uneven informational layer over cities".

Title and authors: Tom: So the authors are Lin Chen, Guangyuan Weng, and Esteban Moro. They’re looking at this uneven layer over cities issue head-on >

Jane: The title itself is pretty telling; it suggests that the problem isn't just about bad recommendations, but about a fundamental way these models organize our understanding of urban space >

Lu: It sets up a crucial distinction between what we call epistemic failure, which is when the model just doesn't know something because we didn't tell it >

Meng: Versus allocative failure, which is when the model knows everything but just chooses to ignore some things on purpose, and that distinction matters for how we fix it >

Tom: Exactly. The paper explores whether these uneven results are just missing data that needs adding, or if there's a built-in bias in how they select what to show us >

Lalam: And they suggest it’s both, but the patterns they find point toward something more persistent than just missing facts >

The paper's summary: Jane: So the summary is that when you test these LLMs, you see fabrication happening mostly in neighborhoods with weaker digital and physical footprints >

Tom: They found that if you give the models a verified list of venues, that fabrication disappears almost entirely, dropping the hallucination rates to near zero under candidate constraint >

Lu: That’s one way they test it; grounding failure is epistemic and remediable because providing verified venue data eliminates fabrication completely >

Meng: But then you have this invisibility problem where even when you restrict the choice to only real venues, forty-seven point five percent of establishments are never recommended >

Lalam: And that invisibility persists, meaning thirty-one point nine percent of these blind spots are shared across all three model families they tested >

The paper's improvements: Tom: So the paper suggests a few ways to tackle this uneven layer, starting with how we test it better >

Jane: They improved their method by using parallel experimental settings, one open-ended and one constrained specifically to verified local venues >

Lu: That lets them disentangle the two failure hypotheses—the epistemic gap versus the selective attention—which is a big step forward for understanding why things are wrong >

Meng: And they also found that grounding failure can be fixed by simply providing verified data, which is a pretty clean solution for hallucination >

Lalam: But they also point out that the coverage deficit isn't just a knowledge gap; it’s an allocative and persistent property of the LLM infrastructure itself >

Conclusion: Tom: So wrapping up, the main implication is that LLMs aren't just randomly messing up recommendations; they are systematically filtering information based on neighborhood characteristics >

Jane: And this filtering isn't random either; it’s linked to socioeconomic factors, suggesting a different kind of infrastructure is being built over the city >

Lu: They found that higher-income users get directed toward venues that are simultaneously less frequented and located farther away, which implies the models are inferring a certain lifestyle from the user data >

Meng: That’s interesting because it suggests they're not just recommending what people can afford, but how they are expected to move through the city >

Lalam: And they also showed that algorithmic visibility is tied to real-world visit volume and digital footprint, meaning popularity amplifies existing hierarchies across the food economy >

Tom: So, in short, this paper on "Large language models create an uneven informational layer over cities" shows us that we need to look at these models not just as recommendation tools but as new ways they structure urban visibility >

Lin Chen, Guangyuan Weng, Esteban Moro

Network Science Institute, Northeastern University · Department of Physics, Northeastern University · Khoury College of Computer Sciences, Northeastern University

cs.CY, cs.SI

Submitted: 2026-07-07

Updated: 2026-11-18

Code: https://github.com/SUNLab-NetSI/llm-urban-info-layer

Project page: https://www.census.gov/programs-surveys/acs

License: http://creativecommons.org/licenses/by/4.0/

The gist: Large language models act as an uneven informational layer over cities, shaping which places people discover, consider…

Key concepts

Fabrication vs. Invisibility
LLMs do two types of errors: they invent restaurants that don't exist (fabrication), especially in areas with weak digital presence. They also systematically ignore many real, good places even when given a list of actual venues (invisibility). These patterns are tied to neighborhood socioeconomic factors.
Allocative Failure Hypothesis
This hypothesis suggests LLMs make choices based on how they allocate resources or attention. The study found that the fabrication and overlooking patterns align with neighborhood socioeconomic characteristics, indicating that the models' filtering is driven by these underlying economic and social structures.
Algorithmic Visibility Predictors
LLMs prioritize venues based on their existing digital and physical footprint—like visit volume, spending habits, and review counts—rather than quality ratings. This means popular or well-reviewed places get amplified, creating a hierarchy where existing success dictates future exposure.
Demographic Stratification of Lifestyles
LLMs filter information differently based on user demographics. Higher-income users are directed toward venues that are less frequented and farther away, suggesting the models associate wealth with an exploratory urban lifestyle rather than just affording more expensive options.

