Right now, a specific group of professionals is having a very good few years. Developers ship more without losing what made them valuable. Lawyers bill the same hours for higher-value work, having offloaded document review. Radiologists read more scans in less time, judgment intact, demand at record highs. This is real, measured, and worth celebrating rather than resenting.

It's also, in a specific and describable sense, temporary. This piece is about three different kinds of safety: one that's already quietly starting to close, one that was never really about capability at all, and one worth paying closest attention to, because it's the only kind with no expiration date.

Three mechanisms do the actual protecting, and it's worth naming them before going further. Liability-based protection is institutional, and institutions can be redesigned, which makes it temporary. Origin-based protection is psychological, built into how humans assign value to a thing's source, and it doesn't erode the way an institution does. Underneath both sits prepared-mind protection, the capacity to notice that something unexpected matters, which is the one this piece argues is hardest to hand to a machine in any meaningful sense.

Liability-based protection: a human decision, not a technical wall

Consider IDx-DR, a diagnostic tool cleared to screen for diabetic retinopathy without a physician in the loop at all. No radiologist reads the image. And yet liability didn't disappear when the human did, it moved. The company behind the device carries a product liability policy and an indemnity clause, so that if something goes wrong within the tool's approved use, the company, not a doctor, absorbs the consequence.

That's the whole mechanism in miniature. Liability isn't a wall that keeps AI out of a profession. It's a question of who is willing to be held responsible, and that question gets answered differently as tasks get narrower and more predictable. Diabetic retinopathy screening is close to binary, a clear yes or no, narrow enough that an insurer could price the risk with confidence.

Reading an ambiguous chest CT for a rare, atypical presentation isn't narrow in that way, and nobody is underwriting that kind of open-ended risk yet. That's why radiology as a whole hasn't been replaced even as this exact liability transfer became real and FDA-approved for a specific slice of it. The transferable part is still small. The untransferable part is still growing.

But "still small" is a snapshot, not a ceiling. As underwriting infrastructure matures, as more narrow, low-variance tasks become priceable, the boundary moves. It moves task by task, not profession by profession, so the erosion won't feel like a single cliff. It will feel like a slow narrowing of what's left that a person has to do.

Some of today's protected professionals will end up doing something adjacent instead: validating the systems that replaced their narrow tasks, consulting for the manufacturers who now carry the liability they used to carry personally. That's not the mass exodus agriculture went through a century ago, when farm labor didn't move to a nearby niche within farming, it left for entirely different industries. This is closer to an expert moving up the stack, from doing the task to standing behind the system that does it.

None of this is a prediction of collapse. It's a description of a window with a shape, wide now, narrowing over time, closing unevenly rather than all at once. But liability only protects what can be priced. Origin protects what can't.

Origin-based protection: a different kind of safety, and it doesn't narrow at all

In 1937, a Dutch painter named Han van Meegeren produced a painting so convincing that it was celebrated for years as a lost Vermeer. Experts admired it. Critics wrote about its mastery. Then van Meegeren, facing a treason charge after the war, confessed to the forgery to save himself, and had to paint another fake in front of witnesses to prove it. The moment the painting's origin changed in people's minds, its value collapsed. Nothing about the object had changed. Not a single brushstroke was different from the day before. Only the fact of who made it had.

Psychologists Newman and Bloom ran a series of experiments in 2012 asking exactly why this happens, and found two real mechanisms: people value an object as evidence of a unique creative act, and they believe that physical or causal contact with a person transfers something of that person into the object, a kind of essence. Tellingly, this effect shows up even in young children, and it does not apply to mass-produced things. A duplicate car isn't devalued the way a duplicate painting is. It's specifically unique, singular, human-made things that carry this weight.

This is why a live concert performed by a machine and a streamed song generated by one aren't the same question, even though both involve music. Recorded, reproducible music was never in the category this effect protects, which is why blind listening tests find most people can't tell AI-generated tracks from human ones, and why AI music is climbing toward half of all daily uploads on some platforms despite listeners saying, sincerely, that they prefer human-made work.

A live performance is the opposite case: unrepeatable, embodied, impossible to duplicate by definition. Ask whether a hundred thousand people would fill a stadium to watch a machine perform, whether they'd wear its merchandise, whether they'd call it a hero. The answer isn't really in question, not because a machine couldn't produce the same notes in the same order, but because heroism and reverence were never about the notes. They were about a person who risked something to make them, who has a life the audience can attach meaning to, who could fail in a way that matters.

