Dr. Melania Poonacha works nights. Real nights — 12, 14, sometimes 16 hours under fluorescent light, charting vitals, fielding pages, keeping people alive in a city that costs more than most people earn in a year just to exist in.
Then she goes home and teaches a machine to be her.
Not in those words, of course. Mercor, the San Francisco company that contracts her time, calls it "AI tutoring." She writes prompts. She grades responses. She quantifies what it means to be a doctor — breaks it into bullet points, ranks symptoms by importance, separates primary from secondary. The AI listens. It learns. It gets better.
She makes $375,000 a year. She had $400,000 in student loans. She's doing this anyway.
The math of being a doctor in San Francisco doesn't add up in the way you'd expect. The city's median salary hovers around $140,000. Physicians earn more than twice that. And still — after debt, after rent, after the compounding arithmetic of a place where even the wealthy feel broke — the numbers come up short. After cost-of-living adjustments, San Francisco ranks fifth-lowest in the nation for physician compensation. Around 40% of doctors hold second jobs.
So they moonlight in the technology that may eventually make them redundant.
There's something almost mythological about it. The craftsman who sharpens the blade that will cut him. The storyteller who writes the book that rewrites him out of it.
Wang — 38, internist, Menlo Park, declined to share his last name — started training AI models in January. He earns $310,000 a year and works 60-hour weeks. After taxes, that works out to roughly $60 an hour. "My friends in the Midwest think I'm rich," he said. "But they own their own houses."
He logs his AI hours at night, after his day job. His reasons are practical. His situation is not unusual. His friends are watching.
The AI companies have their own math. The U.S. spends roughly $5 trillion on healthcare annually. McKinsey estimates that AI could strip out $200 to $360 billion of that through automation — billing, prior authorizations, and clinical guidance. The technology is hungry. It needs real knowledge to eat. It needs board-certified doctors to feed it.
Radiology. Cardiology. Neurology. The specialties most frequently named as ripe for disruption are the ones training the disruptors.
Dr. Krystal Lin didn't sign up for the money. She saw a targeted Instagram ad, felt curious, and went through the intake interview — conducted, fittingly, by an AI. She called the experience surreal. The machine already knew more about occupational medicine than a general practitioner would. She was impressed. She was also a little unsettled.
The work, she found, wasn't hard. What was tricky was the translation — turning the conversational, intuitive language of clinical practice into something a model could parse. No nuance. No bedside manner. Just ranked symptoms and structured criteria, the human logic of medicine flattened into a format a machine could learn from.
She blogged about it. Doctors flooded her inbox. Many asked her to confirm she was human before sending applications.
Dr. Robert Pearl, former CEO of The Permanente Medical Group, puts the reckoning at about 20 years out. By then, he thinks AI replacement is "an inevitability." He also notes that the U.S. is already short 86,000 doctors, with that number growing. The machines and the shortage will meet somewhere in the middle.
In the meantime, doctors keep working the ghost shift — teaching the AI their patterns, their reasoning, their clinical instincts, one graded response at a time. Not because they aren't worried. But because falling behind feels worse than helping it along.
"We're working hand in hand with the AI," Lin said, "to find cures to many of the incurable diseases."
She said it like she meant it. Maybe she does.
The night shift runs until morning either way.