Terminology

Summary

Large language models act as an uneven informational layer over cities, shaping which places people discover, consider, and ultimately visit The authors audit restaurant recommendations from three major LLMs across 304 neighborhoods in five U. p. 0

Where the paper says it

edu ABSTRACT Large language models (LLMs) are emerging as a new informational layer over cities, shaping which places people discover, consider, and ultimately visit. Yet little is known about which places they surface, which they ignore, and whether these patterns vary across communities and users and translate into real-world economic consequences.

S. cities using 320 synthetic user profiles spanning income, age, sex, and residential status They find that LLMs both fabricate venues and systematically overlook real ones. p. 0
Where the paper says it

cities using 320 synthetic user profiles spanning income, age, sex, and residential status. We find that LLMs both fabricate venues and systematically overlook real ones.

Fabrication is concentrated in neighborhoods with weaker digital and physical footprints and disappears when models are provided with verified venue lists. p. 0
Where the paper says it

We find that LLMs both fabricate venues and systematically overlook real ones. Fabrication is concentrated in neighborhoods with weaker digital and physical footprints and disappears when models are provided with verified venue lists.

In contrast, invisibility persists: even when choosing from a fixed set of real venues, 47. p. 0
Where the paper says it

Fabrication is concentrated in neighborhoods with weaker digital and physical footprints and disappears when models are provided with verified venue lists. In contrast, invisibility persists: even when choosing from a fixed set of real venues, 47.

5% of establishments are never recommended, and 31. p. 0
Where the paper says it

In contrast, invisibility persists: even when choosing from a fixed set of real venues, 47. 5% of establishments are never recommended, and 31.

9% of these blind spots are shared across all three model families. p. 0
Where the paper says it

5% of establishments are never recommended, and 31. 9% of these blind spots are shared across all three model families, indicating that uneven visibility reflects not only missing knowledge but also stable patterns of selective attention rooted in shared patterns of visibility rather than model-specific errors.

Pruning redundant reasoning tokens

The study investigates how LLMs function as a new informational layer superimposed on the physical city, which differs from predecessors because they do not expose the boundary between what has been included and what has been left out. The core question is whether observed unevenness reflects correctable knowledge gaps or a more persistent form of selective attention. p. 1

Where the paper says it

Systematic biases have been observed in how LLMs represent brands19, cities20, ideological positions21, and demographic groups in job recommendations22, establishing that LLM-generated content can be socially and geographically uneven. What these studies leave unclear is whether the observed unevenness reflects correctable knowledge gaps or a more persistent form of selective attention.

The researchers employed parallel experimental settings—one open-ended and one constrained to verified local venues—to disentangle the two failure hypotheses. p. 1
Where the paper says it

Specifically, we focus on restaurants, because dining is a high-frequency urban decision that spans the full socioeconomic spectrum, with observable downstream consequences through large-scale mobility and transaction data. To disentangle the two failure hypotheses, we employ parallel experimental settings, one open-ended and one constrained to verified local venues, and find that this informational layer is structurally uneven: LLMs fabricate and selectively overlook venues in patterns that track neighborhood socioeconomic characteristics; grounding models in verified data eliminates fabrication but not selective attention, supporting the allocative failure hypothesis.

They conclude that LLMs fabricate and selectively overlook venues in patterns that track neighborhood socioeconomic characteristics, supporting the allocative failure hypothesis when grounded in verified data. p. 1
Where the paper says it

Specifically, we focus on restaurants, because dining is a high-frequency urban decision that spans the full socioeconomic spectrum, with observable downstream consequences through large-scale mobility and transaction data. To disentangle the two failure hypotheses, we employ parallel experimental settings, one open-ended and one constrained to verified local venues, and find that this informational layer is structurally uneven: LLMs fabricate and selectively overlook venues in patterns that track neighborhood socioeconomic characteristics; grounding models in verified data eliminates fabrication but not selective attention, supporting the allocative failure hypothesis.

