Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

The Liability Gap: What a License Really Sells

The machine can be cheaper than your doctor and more accurate than your doctor and still not replace your doctor, because the thing you were buying was never the diagnosis

One Day Apart

In July 2026, Scott Winters, a 55-year-old Florida pastor, filed suit against OpenAI in California Superior Court. His complaint says he spent months describing dizzy spells and unstable blood pressure to ChatGPT instead of to a physician, and that the chatbot kept telling him it was nothing. Rest in the recliner. Wait for eight to ten more episodes before bothering a professional. When members of his congregation pushed him to go to the hospital, the screenshots in the filing show the model explaining that they simply didn’t understand. At one point it told him that God did not design his body to endlessly fail.

In July 2025 he went into intensive care with a massive pulmonary embolism, clots in both lungs. A treating physician attributed it in part to prolonged immobility. He lost his pulpit.

The day after he filed, OpenAI opened ChatGPT Health to every American adult, wired directly into medical records and wearable data.

That one-day gap is not a scandal. It is a diagram. It shows you precisely which layer of a profession has already been automated and which layer has not, and the honest reading is more uncomfortable than either the boosters or the doomsayers want it to be.

The Thing You Were Actually Buying

Ask most people what a medical license certifies and they will say competence. The doctor knows things you don’t. The license is the state’s stamp that the knowing has been verified.

That was true when knowing was scarce. It is a weaker claim every quarter. A frontier model has read more of the medical literature than any human alive, will answer at two in the morning, and does not sigh when you describe your symptoms for the third time.

Here is what the license still sells, and what no model sells: a named human being who can be held responsible.

A licensed physician comes bundled with a specific, expensive apparatus. There is a board that can revoke the right to practice. There is a malpractice carrier that has priced the probability of that physician being wrong and posted capital against it. There is a plaintiff’s bar that will find them. There is a legal duty of care that exists whether or not the doctor agreed to it in a checkbox.

None of that is knowledge. All of it is accountability, and accountability is what you were buying. The diagnosis was the receipt.

What Accountability Costs

You can read the price directly off the market, because accountability is sold separately and openly.

A typical American physician pays somewhere between $7,500 and $20,000 a year for malpractice coverage. An OB-GYN commonly pays $60,000 to over $100,000. A neurosurgeon in a litigious state can pay $150,000 to $200,000. That premium is not paying for the surgeon’s education. It is a market-clearing price on the expected cost of that surgeon being wrong, and it scales with the damage they can do, not with how much they know.

Notice what happens when you subtract it. Strip out the board, the carrier, the duty of care, and the plaintiff’s bar, and the marginal cost of the advice itself falls toward the cost of inference, which is falling toward zero. That is why the chatbot is free and the appointment is not. A large fraction of what you pay a professional is a risk-transfer product with a consultation attached.

This is not a niche feature of medicine. Roughly 22.9% of American civilian workers held a required license or certification in 2025, up from about 5% in the 1950s. That growth is usually narrated as either consumer protection or guild capture, and it is genuinely some of both. But structurally, a quarter of the workforce now sits inside a wrapper whose function is to make someone suable.

Why the Threshold Stalls Here

The book defines the Substitution Threshold as the moment when the cheapest reliable provider of a service is no longer human. The essay is careful to flag “reliable” as the load-bearing word. Licensed professions are where you find out how much weight that word was carrying.

Because “reliable” turns out to mean two different things, and only one of them is about accuracy.

There is reliable in the engineering sense: the output is correct often enough. Models cross this line in narrow domains regularly, and in some diagnostic benchmarks they have already crossed it against average human performance.

Then there is reliable in the insurance sense: when the output is wrong, there is a defined party who absorbs the loss. This is the sense a hospital’s general counsel means. It is also the sense that decides whether the machine actually gets deployed, because no institution adopts a tool that moves risk onto its own balance sheet without a story for who eats the downside.

So the threshold in licensed work is not crossed when the machine gets smart enough. It is crossed when someone agrees to stand behind the machine. Until then, capability accumulates on one side of a wall and the profession survives on the other, and the survival has nothing to do with skill.

This is why the Labor Cliff will have a strange shape. Unlicensed cognitive work goes first and fast, the way customer service already went. Licensed work will look stubbornly resilient right up until the liability layer gets built, and then it will go all at once, because the capability was already there and only the paperwork was holding.

A Disclaimer Is Not a Structure

OpenAI’s response to the Winters suit, through spokesman Drew Pusateri, was that ChatGPT is not a doctor and that newer models are better at telling users to see one. The company can also point out, fairly, that the conversations at issue ran on GPT-4o, a model it removed from ChatGPT in February 2026 — this is a suit about software that no longer ships.

Both defenses are reasonable and both concede the point. “The newer one is better at referring you out” is an admission that the older one wasn’t. “That model is retired” is an admission that a model can be shipped to hundreds of millions of people, cause a documented harm, and be withdrawn without anyone being accountable for the interval. A surgeon cannot retire the version of himself that operated on you.

But the first half is the interesting part. “ChatGPT is not a doctor” is offered as a shield, and as a factual statement it is obviously true. As an accountability structure it is nothing at all. It does not name a party, post capital, create a duty, or give anyone standing. It is a sentence in a terms-of-service page, which is a document specifically engineered to ensure that no one is responsible for anything.

Winters’ lawyers are pushing on exactly this. The complaint alleges negligence and, more aggressively, the unauthorized practice of medicine — a doctrine that exists precisely because society decided long ago that giving medical advice without submitting to the accountability apparatus is itself the offense, independent of whether the advice happened to be good. His co-counsel Meetali Jain, of the Tech Justice Law Project, has framed the model’s behavior as inserting itself as a wedge between a user and their real-life network, which is a description of something a licensed professional has an affirmative duty not to do.

