68.7% of family medicine residents stay in-state after training. 33.2% of colon and rectal surgery residents do. Nobody has figured out how to make the second number behave like the first.
This week, health systems apparently decided they were done fighting over the same shrinking pool of physicians and started building their own. Sutter Health announced a new medical school with Santa Clara University, aiming to grow from 170 residency slots to 1,000 by 2030. HRSA sent $11.2 million in new rural residency grants to 15 programs across 14 states. Indiana University of Pennsylvania is opening the state’s first public osteopathic medical school. Huntsville Hospital is already recruiting physicians who won’t start until 2029.
The retention data behind these bets is boringly consistent. More than 55% of physicians who finished residency between 2015 and 2024 practice in the same state where they trained. Family medicine and psychiatry — the specialties rural communities most desperately need — keep close to 70% of their graduates in-state. Train locally, retain locally. Health systems are treating GME spending as a core recruiting strategy instead of an academic afterthought, because the alternative is competing for the same finite talent pool with everyone else, forever.
At the other end of the pipeline, something less optimistic is taking shape.
The physicians these systems are working so hard to produce will graduate into a job market where AI proficiency is already influencing hiring — 39% of physicians say it matters — and where 44% believe they should be the primary financial beneficiaries when AI makes them more productive. Only 2% think payors deserve the upside. (This seems low. It is also probably right.)
The compensation question is awkward. The liability question is worse. Ambient AI scribes are becoming mandatory in some systems, and contracts almost never say who is on the hook when the AI-generated note is wrong. The AMA, ACP, and a bipartisan pair of congressional sponsors all moved this week to say physicians should retain meaningful control over these tools — and that AI governance language belongs in employment contracts, not just policy white papers. Physicians signing multi-year deals are starting to ask about indemnification. Employers who can’t answer are going to look dated, and possibly reckless.
The tension here is structural. Systems are pouring billions into producing more physicians, while the conditions waiting for those physicians on the other side — mandatory AI adoption without liability protection, productivity models that route AI-driven efficiency gains to the employer, contracts silent on governance — may drive the exact attrition the pipelines were built to prevent. You can build a medical school. You cannot build a physician who stays after reading a contract that was written for someone else.
Accelerated three-year MD programs are also gaining traction, compressing entry timelines and betting that primary care demand will still be acute in 2030. The American College of Physicians is now framing these programs as necessary infrastructure rather than experiments. One fewer year of tuition changes the debt math, which changes specialty choice, which changes who ends up in a rural clinic versus an urban subspecialty group. Small lever, long arm.
Training pipelines are being expanded, redirected, and accelerated. Employment terms are being contested clause by clause. Whether those two projects are talking to each other is unclear.
Huntsville Hospital is signing physicians for 2029. Somewhere in a law firm this week, someone is drafting the contract they’ll actually be handed.
P.S. PhysEmp is an AI-powered job board for physicians, which mostly means we try to surface real opportunities without the noise. If you’re a physician tracking where the market is actually going — or a recruiter trying to reach physicians without guesswork — we built the platform to make that a little more legible.
All sources are analyzed and curated from PhysEmp’s industry alert network. AI assists with synthesis and pattern recognition; editorial judgment stays with the PhysEmp Editorial Team. [How we make this newsletter →]