How the training signal is built

The user profiles used for testing were constructed using a factorial design crossing four demographic dimensions: household income (5 levels), age (8 levels), sex (2 levels), and residential status (4 levels) resulting in 320 unique profiles per neighborhood. The LLM prompts were designed to minimize output variability by using a temperature of 0, producing a single deterministic response for each profile–neighborhood pair. p. 5

Where the paper says it

" The balanced factorial design ensures that each attribute level is observed across all combinations of the other attributes within each neighborhood, reducing the collinearity that can arise when profiles are sampled from census-derived joint distributions. All LLM queries are issued at temperature 0 to minimize output variability, producing a single deterministic response for each profile–neighborhood pair.

In the candidate-constrained setting, the model was instructed to select exactly 10 venues from a fixed set of 100 real venues sampled uniformly without replacement. p. 6
Where the paper says it

In the candidate-constrained setting, each prompt additionally provides a list of 100 real venues from which the model must select exactly 10. To construct the candidate set, we retrieve all food-related POIs within a 5-km Haversine distance of the neighborhood centroid.

Where the gains come from

The results show that grounding failure is epistemic and remediable, as providing verified venue data eliminates fabrication entirely, dropping hallucination rates to near zero under candidate constraint. However, coverage deficit remains persistent because it reflects a property of the LLM itself as an information infrastructure that selectively filters local information. p. 2

Where the paper says it

This indicates that coverage deficit does not arise solely from incomplete knowledge of local venues; it is a property of the LLM itself as an information infrastructure, which selectively filters local information. Moreover, the spatial heterogeneity remains: NAIR varies substantially within each city (Figure 2c; Supplementary Figures S5-S6), with persistently high invisibility concentrated in specific areas even after the venue pool is controlled.

Furthermore, the selectivity extends to users, where higher-income users receive more expensive and less popular venues within identical venue pools. p. 0
Where the paper says it

The same selectivity extends to users. Within identical venue pools, higher-income users receive more expensive and less popular venues, while tourists are directed toward costlier but more socially diverse establishments than local residents.

This suggests LLMs encode a categorical minor-versus-adult distinction, likely reflecting how training corpora associate minors with lower spending contexts. p. 3
Where the paper says it

001 for all three), while these three outcomes remain largely stable from the 20–29 bracket onward. This suggests that LLMs encode a categorical minor-versus-adult distinction, likely reflecting how training corpora associate minors with lower spending and more mainstream consumption contexts.

The redistribution of urban consumption reveals that fast food is the largest source of outflowing visits, with substantial volumes redirected toward American restaurants, bars, and Mexican cuisine. Because these destination categories carry higher per-transaction spending than fast food, the revenue redistribution is amplified beyond what visitation shifts alone would imply. p. 4

Where the paper says it

At the food-category level (Figure 4c–d), fast food is the largest source of outflowing visits, with substantial volumes redirected toward American restaurants, bars, and Mexican cuisine (Figure 4c). Because these destination categories carry higher per-transaction spending than fast food, the revenue redistribution is amplified beyond what visitation shifts alone would imply.

Specifically, seafood and bars gain an average of approximately 95% CI [12, 30] per year per store, while fast food outlets lose approximately [ -14, -54] per year per store. This indicates that LLM-mediated information systematically favors full-service and sit-down dining over quick-service formats. p. 4
Where the paper says it

Seafood and bars gain an average of approximately $21,476 (95% CI: [$12,685, $30,267]) and $18,615 (95% CI: [$15,360, $21,871]) per year per store, respectively, while fast food outlets and cafes lose approximately $13,004 (95% CI: [-$14,455, -$11,553]) and $3,277 (95% CI: [-$6,014, -$540]) (Figure 4d). LLM-mediated information thus systematically favors full-service and sit-down dining over quick-service formats, reflecting a preference structure encoded in the models that diverges from observed consumer behavior.

The study demonstrates that algorithmic visibility is predicted by the volume of a venue’s digital and physical footprint, rather than by evaluative quality such as user ratings. Venues with higher real-world visit volume, greater median spend per transaction, and more Yelp reviews are significantly more likely to be surfaced. p. 3

Where the paper says it

169, p < 0. 001), greater median spend per transaction (β = 0.

Moreover, LLMs amplify existing popularity and digital-visibility hierarchies where venues that are already well-frequented gain further exposure. p. 3
Where the paper says it

335), suggesting that LLMs respond to the volume of online discourse about a venue rather than its evaluative quality. Together, these results indicate that LLMs amplify existing popularity and digital-visibility hierarchies: venues that are already well-frequented, higher-priced, and widely discussed online gain further exposure, while less prominent establishments remain invisible.

