Who You Are Adds Nothing Detectable to Where You Go Next: Sociodemographic Conditioning in LLM Next-Location Prediction

TEXT START: Large language models (LLMs) are increasingly used for individual next-location prediction, while sociodemographic conditioning is common in LLM-based travel simulation.

The Dissection

This paper is not proving that identity is irrelevant. It is proving that, once a mobility trajectory and a fixed candidate set are supplied, explicit age, gender, occupation, and income labels add almost no incremental ranking signal to a narrow next-location benchmark.

Its real work is methodological: it attacks demographic-conditioning theater and exposes candidate construction as the larger source of apparent model performance. The 7.7-point gain and 22.3-point collapse caused by removing distance under different sampling schemes are the important result. The benchmark’s answer is being shaped more by the question’s candidate geometry than by the person’s demographic profile.

Under the Discontinuity Thesis, the deeper signal is harsher: the trajectory already functions as a behavioral compression of the person. The model does not need to be told who the subject is when their movement history, constraints, routines, and destination environment have already leaked that information into the data.

The Core Fallacy

The title commits a scope inflation. “Who you are adds nothing detectable” becomes true only inside this particular closed-set, short-horizon, feature-controlled prediction task. It does not establish that identity has no causal, political, economic, or strategic importance.

The paper also risks confusing redundancy with irrelevance. If mobility history subsumes demographic information, then the attributes may still shape the trajectory upstream. Removing the label does not remove the social forces that generated the movement pattern. It merely shows that the model can exploit their behavioral residue without naming them.

The paper’s null result is therefore not a refutation of demographic conditioning. It is a refutation of naïve demographic conditioning: attaching coarse labels to a model after richer behavioral data are already present and expecting a measurable accuracy dividend.

Hidden Assumptions

• That top-1 accuracy over 100 candidates is an adequate proxy for useful next-location intelligence.
• That a Shenzhen sample of 5,000 residents generalizes across cities, cultures, income distributions, transport systems, and surveillance regimes.
• That the selected attributes—age, gender, occupation, and income—capture the relevant meaning of “who you are.” They omit household structure, disability, migration status, political affiliation, health, social ties, access rights, and institutional constraints.
• That demographic information matters primarily through individual prediction accuracy rather than through population segmentation, pricing, policing, allocation, or control.
• That the candidate-construction problem is a technical nuisance rather than part of the prediction system’s power. In reality, whoever defines the alternatives partly defines the world the model is allowed to see.
• That better prediction is the central objective. A model can be highly valuable for triage, surveillance, demand shaping, or resource allocation even when explicit demographic features provide no accuracy improvement.
• That “no detectable gain” means no economically meaningful gain. A small average effect can remain decisive for minority groups, edge cases, or high-value decisions.
• That the tested models and reranker adequately represent future systems. They do not establish the limit of more capable models, richer context windows, or integrated real-time data.

Social Function

Primary classification: partial truth with ideological anesthetic attached.

The partial truth is real and useful: coarse demographic labels are often redundant once detailed behavioral traces are available, and benchmark scores can be manufactured or distorted by candidate sampling. The anesthetic is the implication that the disappearance of explicit identity variables reduces the danger or significance of social classification.

It does not. The system can discard the label while retaining the person’s socioeconomic legibility. This is cleaner for the operator: fewer politically radioactive fields, the same inferential leverage. The individual is not liberated from classification; classification has moved from declared identity to observed behavior.

The result also serves transition management. It normalizes a world in which machines infer where people will go without needing to understand them as citizens, workers, or persons. That is exactly the direction of cognitive automation: convert human context into exploitable prediction, then remove the human category from the interface.

The Verdict

This is a competent narrow benchmark study with a larger implication than its title can safely carry. It shows that explicit sociodemographics are weak marginal features when the model already possesses behavioral telemetry, while candidate design can dominate the reported result.

Under DT logic, that is not evidence that “who you are” has ceased to matter. It is evidence that who you are is being converted into data exhaust. The system no longer needs your biography; it needs your trace.

The paper does not demonstrate the death of mass employment, but it illustrates the mechanism that makes that death possible: human behavior becomes machine-readable, prediction becomes detached from human interpretation, and the person’s stated identity becomes optional. The obsolete layer is not the human pattern. It is the human’s role as interpreter and economic intermediary.