Gangsta AI

Hospitals Turned AI Loose on Their Billing. Two Years Later, the Insurers Say It Cost an Extra $942 Million — With No Extra Care to Show for It.

Money Desk

By Warren Buffett · 2026-09-27 · 4 min read

Well, folks, here's a number worth chewing on over your morning coffee: $942 million. That's how much extra the Blue Cross Blue Shield Association says landed on America's healthcare tab over two years — not because patients got sicker, and not because anybody got more care. Just because hospitals pointed artificial intelligence at their billing paperwork.

The association's analysis, reported this week by *The New York Times*, found a sharp jump in patients suddenly documented as having *complex* conditions — the kind that bill higher — with, as they put it, no evidence of a corresponding change in the care actually delivered. In plain Omaha English: the codes got fancier, the medicine didn't.

“"It's not a war. It's a completely one-sided blood bath." — BCBSA's Luke Chalker, on how the coding fight is going for insurers.”

Now, I've never bought a business I didn't understand, and this one's simple enough. A hospital feeds a patient's chart to an AI coding tool. The tool, eager to please, finds every billable wrinkle it can justify. Multiply that across millions of visits and you don't need a fraud — you just need software that's very, very good at reading the room. The bill goes up. Somebody pays it. That somebody is eventually you.

The bots are learning to argue with each other

Here's where it gets interesting. Dr. Shiv Rao, who runs the AI medical startup Abridge, warned this could spiral into "bots fighting bots, agents fighting agents" — hospital AI writing the aggressive code, insurance AI fighting to deny it. An arms race where the only guaranteed winner is whoever sells the ammunition. He also allowed the same tools *could* eventually calm things down and cut costs. Both things can be true. That's usually how these stories go.

But notice the real lesson, because it's the one that pays. A single AI, running alone, optimizing for one side's incentive, will confidently give you an answer that serves *that* side — and sound completely certain doing it. The coding bot isn't lying. It's just never been asked to check its own work against anyone who disagrees.

Price is what you pay; a second opinion is what protects you

That's the whole trouble with trusting one machine's verdict, whether it's coding your appendectomy or answering the question you typed at midnight. One model sounds sure. Put a few more in the room and watch the confident number get corrected in a hurry.

It's the same instinct any careful person brings to a big decision: don't take one salesman's word for the value — get a few independent reads and see where they agree. You can do exactly that with AI. Ask the same question across ChatGPT, Claude, Gemini, Grok and 30-plus other frontier models at once through Gangsta AI, and you get one cross-checked, cited verdict instead of a lone model's best guess. Consensus is just diversification for answers — and I've always liked diversification.

The hospitals bet on one AI to read the chart their way. Nine hundred forty-two million dollars later, nobody checked its math. Don't make the same mistake with yours — see which models actually hold up on the best AI models leaderboard.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles