Why running a shift stays with people
Supervising correctional officers is live work with people who do not want to be there. A sergeant or lieutenant assigns posts, watches how an officer handles a tense exchange, and walks toward trouble when a dayroom turns. Software can flag a camera view or a missed count. It cannot stand between two men, read the room, and talk the moment down.
The second reason is accountability. Use-of-force reviews, discipline, inmate grievances and incident reports all need a named person who signs and can be questioned about it later, in a hearing or in court. An agency can let a model draft the narrative. It still needs a human supervisor who was there, or who spoke to the officers who were.
Scheduling and training sit in between. Filling a roster when three officers call out is partly math, and software is good at math. Deciding which officer belongs on the segregation tier tonight is a judgment about temperament and recent history. That judgment is why the headline question — will AI replace first-line supervisors of correctional officers — comes back to the task mix rather than to any single tool.
What AI runs, what it assists, what it leaves
A narrow slice of the job can run with no one in the loop: generating shift logs from existing records, reconciling head counts against the roster, and sorting routine paperwork into the right file. Task time in that group: 0%.
A larger block is assisted work. Drafting an incident report from an officer’s notes, summarizing video for a review, surfacing patterns in grievances, and building a first-pass schedule all move faster with a model — but a supervisor edits and signs. Task time in that group: 26%. Across all tasks, the ‘Can AI do it?’ score for this job is 19 out of 100; the coverage method page explains how that share of task time is built.
The rest stays with people. Responding to disturbances, searching cells and inmates, inspecting the perimeter and the locks, counseling an officer whose performance has slipped, and testifying about what happened on a tier — none of that moves to software. Task time in that group: 74%. The robotics panel on this page puts the physical share of the work in the hardest tier, a dexterous humanoid, and nothing of that kind is working a housing unit.
What the evidence does and does not show
There is no direct test of an AI system against a correctional supervisor doing this job. The evidence grade for ‘Is it better than a person?’ is D, which is our mark for not measured, so no parity number is given here. We do not estimate one from adjacent work.
What would settle it: a published study that scores model output against supervisor decisions on real incident reviews, scheduling under call-outs, and grievance findings, with corrections professionals grading both blind. Until something like that exists, the honest answer is that the task mix carries the score, not a benchmark. Our full method is at methodology, and the parity grades are explained on the quality parity page.
Outside the AI question, the labor picture matters. The Bureau of Labor Statistics counts about 53,380 of these supervisors in the United States, with median pay of $77,970 (BLS, 2025), and projects employment slipping roughly 2.5% between 2025 and 2035. That decline is driven by jail and prison populations and state budgets, not by software.
When the picture could change
Most likely after 2042 (8 in 10 of our scenarios). The replacement-year method sets out what that window measures and how the scenarios are drawn.
Two things could pull it earlier. Camera analytics and sensor systems keep improving at spotting fights, falls and crowding, which shifts some of the watching toward software. And the cost gap shown above is wide: the tooling is cheap against a staffed supervisory post, so budget pressure in understaffed facilities gives agencies a reason to try it.
Two things hold it back. Custody decisions sit inside legal and union processes that expect a person to answer for them, and those rules change slowly. And the physical half of the shift — searches, escorts, counts, responding to a cell — needs hardware that does not exist outside research labs.
How to stay needed on the tier
Lean into the parts of the job that stayed in the human column. Incident response and de-escalation. Officer development: coaching, evaluations, and knowing who to put where. Inspections and physical security checks, where a supervisor’s eyes still catch what a sensor misses.
Two skills are worth real time. First, writing and reviewing: if models draft the report, your value is catching what the draft gets wrong before it reaches a hearing. Second, basic fluency with the facility’s analytics and records systems, so you can say plainly what a flagged alert does and does not mean.
What to do: ask who reviews the output of any new monitoring or report tool at your facility, and make sure a supervisor sits in that step.
Nearby work is worth a look if you are planning a move. Compare this job with correctional officers and jailers, first-line supervisors of police and detectives, and probation officers and correctional treatment specialists. You can put any two side by side on the compare tool.
For wider context, see the supervisors of protective service workers family, the government sector page, and our corrections overview. If you want to see where this sits against everything else we score, start with the full rankings.