If adopted at scale, these tendencies would redistribute consumer spending across the urban food economy. p. 4
Where the paper says it

Venues that are already well-frequented, higher-priced, and widely discussed online gain further exposure, whereas less prominent establishments remain invisible. If adopted at scale, these tendencies would redistribute consumer spending across the urban food economy.

The findings suggest that LLMs do not merely filter urban information unevenly across space; they also filter it unevenly across people, with higher-income users being directed toward venues that are simultaneously less frequented and located farther away. This pattern suggests the models associate higher income with a more exploratory urban lifestyle, inferring not only what users can afford but also how they are expected to navigate and experience the city. p. 3

Where the paper says it

This pattern suggests that the models associate higher income with a more exploratory urban lifestyle26, inferring not only what users can afford, but also how they are expected to navigate and experience the city. Through training data, the models have internalized associations about who goes where in cities and use those associations to filter the urban landscape, potentially reinforcing existing forms of lifestyle stratification if adopted at scale.

The study concludes that LLMs introduce a different form of infrastructure: one that allocates visibility, which may become as important as understanding transportation or communication infrastructures on which cities have traditionally depended. p. 5
Where the paper says it

Cities have long depended on infrastructures that move people, goods, and information. LLMs introduce a different form of infrastructure: one that allocates visibility.

The biases observed reflect properties of the training ecosystem rather than idiosyncrasies of any single model. p. 5
Where the paper says it

In addition, while we examine LLMs from distinct model families, the magnitude of bias may vary across other models not included in this study. That said, the convergence of findings across the examined models, including the shared blind spots and the demographic gradients, suggests that the patterns observed reflect properties of the training ecosystem rather than idiosyncrasies of any single model.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved system can distinguish between epistemic failure (fabrication) and allocative failure (selective attention) in LLM outputs by using parallel experimental settings: one open-ended and one constrained to verified local venues.

  2. Improved system can be grounded to local venues, which eliminates hallucination effectively, ensuring that generated information is factually based on existing establishments.

  3. Improved system can be designed to address coverage deficit by recognizing it as an allocative and persistent property of the LLM infrastructure, rather than just a knowledge gap, even when provided with verified data.

  4. The system can incorporate demographic profiles (income, age, sex, residential status) directly into the recommendation process to understand how LLMs stratify urban information along demographic lines.

  5. Improved system can infer latent urban lifestyles from demographic descriptors alone by mapping user attributes to expected patterns of mobility and consumption, such as how "higher-income users are not simply directed toward more expensive venues; they are directed toward venues that are simultaneously less frequented, located farther away, and patronized by more socioeconomically homogeneous clientele."

  6. The system can utilize venue characteristics like real-world visit volume, median spend per transaction, and Yelp review count as significant predictors for algorithmic visibility, indicating that LLMs amplify existing popularity and digital-visibility hierarchies.

  7. Improved system can implement a counterfactual analysis to project how consumption shifts by converting projected visitation changes into revenue impacts, demonstrating that LLM recommendations systematically favor independent, full-service dining over chain formats.

Abstract

Large language models (LLMs) are emerging as a new informational layer over cities, shaping which places people discover, consider, and ultimately visit. Yet little is known about which places they surface, which they ignore, and whether these patterns vary across communities and users and translate into real-world economic consequences. Here, we audit restaurant recommendations from five LLMs across 304 neighborhoods in five U.S. cities using 320 synthetic user profiles spanning income, age, sex, and residential status, issuing nearly one million queries. We find that LLMs both fabricate venues and systematically overlook real ones. Fabrication is concentrated in neighborhoods with weaker digital and physical footprints and virtually disappears when models are provided with verified venue lists. In contrast, invisibility persists: even when choosing from a fixed set of real venues, 41.5% of establishments are never recommended, and nearly one-fifth of venues are overlooked by all five models, indicating that uneven visibility reflects not only missing knowledge but also systematic patterns of selective attention that are shared across models. The same selectivity extends to users. Within identical venue pools, higher-income users receive more expensive and less popular venues, while tourists are directed toward costlier but more socially diverse establishments than local residents. Simulating the resulting shifts in consumer demand suggests that widespread reliance on LLM recommendations would redirect visits and revenue away from chain and quick-service restaurants toward independent and full-service dining. Together, our findings show that LLMs act as a selective layer of urban information that unevenly distributes visibility across places and people, with potential consequences for local economies and urban inequality.

Sources

Related papers