Have you ever been stuck at a red light that seemed longer than it needed to be? Or, had to stop for a red at every intersection on a street? Or, given up waiting for a WALK signal and crossed when you had the chance? Most people don’t think too much about traffic signals, but they have a big effect on how it feels to drive and walk through a city. In Boston, a city famous for our traffic and convoluted colonial street grid, getting traffic signals right is essential to keeping the city moving. 

In 2024, the Boston Transportation Department turned to Google’s Project Green Light to optimize signal timing. Boston regularly retimes signals using traditional modeling tools, but we believed that many intersections had opportunities for further improvement. Project Green Light works differently from traditional signal modeling. Rather than starting with a phasing plan and measured traffic volumes, Google’s AI systems look at anonymized movement data from phones to deduce the current signal sequence for an intersection. The system then seeks opportunities for adjustments that would reduce delay and idle time. Most of the recommended changes are minor – for example, adjusting the offset between two adjacent signals to improve coordination or shifting a few seconds from one phase to another. These recommended changes are presented to engineers, and once implemented, the system gathers data and reports back on the effects of the change within a few weeks. 

Since 2024, Boston has implemented recommendations on more than a quarter of our 900 signalized intersections. We saw an average 13 percent reduction of delay at modified intersections, though improvements were not universal — changes were reversed at about a third of locations due to lack of improvement or negative impacts. We estimated that drivers have saved tens of thousands of hours per year because of these changes, including several high-volume intersections in the Back Bay neighborhood, where we’ve seen delay reductions of 27 percent. 

Our experience with Project Green Light offered a few important lessons on how to successfully apply AI technology to city traffic signals: 1) ease-of-deployment, 2) human oversight, and 3) built-in validation and reporting.  

First, Project Green Light is designed for simplicity. It does not integrate directly with our signal control systems. Engineers receive notifications when new recommendations are available, with a link to an easy-to-use dashboard which clearly lays out the proposed optimization. For example, “Reduce phase 1 by 5 seconds and add 5 seconds to phase 3.” Integrating complex technology systems for automated action can be a monumental task, requiring careful engineering, validation, and ongoing monitoring. By avoiding this, Boston was able to get up and running with Project Green Light in weeks rather than months or years that a fully integrated system would require. 

Second, Project Green Light puts a “human in the loop” on any changes to traffic signals. Each intersection recommendation page shows the AI system’s understanding of the signal phasing. Because phasing is deduced from vehicular movements, the AI system occasionally gets it wrong, especially at complex, multi-way intersections. Human review ensures that an engineer with a full understanding of the intersection can consider the accuracy of the recommendation before implementation. Crucially, this approach also allows engineers to weigh tradeoff decisions that the AI system does not consider. For example, a recommendation to increase the cycle length at a busy intersection near a train station might improve traffic flow, but it would come at the expense of increased pedestrian delay where pedestrians are the primary user of the street.  

The third key to success with Project Green Light comes from automated impact measurement and reporting. Boston, like many cities, lacks robust, automated data gathering and analysis infrastructure to quickly measure traffic impacts resulting from signal optimizations. Within a few weeks of a change, Project Green Light provides an easy-to-understand analysis, with a clear recommendation on whether to keep the change or roll it back. By creating a short feedback loop between implementation and assessment, we were able to build confidence in the system and share impacts with the public.  

Beyond these success factors, there are limitations and challenges with the system as it exists today.  While Project Green Light found opportunities to better coordinate adjacent signals, it does not do full corridor or network level analysis. In complex areas and corridors, seemingly beneficial localized optimizations could fail to produce an overall network benefit and might even reduce throughput by shifting vehicle queuing to intersections with less available space, resulting in more blocking and conflict.  

Project Green Light is also only as good as the underlying signal system. At some of our intersections, the optimization opportunities identified arose from equipment failure – such as broken vehicle detectors or incorrectly configured signal plans – rather than a fundamental new insight about timing. While this helped us identify hardware issues, it also served as a reminder that AI requires a strong foundation of reliable infrastructure to deliver its best results. 

Making streets safe and welcoming to pedestrians further constrains AI optimizations. Under Boston’s standards, a 45’ crosswalk receives at least 20 seconds of crossing time, regardless of the volume of concurrent vehicle traffic. Many streets also have signals for bicyclists, particularly in high conflict areas. Providing for pedestrians and bicyclists to safely cross without excessive waiting limits the ability to treat signal optimization as a simple exercise in maximizing vehicle flow.  

These tradeoffs between pedestrian experience and vehicles are not simply optimization problems that a smarter AI could “solve.” Rather they reflect the underlying human values, tradeoffs, and policy choices that underpin many seemingly technical engineering decisions in urban transportation. AI holds tremendous promise to help cities get the most out of the streets they have, but transportation leaders must not let the allure of smarter technology lead to the outsourcing or abdicating of values-based decisions that are so critical to creating livable cities that serve the needs of all road users. 

Jascha Franklin-Hodge is the former Chief of Streets for the City of Boston.