This is a different kind of protection than the liability story. Liability is an institutional arrangement, and institutions change their arrangements. Essence isn't an arrangement at all, it's a fact about how humans assign value to origin, demonstrated in children before they've learned any market logic. There's no version of AI capability, however advanced, that closes this gap, because the gap was never about capability to begin with.

Prepared-mind protection: the last thing worth protecting isn't a task, it's a kind of noticing

In 1896, Henri Becquerel left some photographic plates near a sample of uranium salts, expecting nothing in particular, and found them fogged when he came back to check. A great many people, at a great many points in history, have looked at a ruined result and thrown it away. Becquerel didn't. He recognized that the fogging meant something, and pulled on that thread until it became the discovery of radioactivity.

This is the quieter, less celebrated half of nearly every story like it, penicillin, X-rays, vulcanized rubber, the microwave oven. The accident is never the discovery. The discovery is the person who was paying enough attention, and had enough reason to care, to notice that an accident was interesting rather than merely wrong. Researchers who study this call it a prepared mind, and it's the actual scarce ingredient, not the error itself.

It isn't only a historical pattern. In 2004, physicists Andre Geim and Konstantin Novoselov were using ordinary Scotch tape to peel layers off a lump of graphite, low-stakes tinkering they called "Friday night experiments," unrelated to their actual funded research. Nobody had thought sticky tape and pencil-grade graphite were worth taking seriously. When their hand-made flakes showed unexpected conductivity, they didn't dismiss it as a fluke. That noticing became graphene, a material a hundred times stronger than steel, and won them the 2010 Nobel Prize in Physics.

Machines already produce plenty of errors, unexpected outputs, results nobody asked for. What they don't yet do, and what this piece argues they structurally can't do the same way, is care about one anomaly over another for reasons that come from being a person with stakes in a world that matters to them.

This is the same shape as everything else here, just at a larger scale. The litigator's value was never the paperwork, it was the judgment exercised in a courtroom under real stakes. The writer's differentiator is relocating from drafting to editorial judgment, from making the sentence to deciding whether it's true and worth keeping.

The radiologist's protected core was never the act of looking, it was the accountable call about what the image means for this particular patient. And now, at the widest scale available, the mechanism by which humanity gets smarter over time was never the making of mistakes. It was the noticing that a mistake mattered.

Together, these three protections form the real boundary between what erodes and what endures.

What's left, once the window closes

Picture the long arc all the way out. The tasks that can be narrowed, priced, and underwritten eventually get automated, not because AI became conscious or superior in some cosmic sense, but because someone, somewhere, became willing to carry the risk instead of a person. That's not a tragedy. It's the same story as the calculator, the assembly line, the spreadsheet, each one narrowing what a specific kind of expert has to personally do, without ending the need for the expert entirely.

What doesn't narrow is the territory built on origin, presence, and stakes. A relationship where the other person's specific, irreplaceable attention is the entire point. A live performance where the person's history and risk are inseparable from what's being watched. A discovery that required someone who cared enough to notice.

None of this is safe because a machine technically cannot replicate the surface behavior, in many cases it already can, or soon will. It's safe because the value was never located in the behavior. It was located in the fact of a particular person being on the other end of it, and that fact isn't something capability can manufacture, no matter how far the capability goes.

That's the actual shape of the long run: not humans holding a shrinking set of tasks machines haven't reached yet, but humans finally standing only in the places where being human was always the entire point.

Sources

• Everlaw, 2025 Ediscovery Innovation Report, on liability transfer via IDx-DR and FDA-cleared autonomous diagnostic tools: www.everlaw.com/blog/ai-and-law
• Deena Mousa, "AI isn't replacing radiologists," on IDx-DR's liability structure and radiology employment data: understandingai.org
• Newman, G. E., and Bloom, P., "Art and Authenticity: The Importance of Originals in Judgments of Value," Journal of Experimental Psychology: General, 2012: researchgate.net/publication/51797088
• On the van Meegeren Vermeer forgery case, as discussed via Paul Bloom's essentialism research: screwdowncrown.com
• BPI 2025 music consumer survey and Deezer AI-upload data, on stated preference versus revealed behavior toward AI-generated music: botmemo.com/ai-music-statistics
• On serendipity in scientific discovery, the "prepared mind" mechanism, and the Becquerel radioactivity case: journal.trialanderror.org, azolifesciences.com
• On Geim and Novoselov's 2004 discovery of graphene via "Friday night experiments" and the 2010 Nobel Prize: nobelprize.org/prizes/physics/2010/speedread, oxsci.org