Why the work stays with an officer on the ground
Transit and railroad police spend their shifts in stations, on moving trains and along track right-of-way. They patrol platforms and rail yards, break up disturbances, question people, make arrests and write up what happened. Ask whether AI will replace transit and railroad police and the honest answer starts with that mix: software can read a camera feed, but it cannot step between two people fighting in a train car at 11 p.m.
The legal side carries as much weight as the physical side. Authority to detain, search and use force sits with a sworn person who answers for the decision afterward in a report, a hearing or a courtroom. A model can flag a pattern in fare-gate data; it cannot be cross-examined about why it stopped someone. That accountability is a hard boundary, not a software problem waiting on a better model.
Scale matters too. This is a small, specialized force: the Bureau of Labor Statistics counted about 4,390 transit and railroad police in the United States with median pay of $90,230, and projects employment change of 3.2% over 2025 to 2035 (BLS, 2025). Small agencies rarely rip out their staffing model for new tooling. They bolt the tooling on.
What software handles, what it assists, and what it leaves alone
The first group is the desk work. Drafting an incident narrative from structured fields, pulling records, and searching hours of recorded video for a described person are all jobs a model can take a first pass at. Our share of task time in the AI-does group is 0%. Overall coverage of this job’s task time by today’s AI sits at 13 on a 0 to 100 scale, measured the way we explain on how coverage is scored.
The assisted group is larger and more interesting. Analytics can rank which platforms need a patrol pass, translate a statement, suggest charges from a fact pattern, or alert on a person on the tracks before a train arrives. The officer still decides what to do with the alert. The share of task time where AI helps rather than replaces is 17%.
The rest belongs to a person: foot patrol, de-escalating a mental-health crisis on a crowded car, physically restraining a suspect, controlling a platform during an evacuation, and testifying. The share of task time that still needs a human is 83%, which is why the headline figure lands where it does on the Still needs a human scale.
What has actually been tested
Not much, and we say so plainly. Our evidence grade for quality parity on this job is D, and a D grade means there is no direct, published test of an AI system against a qualified transit or railroad police officer on this job’s real tasks. So we publish no parity number for it. Guessing one would be worse than leaving the cell empty.
What would settle it? A controlled comparison on narrow slices of the work: report accuracy from bodycam audio against officer-written reports, suspect-identification accuracy from station video against trained reviewers, or dispatch prioritization against a watch commander’s calls. Each is measurable. None of it would touch arrest authority. Our rules for grading that kind of test are on the quality-parity method page, and the full approach is on our methodology page.
When the picture could shift
Most likely after 2042 (8 in 10 of our scenarios). The chart above shows the spread; how we build the replacement-year range explains what that window is measuring.
Two things could pull it earlier. Transit agencies keep adding camera analytics, gate sensors and track-intrusion detection, and each rollout moves monitoring work off a person’s plate. Cheap, reliable mobile hardware that can move through a packed car would matter even more, because it would take on the presence part of presence.
Two things push it later. The hardware is not close: the robotics panel above places this job in the hardest physical tier, and nothing on the market handles a stairwell, a crowd and an uncooperative person safely. And the legal structure resists. Sworn authority, use-of-force rules, union agreements and evidence standards all assume a named human. Changing them takes legislatures, not product releases.
What to do: treat AI tooling here as a reporting and monitoring assistant, and get good at auditing what it flags.
How to stay needed in transit policing
Lean into the parts of the job that no feed can do. First, crisis response on a platform or in a car, where outcome depends on tone and timing. Second, building a route and a beat where riders and station staff know you and talk to you. Third, case work that ends in testimony, where your notes and your credibility are the evidence.
Two skills are worth real effort. One is evidence handling in a video-heavy environment: knowing how analytics reach a match, what they miss, and how to document a review so it holds up. The other is crisis intervention and de-escalation training, which is both the hardest task to automate and the one agencies are most often short on.
If you want to see how close jobs compare, police and sheriff’s patrol officers, detectives and criminal investigators and transportation security screeners sit near this one in the data. The wider law enforcement workers family and the government sector page give the pattern across similar roles, and the jobs least exposed to AI list shows where hands-on, licensed work clusters. You can also put any two roles side by side on our compare tool.