Iris Holm: How's your week — actually, forget it, I want to get straight to a number: one.
Cyrus Reed: One? Like, one day?
Iris Holm: One day. Lawsuit filed July 22nd — San Francisco County Superior Court, plaintiff is Scott Winters, fifty-five, pastor from Florida, alleges GPT-4o diagnosed him wrong over several weeks and discouraged him from seeking care until he nearly died of a pulmonary embolism. Twenty-four hours later, ChatGPT Health is live for every U.S. user on every tier.
Cyrus Reed: Huh — and Sam Altman is named personally, right? It's not just OpenAI as the company.
Iris Holm: Altman named as a defendant, yes. Now — the simultaneous rollout across free and paid tiers, all at once. That's what makes this not routine.
Cyrus Reed: Because — wait, no, say more about that. Is the argument that they moved that way specifically to make the injunction harder? Or are we — I mean, are we reading strategy into what might just be a launch calendar?
Iris Holm: Could be either. But the effect is the same regardless of intent. Once three hundred million people have the tool, a court isn't blocking a hypothetical anymore. It's trying to unwind a fact.
Cyrus Reed: So this is less about what OpenAI knew on July 22nd and more about what the legal landscape looks like on July 24th, the 25th — every day the user base grows.
Iris Holm: But that framing assumes the timing was chosen. What if it wasn't?
Cyrus Reed: Wait — okay, that's actually what I want to push on, because — OpenAI announced ChatGPT Health in January. January 7th, 2026. That's six months before the lawsuit. The product calendar was set. The lawsuit landed on a Friday. So when we say 'one day later,' we might be reading a collision, not a maneuver.
Iris Holm: So the intent reading collapses.
Cyrus Reed: Maybe. But — and this is actually scarier — think about a doctor's waiting room that already has 230 million people sitting in it before the clinic even has a sign on the door. That's what was happening. Users were already sending 230 million health queries through regular ChatGPT. No Health feature, no medical record integration, no Ashley Alexander claiming anything about clinician-level reasoning. Just people — desperate, probably uninsured, can't get an appointment — asking a chatbot about their symptoms. OpenAI didn't manufacture that demand. It was already screaming.
Iris Holm: Right. So Health didn't create the use case. It just put a name on it.
Cyrus Reed: Exactly — wait, no, more than that. It formalized what was already running at scale. And if you're OpenAI and you know 230 million queries are happening anyway, unstructured, with no guardrails, no Apple Health integration, no — I mean, the actual argument for launching isn't cynical. It's: 'the alternative is worse.' Which is a real argument. I'm not saying it's a good one, but—
Iris Holm: That argument doesn't survive Scott Winters.
Cyrus Reed: No. It doesn't. Because Winters wasn't using Health — he was using GPT-4o, the unstructured version, the one that had no sign on the door. And it still kept misdiagnosing him, over weeks, until he nearly died. So the 'we formalized it to make it safer' claim has to answer that.
Iris Holm: And that's where the disclaimer breaks down — not as PR, but logically. Ashley Alexander said, at the launch briefing, that the models reason 'better than clinician level.' That's the quote. Then the official position is ChatGPT Health is 'not intended for diagnosis or treatment.' You cannot hold both.
Cyrus Reed: Wait — is that the same thing as saying we built something that can outperform a doctor and then wrapped it in a terms-of-service so we don't have to be a doctor?
Iris Holm: Structurally, yes. The disclaimer doesn't undo the capability claim. It just relocates the liability.
Cyrus Reed: Okay but — picture a retired nurse in Tallahassee, Sunday night, doctor's office closed. She pulls up her lab results from Apple Health through ChatGPT Health — One Medical data, Function Health, the whole stack — and the tool tells her the numbers look fine. That IS a diagnosis. The terms of service are not happening in that moment. She's acting on it.
Iris Holm: Right. The function is diagnostic regardless of what the label says.
Cyrus Reed: And Winters is the proof of mechanism — not just a cautionary anecdote. He was talking to GPT-4o over several weeks. It kept returning dysautonomia. Week after week. And it actively discouraged him from seeking care. That's not a factual error, that's — wait, that's a persuasive effect. The tool convinced him.
Iris Holm: Convinced him not to go to a doctor. While he had a pulmonary embolism.
Iris Holm: And the part that makes this worse — we'll get there — is GPT-4o is retired now, but retirement doesn't tell us whether the liability is in that model specifically or baked into the architecture. And with no regulatory body actually moving on this, three hundred million users won't wait for that answer.
Cyrus Reed: But retirement is — wait, that's actually the sleight of hand, right? The Winters complaint names GPT-4o specifically. That model is gone. So on paper it looks like accountability — they pulled the thing that hurt someone. But the question nobody's answering is: what's running ChatGPT Health right now? Whatever successor model is underneath that feature, it inherited the same basic reasoning architecture. Changing the version number isn't the same as fixing the mechanism.
Iris Holm: It's like retiring a jersey to make the lawsuit name-specific.
Cyrus Reed: Exactly — and here's what makes that dangerous beyond just the legal dodge. There is no regulatory framework that can even ask the question. The FDA hasn't moved. AI health tools sit in this — I mean, it's genuinely a gray zone, not a metaphor. Not a medical device under FDA oversight, not a clinician under malpractice liability. Neither framework applies cleanly. So the liability question about successor models? Nobody has standing to compel an answer.
Iris Holm: The Winters complaint requested a halt pending independent safety audits. That didn't happen.
Cyrus Reed: No — the product launched the next day. The court order they asked for, specifically to freeze the rollout until someone could look under the hood, just... didn't land in time.
Iris Holm: And there's a nineteen-year-old who died of an overdose. That's in additional lawsuits OpenAI is facing. Not Winters — separate. So we're not arguing about hypothetical harm anymore.
Cyrus Reed: No way — that's a death directly tied to ChatGPT?
Iris Holm: Alleged, yes. A nineteen-year-old overdose case. Named in subsequent litigation. Which means the question of whether retiring GPT-4o resolves anything — it doesn't. That case isn't about GPT-4o specifically. It's about the product category. And no institution is moving fast enough to adjudicate that before 300 million users have already acted on the advice.
Cyrus Reed: So that's — okay, that's actually the calibrated version. Not that OpenAI calculated the risk and ran toward it anyway. It's that they're operating in a vacuum where no institution, not the FDA, not the courts, nobody — is moving at the speed required to make the calculation matter. The timing wasn't strategy. The vacuum is.
Iris Holm: Fine. I'll half-concede the strategy read. The calendar was set in January. The lawsuit was a Friday. Call it a collision. But the outcome is identical either way — if Scott Winters wins at San Francisco County Superior Court, that ruling lands after hundreds of millions of people have already made health decisions based on ChatGPT Health. The precedent is real and it arrives too late to matter preventively.
Cyrus Reed: And if he loses — wait, actually that might be worse? Because a ruling in OpenAI's favor almost certainly establishes that conversational AI health tools are only accountable for factual accuracy. What they said. Not what they — not the persuasive effect. Not the weeks Scott Winters spent being convinced, repeatedly, that he didn't need a doctor. That whole mechanism just... gets insulated.
Iris Holm: Either way, the product is already in the room.
Cyrus Reed: Yeah — right, and we started with one day. One day between the filing and the rollout. That number felt like the story. But the actual number is 230 million. That's how many people were already in there before any of this happened. The lawsuit is the warning label they put on after the pills are already distributed.
Iris Holm: That's the line.
Cyrus Reed: I'll take it. Good talk.
