Yokohama: AI Cuts Taxi Cruising Time 5.3% and Skill Gap 13.4%

The Challenge

Taxi drivers in Yokohama were spending much of their shift simply looking for passengers. Researchers wanted to find out whether a predictive AI system could help cut down empty cruising time. More importantly, they asked: does this technology level the playing field between experienced and novice drivers, or does it amplify existing skill gaps?

The trial faced several hurdles. Uptake was modest—fewer than 200 out of 520 drivers tried the tool, and it was active in only 12.6% of journeys. When cabbies did switch it on, conditions were typically harder: median search time with AI running stood at 12.20 minutes, compared to 7.75 minutes when it was off. That selective usage made it difficult to isolate the tool’s true effect. The research team also lacked basic driver data such as age or tenure, and had to account for minute-by-minute switching during individual shifts.

The Approach

To tease out causality, the researchers used sophisticated statistical controls. They compared the same driver’s performance with and without AI, holding constant location and time of day. A skill index was built from each driver’s historical record, and a demand index tracked local conditions hour by hour.

Because drivers tended to activate AI in tricky situations, a standard comparison would be misleading. The team tackled this by using an instrumental-variable method: drivers unfamiliar with a drop-off area were more likely to turn on the tool, yet that unfamiliarity didn’t directly affect cruising success. Additional checks—including propensity-score trimming and placebo tests—confirmed the findings were robust. The analysis also looked at whether drivers followed AI recommendations and tracked changes across the trial’s first and second fortnights.

What the Data Revealed

Cruising time fell modestly

Switching on the AI cut empty search time by 5.3% on average. That figure stayed consistent across different analytical methods, ranging from 4.8% to 6.5%.

Lower-skilled drivers gained most

The benefits were far from equal. Less experienced drivers saw their search time drop by roughly 7%, with the least skilled third enjoying an 8% improvement. Meanwhile, top performers saw virtually no change. In effect, the AI acted as a substitute for hard-won local knowledge.

Productivity spread narrowed

The performance gap between the best and worst drivers shrank by 13.4%. This compression suggests the tool is redistributing efficiency gains downwards rather than magnifying existing advantages.

Fare quality held steady

There was no sign that drivers were chasing quicker, cheaper rides when using AI. Average fares remained stable, ruling out one potential unintended consequence.

Benefits plateaued early

All the improvement came within the first two weeks; no further learning occurred thereafter. Compliance with AI suggestions hovered around 55% for less skilled drivers and 53.5% for their more experienced peers—remarkably similar rates.

Broader Implications

This study offers rare real-world evidence that AI can directly replace human expertise in tasks that rely heavily on prediction. The gains flowed almost entirely to those with less skill, narrowing inequality within the workforce.

That compression has workforce implications. Firms might recruit more novices for AI-augmented roles, knowing the technology compensates for lack of experience. At the same time, it underscores the importance of developing skills that machines can’t easily replicate—particularly interpersonal abilities and customer service. As predictive tools become ubiquitous, the premium on social and adaptive skills is likely to rise.