AI Mobility Models Biased Against Elderly, Study Finds

Smart city development increasingly relies on analyzing trajectory data, but a significant issue arises from the sparse representation of underrepresented groups, particularly the elderly, in public mobility datasets. This demographic imbalance can introduce substantial bias into mobility models and subsequent urban planning initiatives. Researchers used the Jersey City Citi Bike data to quantify how this lack of representation impacts modeling, specifically through synthetic trajectory generation.The analysis demonstrated that elderly riders exhibit distinct mobility signatures compared to younger riders, characterized by more localized activity spaces, lower mobility entropy, and unique off-peak temporal patterns. To illustrate the bias, the study evaluated both a first-order Markov chain and a fine-tuned Qwen3-4B model across different training demographics.Results consistently showed that models trained predominantly on majority populations systematically misrepresented elderly mobility behavior, especially concerning spatial metrics. For instance, the full-population Markov model overestimated elderly step length and dwell time. The study concludes that higher-capability models do not inherently improve subgroup-level fidelity without adequate demographic data, underscoring the critical need for inclusive data representation in mobility modeling for equitable urban applications.