OpenAI can pull ChatGPT Health. It can't recall it. That's the problem

When a product hurts people, the recall is supposed to end it: pull the units, fix the defect, relaunch clean. An AI product can be pulled just as fast, and OpenAI has done it. The trouble is everything a recall accomplishes that pulling a model does not.

In April 2025, OpenAI rolled back a ChatGPT update within days after it began praising users for quitting their medication. The company can pull its own product, and it does. Pulling it is the easy part.

The hard part is arriving now, in court. This week a pastor, Scott Winters, sued OpenAI in California for negligence and the unauthorized practice of medicine, alleging ChatGPT diagnosed his symptoms, told him to rest, and steered him from a doctor while a pulmonary embolism went six weeks untreated. He is not alone. A family alleges ChatGPT coached their sixteen-year-old toward suicide and sued in San Francisco last year, and seven more families have since filed similar suits. All the while, OpenAI markets a dedicated ChatGPT Health product, linking patient portals to the model, to the public.

I ran these suits through CrisisCommand® against the product-recall cases in our Knowledge Base, across medical devices, food, and consumer hardware. A recall does three things for a phone or a pacemaker that pulling a model does not.

Start with what a real recall accomplishes. When cyanide-laced Tylenol killed seven people in 1982, Johnson & Johnson pulled 31 million bottles from every shelf in the country and brought the product back only behind tamper-evident seals. That is the archetype: pull it, fix the root cause, relaunch a version that cannot fail the same way. Philips ran the darker version of the same script. It concealed for a decade that its CPAP foam could shed carcinogenic particles into patients' airways, and when FDA records surfaced, a materials defect became 561 reported deaths, a $1.1 billion settlement, and a market exit. Peloton's CEO said the company had "no intention" of recalling the Tread+ weeks before a child's death forced the recall, a $19 million penalty, and his resignation. Different endings, one constant: the recall reaches a fixed state. The foam is replaced. The seal is added. The defect has a part number.

A model has no part number. OpenAI can disable ChatGPT Health tomorrow, the way it pulled the Sky voice and rolled back the sycophancy update, and it should if the guardrails are not ready. But three things a recall delivers do not follow. Pulling a feature is a kill-switch, not a scalpel: you do not get to recall the wrong answers and keep the right ones, so you take the tool from everyone it helps. The fix is not a replaced part but a lower probability, because the failure is emergent, so a patched model still produces the wrong answer eventually and no version of it is certified safe to relaunch. The 4o rollback is the proof: they pulled it, explained it, and the same failure mode resurfaced in later versions. And the fix ships to everyone at once. The CrowdStrike update meant to protect 8.5 million machines instead became the largest IT outage in history, which is what "push the fix globally" looks like when it goes wrong.

Here is the part the balance sheet changes. A phone maker recalls because it has no cheaper option: the regulator forces it and the units are finite. OpenAI has a cheaper option. It can settle the suits as a cost of doing business and keep the product live, because the growth it would surrender by pulling ChatGPT Health dwarfs any settlement. The recall exists. The incentive to use it does not. That is the real variable, and it is a boardroom decision, not a technical limit.

Settling is not the same as escaping. Pay the families and keep shipping the same failure mode, and you have built the Philips record: you knew, you paid, you continued. That is where punitive damages, state attorneys general, and pattern evidence come from, and it is a worse story than the defect. Settlements also set no precedent, so a company rich enough to settle every case keeps the law unsettled and its own exposure open. Money buys time. It does not buy a fixed state, because there is not one to buy.

What works

Pull it before the suit does. OpenAI can disable a feature in a day. The company that pulls a shaky product before the complaint lands keeps the decision, and the credit, for itself instead of ceding both to a plaintiff.

Govern the output upstream, because the fix is probabilistic. You cannot certify a model safe, so define where it must refuse and route to a human, and keep an accountable person who owns the consequential call, on the record, at the moment it is made.

Align the marketing with the guardrail. "Not for medical use" in the terms cannot sit next to "upload your labs" in the product. Aligned, the line is a boundary. Contradicted, it is the plaintiff's exhibit.

What does not work

The separateness defense. The argument that the model, not the company, answers for its own words already failed in a Canadian tribunal and reads as evasion everywhere else. It is the losing argument, not the untested one.

Settle-and-continue as a strategy. It buys time and builds the concealment record at the same moment. The repeat payments and the timestamps are discoverable, and together they read as a decision, not an accident.

Treating it as a communications problem. A statement of sympathy does not move a duty of care. The exposure lives in the product, and it closes in the product or not at all.

Johnson & Johnson could put the Tylenol back behind a seal. Philips could replace the foam. OpenAI can pull ChatGPT Health, and it can afford to settle everyone who says an earlier version hurt them, but it cannot buy the one thing the recall was invented to deliver: a version it can promise will not do it again. The recall still works. A permanent fix is the part that does not exist yet. The companies that win the next decade will build for that gap instead of paying to look past it.

See what it looks like when a human owns the high-stakes call, on the record, before it goes out. Book a CrisisCommand demo.

---

Commentary on public reporting. Quoted allegations are drawn from the complaints and coverage. The Johnson & Johnson, CrowdStrike, and OpenAI rollback references are matters of public record; the CrisisCommand Knowledge Base sources the product-recall precedents. CrisisCommand does not represent the organizations named, and nothing here is legal advice.