AI traffic lights are already deciding, intersection by intersection, how long you sit at red before your city’s engineers ever look at the data themselves. Pittsburgh’s Surtrac system, the first large-scale deployment of this technology in the United States, cut travel times through its network by 25 percent, wait times by more than 40 percent, and emissions by 21 percent in its initial East End rollout, according to a case study compiled for the National Science Foundation’s Computing Community Consortium.

A decade later, similar systems are running in cities on four continents, and the pitch to every one of them is the same: better signal timing, using infrastructure the city already owns, can meaningfully cut congestion and pollution without a single new road being built.

What has changed since Pittsburgh’s early deployment is scale and ownership. Google’s Project Green Light now operates across dozens of cities from Seattle to Bangalore, using aggregated Google Maps driving data rather than city-owned sensors, and a growing number of critics are asking a harder question than whether AI traffic lights work.

They are asking who decides what these systems optimise for, whose neighbourhoods get the upgrade first, and what happens to the driving and movement data these systems inevitably collect along the way.

How AI Traffic Lights Actually Work

Two distinct approaches dominate the field. Pittsburgh’s Surtrac, developed at Carnegie Mellon University, uses real-time sensor data at each intersection to allocate green time dynamically as traffic actually arrives, rather than following a pre-set schedule built around average historical conditions.

Google’s Project Green Light takes a different route entirely: instead of installing new hardware, it uses over a decade of Google Maps driving trends to build a model of how traffic actually flows through an existing intersection, then hands city engineers a dashboard of specific timing recommendations they can implement, according to Google, in as little as five minutes using infrastructure the city already has.

Google says Green Light’s early results show up to a 30 percent reduction in stops and a 10 percent cut in greenhouse gas emissions at optimised intersections, with the tool live in more than 70 intersections across cities on four continents and estimated to affect up to 30 million car trips a month.

The company is explicit that user location data is never shared with city partners directly; Green Light provides only aggregated timing recommendations, not access to the underlying driving data itself. City engineers see a dashboard of suggested timing changes and an impact report after implementation, not raw movement data about individual drivers.

The Case for Adaptive Signals: Real Numbers, Genuine Gains

The evidence that adaptive signal control reduces delay and emissions is not seriously disputed in the traffic engineering literature. Beyond Pittsburgh and Google’s own figures, a federally tracked adaptive signal pilot in Maricopa County, Arizona cut average vehicle delay by 46 percent and pedestrian waiting time by 22 percent at the project intersection, according to a US Department of Transportation Intelligent Transportation Systems Joint Program Office update.

Los Angeles’s citywide adaptive signal network is credited with saving roughly 9.5 million driver hours annually. Congestion at US intersections costs an estimated 85 billion dollars a year in wasted time and fuel, and pollution at a typical city intersection can run 29 times higher than on an open road, largely from vehicles accelerating after a full stop, which is precisely the pattern adaptive signals are designed to reduce.

The deployment cost varies enormously depending on approach. Software-only systems like Green Light require no new hardware and can be rolled out to an existing signal in days. Sensor-based systems like Surtrac cost roughly 20,000 dollars per intersection to equip, and full citywide hardware overhauls, such as Nashville’s, have run into the hundreds of millions of dollars.

That cost gap is itself a policy-relevant fact, since it shapes which cities, and which parts of a given city, get access to the more sophisticated version of the technology first. A wealthy commercial corridor and an underfunded residential district are rarely competing on equal footing for the same upgrade budget.

What This Means for You: The Trade-Offs Rarely Mentioned in the Press Release

The most consistent criticism of AI traffic lights is not that they fail to reduce vehicle delay. It is that vehicle delay is the metric they are built to optimise, and pedestrian safety, noise, and neighbourhood character are values that do not show up in that objective function unless a city deliberately writes them in.

Research and advocacy groups tracking Vision Zero pedestrian safety programmes warn that AI-driven traffic systems can unintentionally prioritise vehicle throughput over pedestrian safety if congestion reduction becomes the metric a city’s engineers are judged against, and that advanced traffic technology tends to appear first in wealthier commercial corridors even though lower-income neighbourhoods often carry higher crash exposure.

A widely cited essay on AI and smart cities put the underlying mismatch plainly: technologists train a model to optimise commute times, while residents often prioritise pedestrian safety, noise reduction, or neighbourhood character, values that are harder to quantify and therefore routinely excluded from what the algorithm is actually optimising for.

The essay points to the 2020 collapse of Alphabet subsidiary Sidewalk Labs’ Toronto waterfront project as the clearest case of a community successfully rejecting an AI-driven smart city proposal before it was built, specifically over unresolved questions about who would own the movement data the sensors would generate and who could audit the algorithms making decisions about the neighbourhood.

