Will AI replace radiologists?

Some tasks are already being done by AI or machines; the rest still needs a person. About 4h 26m of a typical shift is made up of tasks that AI or robots have done on their own in real use somewhere. That doesn't mean those hours are automated in a typical workplace. The other 3h 34m (45%) stays with people. Separately, BLS expects about 3.4% more of these jobs by 2035 (all jobs: +3.5%).

AI shows up regularly in 3 of 30 tasks, about 4h 26m of a typical shift. The biggest ones:

• Perform or interpret the outcomes of diagnostic imaging procedures including magnetic resonance imaging (MRI), computer tomography (CT), positron emission tomography (PET), nuclear cardiology treadmill studies, mammography, or ultrasound. (2h 32m a shift, AI mostly does it alone)
• Prepare comprehensive interpretive reports of findings. (1h 45m a shift, AI mostly does it alone)
• Document the performance, interpretation, or outcomes of all procedures performed. (10m a shift, AI mostly does it alone)

Not in workplaces like this one, based on a 2026 review of robot deployments. 1 more task has only been done in factory setups built around the machine, so it stays with people here. For 3 physical tasks, no robot has shown it can do the work yet.

3h 34m of a typical shift, or 45%, goes to tasks with no AI or robot evidence yet. No AI usage or robot deployment has shown up for these so far. The biggest:

• Communicate examination results or diagnostic information to referring physicians, patients, or families. (29m a shift)
• Obtain patients' histories from electronic records, patient interviews, dictated reports, or by communicating with referring clinicians. (24m a shift)
• Confer with medical professionals regarding image-based diagnoses. (24m a shift)
• Perform interventional procedures such as image-guided biopsy, percutaneous transluminal angioplasty, transhepatic biliary drainage, or nephrostomy catheter placement. (24m a shift)
• Review or transmit images and information using picture archiving or communications systems. (14m a shift)

Much of it comes down to where and how the work happens. In-person, hands-on and high-stakes work is where AI and robots have made the least headway so far. In the official O*NET survey of people in this job:

• 92% say dealing with the public is very or extremely important
• 70% say a mistake would be very or extremely serious
• 67% work at arm's length or closer to other people

We track 1 sourced company move that affects this job. Most recent:

• Feb 3, 2025: Mercy rolls out Aidoc AI to review most diagnostic imaging (Rolling out)

For the 2 tasks in this job where AI shows up and timing data exists, a typical request takes about 4h 17m for a person alone and about 37m with AI. These are Claude's own estimates from real conversations in April and May 2026 (Anthropic Economic Index). They describe single requests, not whole workdays, and time saved on a task doesn't mean jobs are cut. The biggest gaps:

• Prepare comprehensive interpretive reports of findings (about 6h 14m alone, 43m with AI)
• Perform or interpret the outcomes of diagnostic imaging procedures including magnetic resonance imaging (MRI), computer tomography (CT), positron emission tomography (PET), nuclear cardiology treadmill studies, mammography, or ultrasound (about 2h 20m alone, 32m with AI)

On Indeed, job postings in the Physicians & Surgeons sector are at 202% of their February 2020 level (all U.S. postings: 104%), up 8.8% from a year earlier, as of Oct 9, 2026. BLS separately expects about 700 openings a year in this job through 2035. Across all U.S. postings, 7.3% now mention AI skills. Sector postings cover more than this one job and move for many reasons besides AI.

The median pay is about $420,860 a year ($202.34 an hour), 726% above the median for all U.S. jobs ($50,980). Most people in this job earn between $89,010 and $594,410 or more a year; that range covers the middle 80%. Pay figures are from BLS, May 2025, and vary a lot by state and employer. Pick a state to compare:

BLS projects growth of 3.4% from 2025 to 2035, compared with 3.5% growth for all jobs. That's about 30,000 jobs in 2035. It expects about 700 openings a year, mostly from people retiring or changing careers. These projections don't single out AI as a cause.

About 27,000 people in the U.S., roughly 0.02% of all workers (BLS, May 2025). That makes it the 497th largest of the 796 jobs on this site. BLS expects about 30,000 in 2035.

The most common requirement is post-doctoral training: about 58% of people in the job said so in the official O*NET survey. O*NET rates the job at level 5 of 5 for preparation, which usually means: graduate degree. Requirements vary by state and employer, especially for licensed work.

These come from the job's official skill ratings and from the tasks that AI and robots touch least today. Building them keeps you strong in the parts of the work that still need a person:

• Skills that matter more here than in most jobs: Science, Judgment and Decision Making, Monitoring, Social Perceptiveness
• Get strong at the parts of the job that stay with people (listed above), which take 3h 34m of a typical shift
• Tools employers list: Epic Systems, eClinicalWorks EHR software, MEDITECH software, Microsoft Excel, Microsoft Word
• Learn to use AI well on the tasks where it already shows up in this job (listed above)

These use similar skills, with less of a typical shift touched by AI or robots. Matches come from 109 official ratings of skills, knowledge and work activities, and each needs the same or less preparation unless the official list relates the jobs directly:

• Anesthesiologists (57m touched, $391k median pay)
• Ophthalmologists (44m touched, $300k median pay)
• Urologists (1h 17m touched, $266k median pay)
• Sports Medicine Physicians (0m touched, $266k median pay)
• Pathologists (2h 16m touched, $312k median pay)

Each task carries its own evidence label. For this job: 3 observed in real AI use, 1 with robots in real deployments, 26 with no AI or robot evidence found. The minutes per task are estimates, not measured hours, and "no evidence" doesn't prove a task is safe.