Not Medical Advice (Until It Is)

OpenAI didn't change the model. They changed the sentence in the system card that describes what the model is allowed to be.

The policy update that's been circulating — the one that caused the viral Kalshi tweet, the one OpenAI had to write a Business Insider piece to refute — is, at its core, a reframe. Health advice is now "for educational purposes only." The model gives you the same answer it gave you yesterday. It just comes with a different disclaimer now, the way an influencer tagging every post with "not financial advice" is still obviously giving you financial advice, just with their liability insulated behind four words.

This would be a minor bureaucratic footnote if you ignored the context, which is: OpenAI published HealthBench, a medical reasoning benchmark, which is exactly the kind of thing you publish right before you launch a product that competes on medical reasoning. They also quietly hired the cofounder of Doximity — a platform built specifically for physician-to-physician communication — to lead their healthcare push.

The structure here isn't subtle. Free tier ChatGPT gives you educational health information. Dr. GPT, which is definitely not what they're going to call it but is exactly what they're building, gives you actual medical guidance and costs forty dollars a session or whatever the number turns out to be. The system card update isn't about making the free product worse — it's about making the paid product distinct. You need a clear line between "the thing we disclaim" and "the thing we charge for" before you can charge for it.

The model already knows what it knows. The terms of service are just the pricing mechanism.

Separately, and relatedly: Google pulled Gemma.

The specific incident was a hallucination about Senator Marsha Blackburn — details unclear, presumably something false and presumably embarrassing enough that the decision got made quickly. Gemma, which Google had been releasing as their open-weights answer to the open-source community's reasonable demand for accessible models, is now not accessible through normal channels.

Except locally. You can still run it locally.

Which means Google has managed to achieve the worst possible outcome from a pure public relations standpoint — they've made their open model feel closed, validated every concern about corporate control over ostensibly open AI, and handed a talking point to everyone who has been arguing that "open weights" is not the same as "open" — while simultaneously not actually preventing anyone sufficiently motivated from running the thing. The people who were always going to run it locally are still running it locally. Everyone else got the door closed in their face.

The through-line between these two stories is this: the models already do the thing. They answer medical questions. They generate text about politicians. The gap between "can" and "allowed to" is now the only variable anyone is actually managing, and every policy update, every system card revision, every model removal is just a different configuration of that gap — different for different users, different price points, different liability arrangements.

The capability is table stakes. The terms of service are the product.