The Camera and Data Question

Camera-based AI traffic systems raise a second, separate concern that has nothing to do with signal timing itself: many of the sensors that make adaptive control possible are also capable of classifying vehicle types, estimating speeds, and in some deployments, reading licence plates, generating a movement dataset that has uses well beyond traffic engineering.

That capability sits inside a much larger expansion of AI-enabled monitoring in American cities, one that LiveAIWire’s coverage of the AI surveillance state found is currently expanding faster than the legal frameworks meant to govern it, with state legislatures passing 145 AI-related laws in a single year specifically because deployment has consistently outpaced deliberate oversight.

Cities piloting traffic AI have generally responded to this concern with technical fixes such as anonymised sensor data and thermal rather than optical cameras, but the underlying governance question, who can access the data a traffic camera collects and for what secondary purposes, is rarely settled before a system goes live rather than after.

That pattern echoes LiveAIWire’s broader reporting on AI in real estate and urban planning, which found that smart city sensor networks installed for legitimate services like air quality monitoring routinely create surveillance capabilities that operate without the transparency or legal authority a system of equivalent intrusiveness would require in most other public sector contexts.

Reading the Emissions Numbers Correctly

The emissions figures vendors publish for AI traffic lights are generally credible as far as they go, but they answer a narrower question than the one most coverage implies. A 10 percent emissions cut at an optimised intersection is a real, independently measurable local effect.

It is a different claim entirely from saying AI traffic technology is shrinking transportation’s overall carbon footprint, since the same efficiency gains that make traffic flow smoother can also make driving marginally more convenient and therefore more frequent, the same rebound dynamic LiveAIWire has documented in its coverage of the green AI myth in the context of AI’s own energy consumption.

Reducing stop-and-go idling at a given intersection is a genuine local win worth having regardless, but it is not, by itself, evidence that AI traffic systems are reducing total vehicle miles travelled or total transportation emissions at the city level. The two claims get blurred constantly in coverage of this technology, and they are not the same claim.

Who Gets to Audit the System

Barcelona’s approach under former chief technology officer Francesca Bria is frequently cited as a workable middle path: the city used machine learning for traffic and air quality analysis but required that any data generated by public infrastructure remain under public control.

Barcelona also favoured open-source models that could be independently audited over proprietary black-box systems, and treated AI recommendations as inputs for accountable human planners rather than as self-executing decisions, a governance choice rather than a technical one.

That model depends on a level of in-house technical capacity and political will that most mid-sized cities adopting AI traffic tools do not currently have, which is precisely why LiveAIWire’s coverage of the accountability infrastructure being built to audit government AI systems treats this as an active, unresolved problem rather than a solved one.

Most cities adopting AI traffic control are, in practice, taking a vendor’s efficiency numbers on trust because they lack the independent capacity to verify them, which is a governance gap rather than a flaw in the underlying technology.

The Feature Vendors Rarely Lead With: Emergency Priority

One capability of AI traffic lights gets far less press coverage than emissions or delay statistics, and it may matter more to the people directly affected by it: automatic priority for ambulances, fire trucks, and transit buses. Adaptive systems that already track approaching vehicles in real time can extend a green light or shorten a red cycle the moment an emergency vehicle’s transponder is detected, shaving critical seconds off response times without a dispatcher or driver doing anything differently.

Cities that have layered this feature onto existing adaptive signal networks report meaningfully faster emergency response times on corridors where the technology is active, and transit agencies report improved on-time performance for buses using the same priority mechanism. Neither benefit shows up in the headline emissions or congestion figures vendors typically publish, which is itself a useful reminder that a traffic AI system’s most valuable feature for a given city may not be the one its marketing leads with.

That gap between what gets measured publicly and what a community actually values most is the same gap Barcelona’s model was built to close by keeping the choice of what to optimise for in the hands of accountable local officials rather than a vendor’s default settings.

What This Means for Your Commute

If your city has adopted or is piloting AI traffic light technology, the concrete question worth asking your local transportation department is not whether the system reduces average delay, since the evidence that adaptive signal control generally does is solid.

The more useful questions are which neighbourhoods received the technology first and why, whether pedestrian and cyclist safety metrics are tracked alongside vehicle throughput or subordinated to it, and who, other than the vendor, has access to audit the sensor data the system generates.

Those questions determine whether a given deployment looks more like Barcelona’s accountable model or more like the surveillance-first version civil liberties advocates warn against, and the technology itself does not answer that question. The governance choices a city makes around it do.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.