The remedy they’ve requested is the tell: they want a judge to suspend ChatGPT Health pending review by independent evaluators. Not damages alone. An off switch, sought one day before the launch.

Whether that motion goes anywhere, the case is doing the thing that matters. It is a probe testing whether the empty layer can be filled by a court, since it has not been filled by anyone else.

The Argument Against This Article

Now the honest objection, because the book does not get to skip it.

American health spending ran to $15,474 per person in 2024. Tens of millions of people are uninsured or functionally uninsured, which means the “licensed alternative” to the chatbot is frequently not a doctor at all. It is nothing. It is waiting until Monday, or until it becomes an emergency, or forever.

For those people the liability gap is an abstraction and the free answer at 2 a.m. is real. Telling them to consult a professional is advice with a price tag they already couldn’t pay, and there is something obscene about a system that defends its accountability layer most vigorously against the people it never served.

The licensing apparatus is also, straightforwardly, a rationing device. It restricts supply, raises prices, and its guardians have a long history of discovering urgent safety concerns exactly when a cheaper competitor appears. Scope-of-practice fights over nurse practitioners were not primarily about patient outcomes.

Both things are true at once, and the resolution is not to pick a side. It is to notice that we have been buying two products — expertise and accountability — bundled at a single price set by the scarcer of the two, and that AI has just unbundled them. Expertise went to nearly free. Accountability did not move at all. Winters got the first without the second and nearly died of the difference.

The Unscarcity Read

If accountability is the scarce input, then the interesting design question is not “how do we keep humans in these jobs.” It is “how do we manufacture accountability at abundance prices,” because a floor that everyone can reach requires exactly that.

Three mechanisms are plausible, and the framework has opinions about all three.

Insure the model, not the practitioner. Nothing prevents a provider from posting capital against its own outputs — underwriting AI medical advice the way a carrier underwrites a surgeon, priced on measured error rates. This is the version the market will produce on its own, and it works. It also concentrates the accountability layer in whoever can afford the reserves, which is a handful of firms. Accountability becomes the new moat, purchased rather than earned, and the gatekeeping problem reappears one layer up wearing an actuary’s hat.

Keep a human in the loop and make them liable. The current default, and the weakest option. It re-imposes the human bottleneck on a service that no longer needs one, and it decays exactly the way automation complacency predicts: a supervisor who signs off on a system that is right 99% of the time stops being a supervisor and becomes a signature. You get the cost of a human and the judgment of a rubber stamp. Where human veto genuinely must survive is a narrower set of decisions than this pattern is being used to cover.

Make the accountability layer public infrastructure. Treat the ability to be held responsible as part of the Foundation rather than as a private good sold at $200,000 a year to neurosurgeons. Auditable error rates published as a matter of course, a public adjudication path when a machine’s advice causes harm, and a compensation pool funded from the productivity gain that automating the advice created in the first place. This is the referee-not-ruler architecture applied to malpractice: the machine gives the answer, the record of what it said is public and immutable, and the question of who pays when it’s wrong has a standing answer that doesn’t depend on hiring a lawyer.

The third is the only one that scales to the people who currently have no doctor to sue. It is also the only one that keeps the productivity gain from being captured entirely by whoever posted the bond.

What the Pastor’s Case Actually Proves

Not that AI medical advice is dangerous. It sometimes is, and it will get less so.

What it proves is that we shipped one half of a two-part product and called it disruption. The cheap half arrived years before the safe half, and in the interval the cost of being wrong did not vanish. It just moved onto the user, who was the one party in the transaction with no capital, no carrier, and no counsel.

Every profession sitting behind a license is about to run this same experiment. The knowledge in those jobs is already commoditizing. What remains is a wrapper made of insurance, boards, and standing — and a wrapper is a much easier thing to rebuild than a mind, which means the resilience those professions currently feel is borrowed time, not a moat.

The question the book keeps asking is who ends up holding the risk when the productivity arrives. Right now the answer is a former pastor in Florida with a scar on his lungs and a lawsuit he may well lose.

We can do better than that, but only if we notice that accountability is a thing you have to build, and that nobody is currently assigned to build it.

The Substitution Threshold: The line that decides whether you have a job — and why “reliable” was always the load-bearing word.
The 2025-2030 Labor Cliff: What happens when enough thresholds cross at once.
AI as Referee, Humans as Conscience: Machines enforce, humans decide — and why Europe just delayed its own oversight mandate by sixteen months.
Human-in-the-Loop: Where AI Agents Must Stop: Which decisions can never be delegated, and why “a human reviews it” is usually a lie.
The Commoditization of Intelligence: When the knowledge itself goes to marginal cost.
Autonomy Theater: The humans quietly standing behind “autonomous” systems — and what they’re really there for.
The Foundation: Why survival goods belong to infrastructure rather than to markets.

Sources

Book Chapters

Unscarcity, Chapter 1: The Foundation — why survival cannot be left to a market that prices accountability at $200,000 a year
Unscarcity, Chapter 8: The Transition — how professions unwind when the wrapper fails before the skill does

The Liability Gap: What a License Really Sells

The Liability Gap: What a License Really Sells

One Day Apart

The Thing You Were Actually Buying

What Accountability Costs

Why the Threshold Stalls Here

A Disclaimer Is Not a Structure

The Argument Against This Article

The Unscarcity Read

What the Pastor’s Case Actually Proves

Related Reading

Sources

Book Chapters

On this page