Cities start to test AI traffic lights as congestion and emissions targets tighten

From London to Los Angeles, a new wave of pilot projects is quietly putting artificial intelligence into one of the most familiar pieces of urban infrastructure: the traffic light. Instead of following rigid pre‑programmed schedules, these upgraded systems use live data to decide when to turn red or green.

Supporters say AI‑driven signals can cut journey times, reduce idling at junctions and help cities meet air quality and climate goals. Early tests show promise, but also highlight questions about transparency, data use and what happens when algorithms control something as fundamental as who gets to move.

How AI traffic lights actually work

Traditional traffic signals usually run on fixed cycles that are tweaked occasionally using historical traffic counts. Modern adaptive systems already react to sensors or cameras, but they still follow limited rule sets defined in advance by engineers.

AI‑based systems go further. They take continuous input from cameras, radar, induction loops and sometimes connected vehicles, then use machine learning models to predict flows over the next few seconds or minutes. Based on those predictions, they adjust phase length, sequence and coordination with nearby junctions.

Some deployments use reinforcement learning, where algorithms are trained in simulation on years of virtual traffic. Once deployed, they keep updating their strategies as patterns evolve, for example when a new shopping center opens or a street is closed for construction.

Global pilots and early results

A growing list of cities has started real‑world trials. In Pittsburgh, a long‑running project on dozens of intersections has reported shorter travel times and fewer stops at red lights on main corridors compared with static timing plans, according to summaries published by transport researchers.

In London, Transport for London has been testing AI‑controlled signals at selected junctions with the goal of smoothing bus movements and reducing congestion hotspots. Authorities in Sydney and several European mid‑sized cities have launched similar pilots focused on peak‑hour bottlenecks and corridors serving major hospitals.

Real‑world gains vary by location, but municipal reports often mention improvements such as more consistent travel times, better handling of unusual events and fewer long queues on side streets. Even modest percentage cuts in average delay can translate into millions of hours saved annually in larger metro areas.

Potential benefits for emissions and safety

Less stop‑start traffic can mean lower emissions, especially in cities where older petrol and diesel cars are still common. Vehicles waste fuel when accelerating from a standstill, so fewer full stops on arterial roads can have a noticeable impact on local pollution hotspots.

Some pilots feed data from air quality sensors into their optimization goals. The system can then favor patterns that keep traffic moving steadily through sensitive areas such as school zones, even if that means slightly longer waits elsewhere.

Safety is another target. AI systems can be tuned to prioritize longer all‑red phases for pedestrians at high‑risk crossings or to react faster when sensors detect vulnerable road users. Several trials integrate data from pedestrian buttons, bicycle detection loops and near‑miss analytics from cameras to adjust timings around schools and busy shopping streets.

How it changes the commuter experience

Drivers and cyclists may not notice that AI has taken over, beyond the sense that certain journeys feel smoother. The biggest visible change often comes from coordination: green waves along major corridors that adapt more fluidly to actual demand instead of fixed peaks.

Public transport operators are also paying attention. Some buses already communicate their location and delay status to traffic systems. AI‑based controllers can then decide whether to hold a green light for a bus running behind schedule or to prioritize a streetcar line carrying large numbers of passengers over lightly used side roads.

For pedestrians, the impact depends heavily on how cities set priorities. Well‑configured systems can cut the wait for a walk signal during off‑peak periods and respond more quickly when people press a button. Poorly tuned ones risk extending waits if vehicle throughput is treated as the dominant goal.

Data, privacy and bias concerns

Most advanced systems rely on video feeds, which immediately raises privacy questions. Cities and vendors say they increasingly use computer vision that analyzes objects on the fly without storing identifiable images, but policies vary and are not always clearly communicated.

There is also the issue of algorithmic bias. If an AI is trained mainly to optimize vehicle throughput, it may systematically disadvantage pedestrians, cyclists or neighborhoods with different traffic patterns. That might translate into longer waits in lower‑income areas or worse service for non‑car users.

Transport planners and civil society groups are calling for explicit policies that state which users are prioritized, as well as public reporting on outcomes in different districts. Some pilot programs now publish metrics such as average pedestrian wait times and bus delay changes alongside car‑centric indicators.

Who is in control when something goes wrong

AI traffic control also raises questions about accountability. When timings are set by human engineers, responsibility for bad outcomes is straightforward. When decisions emerge from a trained model, tracing why a particular pattern occurred can be much harder.

Most deployments keep human operators in the loop and allow signals to fall back to conventional timing plans if systems malfunction or if roads need to be managed manually, for example during a major incident or protest. Regulators are still working out how to adapt safety standards and certification processes originally designed for simpler hardware.

Experts in transport governance argue that clear audit trails, simulation testing under unusual scenarios and independent oversight will be needed if AI signal control expands to entire city networks rather than small test zones.

What comes next for AI on the roads

As connected vehicles become more common, traffic lights may eventually coordinate directly with cars, buses and bikes. Pilot schemes already use short‑range communications to tell approaching vehicles how long a light will stay green, which can help drivers adjust speed to avoid stopping.

Future systems could combine AI at junctions with city‑wide digital twins, where planners test changes virtually before making them in the real world. That might include varying signal priorities by time of day, weather conditions or special events, and measuring the knock‑on effects on congestion and emissions.

For now, most cities are moving cautiously, expanding pilots only when they show tangible benefits and fit within existing budgets. As AI traffic lights spread, the biggest question will not be whether they can make journeys faster, but how fairly and transparently they serve everyone who uses the streets.