Traffic-signal AI must optimize for the whole intersection

Boston's Green Light programme supplies an implementation example, while pedestrian, transit and neighboring-road effects remain essential checks.

Changing a signal by a few seconds can help one stream of vehicles and create a longer wait elsewhere. An AI timing recommendation therefore needs to be judged as a change to the whole intersection and surrounding corridor. Reduced stops are useful, but they are not the only public-interest outcome a city is responsible for.

Boston's June 2025 report describes an expanded partnership with Google's Project Green Light and implementation of recommendations at 114 intersections. The city reports reductions in delay and unnecessary stops at retimed locations. Google's mechanism uses traffic patterns to recommend adjustments for review by city engineers. The programme is an implementation example with reported outcomes, not evidence that every recommendation benefits every road user. [City report](https://content.boston.gov/news/mayor-wu-announces-expansion-project-green-light-signal-optimization-program), [project description](https://sites.research.google/gr/greenlight/).

The human decision remains explicit. A recommendation identifies a possible timing change; the traffic authority reviews it and implements an accepted plan using its own systems. That is different from giving a model unrestricted real-time control. The distinction allows a city to test whether the analysis helps engineers without obscuring responsibility for the signal plan.

An illustrative corridor pilot should begin with a clear objective and constraints. Record the problem intersections, neighboring signals, pedestrian movements and transit operations. State which outcomes the project seeks to improve and which must not deteriorate. If the objective mentions only vehicle delay, the model can recommend a technically successful change that misses the city's broader purpose.

Before-and-after evaluation should use comparable periods. Weather, roadworks, school calendars and demand can change travel patterns independently of a signal adjustment. Preserve the implementation date and compare relevant days and times. Where possible, use an appropriate comparison corridor or other study design to help distinguish the timing effect from wider changes.

Measure several stages of the journey. An intersection can show shorter local queues while moving delay downstream. Check corridor travel, spillback and route effects alongside the selected junction. A driver avoiding one stop is not a complete mobility evaluation if buses or crossing pedestrians experience a worse service.

Pedestrian and accessibility needs deserve direct observation. Review crossing opportunities, waiting and the relevant approved timings with the responsible engineers. Do not assume that smoother vehicle flow automatically improves every user's experience. A city can incorporate these constraints before selecting recommendations and examine them afterward.

Transit is another distinct requirement. Reliable bus movement can serve many people in one vehicle, while a general traffic metric may count it similarly to a car. The evaluation should reflect the city's actual priorities. Keep any transit-priority logic and emergency requirements visible in the plan rather than allowing a generic optimization goal to override them silently.

The city budget material helps place Green Light within ordinary traffic-management work. That context matters because AI recommendations supplement existing engineering, equipment and maintenance. An observed programme benefit may depend on those foundations being in good order; software analysis is not a replacement for functioning signals and communications. [Budget presentation](https://www.boston.gov/sites/default/files/file/2025/05/Streets%20FY26%20Budget%20Presentation_0.pdf).

Costs include engineer review, implementation, monitoring and any required integration. A supplier's statement that no new hardware is needed does not mean that the city incurs no effort or expense. The sources here do not establish a universal implementation price. Request terms and assess the workload using the city's actual signal infrastructure.

A recommendation record should explain what changed and why it was accepted. Retain the previous plan and a defined way to restore it if monitoring reveals a problem. When traffic data are unavailable or stale, the system should show that status clearly. The absence of a recommendation must not be presented as evidence that an intersection has no problems.

Review recommendations after a demand change such as a new school opening or a diversion. A previously useful timing plan can become less suitable as patterns shift. Keeping the implementation date and subsequent context together helps the city decide when another engineering assessment is justified.

Traffic AI can make patterns easier to find and help engineers focus their effort. Its strongest use is a reviewed, measured improvement to a city's existing process. The acceptance decision should follow outcomes for the full intersection and corridor, including the people least visible in a vehicle-focused dataset.