AI reads the mood of metro passengers in São Paulo and Porto, revealing how crises and tourism shape public value

Every day, roughly 405,000 people pass through Sé station in São Paulo, and about 69,000 validations are recorded at Trindade in Porto. Behind those numbers sit thousands of spontaneous reviews on Google Maps, written in frustration or relief by passengers who have just experienced the metro as a public service. A new study has now turned those reviews into a longitudinal dataset, using artificial intelligence to track how passengers in two very different cities perceive the same essential infrastructure over eight years.

The research, published in Frontiers in Future Transportation, was conducted by Hadassa Souza and Rodrigo Ribeiro de Oliveira at the Federal Institute of Science and Technology of São Paulo and Álvaro Costa at the University of Porto. It collected 2,196 reviews of Sé station and 1,146 reviews of Trindade station between January 2017 and December 2025 through an automated extraction system, then processed them with DeepSeek-V3, a large language model, to classify sentiment and identify recurring themes. The study is framed around two concepts: Public Value, the collective legitimacy of a public service, and Travel Experience, the subjective quality of an individual journey.

The two systems could hardly be more different. The São Paulo Metro opened in 1974 and now operates seven lines across 104.4 kilometres, serving a metropolitan population of 11.45 million. The Metro do Porto began commercial operations in December 2002, runs six light rail lines over 70 kilometres, and recorded 89.8 million passengers in 2024. Sé station is a heavy metro interchange between Lines 1 and 3, accounting for 22 percent of the system’s ridership. Trindade is the main hub of the Porto network, handling 18 percent of all validations and connecting every line in the system.

To validate the AI classifier, the researchers drew stratified random samples of 331 comments from Sé and 292 from Trindade, distributed proportionally across the years. Human coders classified the same comments independently, and agreement was measured with Cohen’s Kappa coefficient. The overall average Kappa was 0.85 for Sé and 0.89 for Trindade, both classified as almost perfect on the Landis and Koch scale. The final classifier achieved an average F1-score of 0.94 for Sé and 0.95 for Trindade, with the highest performance on safety-related content.

The prompt given to DeepSeek-V3 asked the model to classify each comment on nine criteria: efficiency, accessibility, customer service, amenities, comfort, cleanliness, infrastructure, fare satisfaction, and safety. It also asked the model to identify whether the comment demonstrated Public Value across three sub-dimensions — accessibility, safety, and cost-benefit — and Travel Experience across physical and sensory perception, service quality, and emotional and affective response. The researchers deliberately chose a large language model over traditional dictionary-based approaches such as VADER because reviews often contain contextual ambiguity, regional expressions, and idiomatic language that rule-based systems struggle to interpret.

The results reveal a striking dissociation between the two dimensions. At Sé station, 42 percent of comments combined positive Public Value with a negative Travel Experience, more than double the 18 percent observed at Trindade. In other words, passengers in São Paulo frequently acknowledged the station’s collective utility — its role as an essential hub connecting the city — while simultaneously describing their individual journey as unpleasant, unsafe, or dirty. At Trindade, by contrast, 52 percent of comments recorded positive ratings in both dimensions simultaneously, suggesting that passengers there perceive the station’s public value and their own experience as tightly integrated.

The temporal patterns are equally revealing. At Sé, the average star rating dropped from 3.7 to 3.9 between 2017 and 2019, coinciding with a post-2015 economic crisis and higher rates of negative sentiment, with comments frequently referencing perceived insecurity, dirtiness, and overcrowding. During the pandemic years of 2020 to 2022, ratings rose to 4.1, with 80 percent positive comments and frequent mentions of the station being clean and organized, likely reflecting reduced passenger flow and intensified health measures. In 2023, the rating fell again to 3.9 amid the Cracolândia crisis in downtown São Paulo, with 34 percent of negative comments highlighting homelessness, drug use, and robberies. By 2024, investments in security and maintenance pushed the average rating to its highest point of the entire period at 4.3 stars.

Trindade tells a different story. The median rating remained at 5 stars throughout the entire period, and the average never fell below 4.10. In 2021, the station achieved a peak of 4.67 stars with 92.2 percent positive reviews and no negative comments at all, likely reflecting post-pandemic appreciation of restored services. However, from 2022 onwards, criticism increased as tourism surged. Comments began highlighting terms such as crowded, noise, and tourists, indicating that the rise in visitor numbers was straining infrastructure and degrading passenger comfort. The researchers also identified persistent frustration with the ticket vending machines, described as confusing, and with counter service perceived as insufficient or unempathetic — issues that fall within the operator’s direct remit.

The cross-tabulation of 1- and 2-star ratings shows which sub-dimensions drive the most severe dissatisfaction. At Sé, the physical and sensory dimension — encompassing cleanliness, comfort, crowding, and safety — was present in 78 percent of the lowest ratings, confirming that exposure to adverse conditions is the primary driver of collapse in the travel experience. At Trindade, service quality — machines, customer service, signage — appeared in 88 percent of 1- and 2-star ratings, making it the main bottleneck for satisfaction. The researchers note that collapsing both patterns into a single explanation would misattribute responsibility: at Sé, negative sentiment is largely driven by urban management and public safety conditions surrounding the station, while at Trindade it concentrates on rail-operation factors within the operator’s control.

Modal integration emerged as the most resilient sub-dimension of Public Value in both systems, maintaining positive perception above 85 percent even during crisis years — 82 percent at Sé during 2018 and 2019, and 88 percent at Trindade during the years of highest tourist pressure. This finding has direct policy implications: investments in line expansion, improved intermodal signage, and fare integration yield greater returns in perceived legitimacy than other dimensions. The researchers propose that the same UGC methodology could serve as a low-cost monitoring tool, tracking quarterly safety mentions, night-time complaint shares, and ticketing complaint frequencies to guide seasonal staffing and multilingual interface upgrades. The study aligns with UN Sustainable Development Goals 9.1 and 11.2, which call for quality infrastructure and access to safe transport for all.

Subject of Research: Operationalizing public value and the travel experience with artificial intelligence: a longitudinal analysis of user sentiment at subway stations in São Paulo and Porto

Article Title: Operationalizing public value and the travel experience with artificial intelligence: a longitudinal analysis of user sentiment at subway stations in São Paulo and Porto

Article References: Souza, H., Costa, Á., & de Oliveira, R. R. (2026). Operationalizing public value and the travel experience with artificial intelligence: a longitudinal analysis of user sentiment at subway stations in São Paulo and Porto. Frontiers in Future Transportation, 7, Article 1924380. https://doi.org/10.3389/ffutr.2026.1924380

Image Credits: AI Generated

DOI: 10.3389/ffutr.2026.1924380

Keywords: artificial intelligence, sentiment analysis, urban mobility, metro, public value, travel experience, DeepSeek-V3, São Paulo, Porto, Operationalizing, public, value