New York
CNN
—
Rob Sarno has been with the New York City’s Metropolitan Transit Authority (MTA) for 14 years. As assistant chief track officer, he assists maintenance and emergency response — which also meant teaching artificial intelligence systems what a damaged rail sounds like last year.
For a few months starting in September, he helped a pilot program between the MTA and Google Public Sector, the search giant’s division that works with government agencies and educational institutions. The project involves retrofitting Google’s Pixel smartphones to certain subway cars to collect sounds and other data and feed it into Google’s Cloud. The data is then analyzed to spot patterns that could indicate track defects before they become a problem.
“By being able to detect early defects in the rails, it saves not just money but also time – for both crew members and riders,” Demetrius Crichlow, New York City Transit president, said in a statement released February 27.
New York is just one major city to have implemented AI in the hopes of improving transit systems. In 2023, infrastructure consulting firm Aecom completed a pilot program for the New Jersey Transit system that used the technology to analyze customer flow and crowd management, and in 2024, the Chicago Transit Authority (CTA) uses AI to enhance security by detecting guns. Also in 2024, Beijing introduced a facial recognition system to be used in place of transit tickets and cards to reduce lines during rush hours.
The pilot program between the MTA and Google — dubbed TrackInspect — is just the latest indication that companies are exploring whether the technology can make transit more efficient, although whether such an initiative will ever be deployed widely remains to be seen.
TrackInspect which was announced last week, started as a proof-of-concept developed by Google Public Sector in partnership with its Rapid Innovation Team for the MTA at no cost, according to the transit agency. However, it’s uncertain whether the project will expand into a permanent program since it’s unclear how much it would cost the MTA, which already needs billions of dollars to complete existing projects.
Google has partnered with other transportation agencies in the past. The tech giant has developed a chat box for the Chicago’s CTA, launched direct data integration for Amtrak departure and arrival times and has partnered with tech providers Passport and ParkMobile to connect street parking meters to Google Maps.
But the MTA’s reach is massive; it’s the country’s largest public transit system with 472 subway stations and 237 local bus routes, according to MTA data. In 2024, the agency says there were more than 1 billion trips on the subway.
Yet service disruptions continue to be a problem for the aging 120-year-old transit system. There was a total of 38,858 total delays in September; 39,492 in October; 36,971; in November and 42,862 delays in December last year, according to data from the MTA. Chicago’s transit system, by comparison, only experienced about 200 delays each month with a wait time of 10 minutes or more in September, November and December. But only two of Chicago’s subways run 24 hours a day, unlike in New York, where most of the train lines runs at all hours every day or have other train lines fill in the gaps.
The goal of a program like TrackInspect is to figure out how to cut down on service disruptions.
Between September 2024 and January 2025, six Google Pixel smartphones with “standard, off-the-shelf plastic cases” were installed on four R46 subway cars — better known as the cars with the orange and yellow seats. The smartphones collected 335 million sensor readings, 1 million GPS locations and 1,200 hours of audio, according to the MTA.
The smartphones, which were located inside and underneath the subway cars, detected subtle sounds and vibrations using sensors and microphones. The smartphones located inside cars had their native microphones disabled and did not capture audio or customers’ conversations, only vibrations, whereas the smartphones outside the cars had additional attached microphones, according to the MTA. New York City Transit inspectors would examine areas highlighted by the AI system, manually check for issues and then feed those findings into the model to train it, the MTA said.
The system highlighted “areas that were above a certain threshold for decibels,” which could indicate a defect, according to Sarno. His role involved listening to clips ranging from five to 30 minutes and marking snippets that could signal an issue.
“Maybe a loose ball, maybe a loose joint, maybe a battered rail,” he said.
When asked why the devices were retrofitted on older models instead of newer ones, Sarno said the MTA typically uses older car models when making modifications in case there are any unwanted effects.
The MTA chose the A line because its cars go above and below ground. It also has areas with new construction, which provided a baseline for the MTA, according to Sarno. And there is no shortage of disruptions on the A line: Data from New York Open Data — an online portal where city agencies provide raw data to promote transparency — shows there were 2,252 delays in September, 2,368 in October, 2,643 in November and 2,572 in December. But not all delays were caused by mechanical or track problems; factors like crew availability, people on the track and construction played a much bigger role in setbacks.
After NYCT track inspectors examined the tracks in person, they compared their findings with Sarno and the system’s discoveries.
“That’s how we were teaching the model,” Sarno said. If his estimate based on the audio captured by Google’s phones matched the inspector’s findings, that was considered a positive prediction, and the AI model would be taught accordingly. Sarno said that his own positive prediction success rate was about 80 percent.
In addition to capturing and analyzing data for potential issues, the TrackInspect program included an AI system based on Google’s Gemini model that inspectors could use “to ask questions about maintenance history, protocols, and repair standards, with clear, conversational answers,” according to the MTA.
The TrackInspect system identified 92 percent of the defect locations found by the MTA’s inspectors and is considered to have been a success, a Google Public Sector spokesperson told CNN, adding that other transit systems have already expressed interest in similar programs.
New York Open Data showed that certain types of delays, such as those related to braking issues, rail and roadbed problems and service delivery, decreased on the A line from September to December. But it’s too soon to tell whether the pilot contributed to that change without further analysis, the MTA said.
The trial with Google may be over, but the MTA isn’t finished yet. It’s now trying to court other companies with technology that could help develop track improvement software.
