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

TEXT START: 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.

The Dissection

This is an audit, not a breakthrough. It tests whether zero-shot LLMs can infer block-group mobility from urban context, then measures them against supervised models and empirical directional trends. The result is narrow but useful: LLMs recover coarse urban priors, yet those priors remain stubbornly similar across cities and outcomes and treat protected-group predictors asymmetrically.

The paper’s real function is to attach a warning label to synthetic mobility inference before institutions mistake plausibility for knowledge.

The Core Fallacy

It locates the main danger in prediction error and structural misalignment. The Discontinuity Thesis locates the decisive issue in scalable substitution and control.

A best-LLM accuracy of 0.435 versus 0.580 for supervised models is a competence gap, not a defense of human economic relevance. An LLM does not need census-grade accuracy to displace analysts, reduce demand for proprietary data collection, or support cheap administrative guesses. It only needs to be inexpensive, scalable, and acceptable to whoever controls deployment.

The paper also treats alignment with OLS and Jonckheere-Terpstra trends as evidence of structural grounding. That is weaker than it sounds. Agreement with existing statistical regularities can mean conformity to inherited measurement systems—or reproduction of demographic stereotypes—not causal understanding.

Hidden Assumptions

• That supervised accuracy represents neutral ground truth rather than the advantage of whoever already owns better data and infrastructure.
• That auditing model priors can contain the problem without confronting control over deployment.
• That weak aggregate predictions will remain too weak to matter institutionally.
• That neighborhood mobility can be adequately represented by the selected outcomes and predictors.
• That four metropolitan areas and Cuebiq-derived measures generalize to broader urban systems.
• That protected-group asymmetry is merely a bias to be corrected, rather than evidence that models are importing socially entrenched classifications into automated allocation.
• That the relevant question is whether LLMs “know” neighborhoods, rather than whether they can cheaply produce actionable approximations.

Social Function

Primary classification: partial truth and transition management. Secondary classification: ideological anesthetic.

The paper correctly punctures the fantasy that an LLM possesses reliable, fine-grained urban knowledge. But its framing risks making auditing look like containment. The system does not need to become epistemically pure before it becomes economically useful. Technical warnings can coexist with deployment, especially when institutions value speed, cost reduction, and plausible outputs over causal validity.

The Verdict

The empirical warning is valid: current LLMs are inferior to specialized supervised systems and recycle coarse, potentially biased priors. But this is not a rebuttal of the Discontinuity Thesis. It is reconnaissance on an immature instrument.

The paper shows that LLMs do not yet replace high-quality mobility models. It does not show that human-only economic domains can be preserved. Under DT logic, the danger is not that the model is omniscient; it is that approximate inference becomes cheap enough to automate decisions once performed by people. The paper’s warning label is accurate. Its implicit comfort is not.