Accuracy as Autonomy: How Reliable Information Enables Choice in Wheelchair-Accessible Taxi Services
Hyunseon Won, Seoyoung Park, Sejung Son, Taenyun Kim, Jinyoung Han · 2026 · Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26) · doi:10.1145/3772363.3798989
Summary
Won and colleagues (Sungkyunkwan University and Michigan State) investigate how the accuracy of AI-predicted information in accessibility-oriented services affects users' psychological needs and technology acceptance, using Seoul's Wheelchair-Accessible Taxi (WAT) service as a case study. Only 39% of Seoul's city buses are wheelchair-accessible, making paratransit essential, yet the official SWAT app's dispatch-time estimates are widely perceived as unreliable — 2023 averages exceeded 40 minutes, with predicted-vs-actual discrepancies ranging from 30 minutes to over two hours in a formative interview study (N=5). The authors build Greum, a prototype Android app that sits alongside SWAT and predicts dispatch time from user-entered departure, destination, and intended reservation time using an XGBoost model trained on Seoul Open Data Plaza WAT call data (MAE 13.12 minutes, substantially better than the existing system's 30-120-minute error). They then evaluate Greum with 41 wheelchair users recruited via two disability organisations, measuring perceived system quality (prediction accuracy, interface adequacy), basic psychological needs (autonomy, competence, relatedness from Self-Determination Theory), and Technology Acceptance Model outcomes (perceived usefulness, perceived ease of use, usage continuance intention) via a 7-point Likert survey. The model is analysed with partial-least-squares structural equation modelling (PLS-SEM, 5000 bootstrap iterations).
Key findings
Prediction accuracy significantly improved all three basic psychological needs — autonomy (beta=0.63, p<.01), competence (beta=0.76, p<.001), and relatedness (beta=0.76, p<.001) — but among the three, autonomy alone mediated the effect on technology acceptance, significantly predicting both perceived usefulness (beta=0.57, p<.01) and perceived ease of use (beta=0.56, p<.05). Interface adequacy showed no significant effect on any psychological need or acceptance outcome, reversing the ordering familiar from general-population TAM studies. The direct path from prediction accuracy to perceived usefulness was not significant, confirming that accuracy operates through the psychological-need pathway rather than as a direct utility. The authors interpret this as evidence that in transportation contexts where alternatives are severely limited, reliable information is what enables wheelchair users to plan independently rather than depend on personal experience or external help — 'when users cannot trust system-provided information, meaningful choice collapses.' Demographic data reinforced the stakes: 95.1% of participants reported facing transportation barriers 'often' or 'very often/always', and 78% used electric wheelchairs, making independent trip planning especially dependent on accurate dispatch estimates.
Relevance
The paper's core argument — that for disabled users, prediction accuracy is not merely a usability concern but a precondition for autonomy — generalises well beyond paratransit. Any accessibility-oriented system that delivers time, cost, or availability estimates (accessible-route mapping, captioning-availability indicators, assistance-request queues at airports, elevator status dashboards) operates under the same constraint: when the underlying information is unreliable, the user loses the ability to make meaningful independent choices and falls back on personal workarounds or human assistance. This reframes information reliability as a disability-justice issue, not just a product quality issue, and suggests that accessibility standards and procurement criteria for government-operated services should include explicit accuracy requirements alongside interface-level compliance. The inversion of the typical TAM pattern (where ease of use usually drives acceptance) is a useful caution against importing general-population usability findings uncritically into disability research. Limitations include the Seoul-specific training data, the relatively small sample (N=41), reliance on self-report without a real dispatch-time comparison condition, and the cross-sectional design that captures only short-term responses to prototype use.
Tags: transportation accessibility · wheelchair users · paratransit · self-determination · autonomy · technology acceptance · machine learning · mobility · physical disability · psychological needs