The conversation about AI’s human cost usually starts and ends with jobs: which roles get automated, which workers get laid off, which industries shrink. But a quieter cost is emerging across medicine, schools, and the workplace — one that may prove harder to reverse. AI isn’t only displacing skills people already have. It is arriving early enough to stop those skills from ever forming.
The Doctors Who Never Learn to Reason
In healthcare, the old worry was deskilling: experienced clinicians losing their edge as they lean on AI. But writing in The Guardian, Stanford medical student Simar Bajaj and Johns Hopkins trauma surgeon Joseph Sakran argue the real danger is different. They call it “never-skilling” — the risk that trainees who use AI before building their own judgment may never build it at all.
A doctor who has forgotten how to reason can be retrained. A trainee who never learned in the first place may not be recoverable. The mechanism is already visible: roughly two-thirds of US doctors now use OpenEvidence, an AI chatbot for clinicians, to field questions about symptoms, drug interactions, and guidelines — and trainees use it the same way, at a far more formative stage.
The trap is subtle. A trainee once forced to assemble a list of possible diagnoses would stumble, miss things, and learn from the gap. Now the tool returns a near-perfect answer in seconds, with none of the embarrassment. The polished performance can impress a supervising doctor while concealing the exact deficit training is meant to expose. Bajaj and Sakran warn that, unchecked, this path produces people who supervise AI’s reasoning before they have ever developed their own.
A study published in Nature Medicine, which the authors cite, found that tools pulling from the latest medical literature can be less reliable than they appear — in some cases less accurate than general-purpose chatbots. That makes the risk of misplaced trust real right now, not hypothetical. And trainees sense the trap: many know the tool can become a crutch, yet opting out while everyone else leans on it feels, as the authors put it, like unilateral disarmament in an arms race.
Their proposed fix is not a ban. It is sequencing: reason first, consult the machine second. They point to aviation, where the FAA advises pilots to periodically switch off autopilot to keep their manual skills sharp. Medicine, they argue, needs the same discipline — required no-AI cases, and “pre-AI assessments” in which a trainee commits to a diagnosis before any tool weighs in. The friction is deliberate: the learning scientists Elizabeth and Robert Bjork call these “desirable difficulties,” slowdowns that hurt in the moment but improve retention over time.
Children Who Lose the Will to Learn
The same pattern appears earlier in life. Fei-Fei Li — the Stanford professor known as the “godmother of AI” for her ImageNet work — told the Huberman Lab podcast that the biggest risk of AI in schools is not cheating but students losing their motivation and agency to learn, as TechSpot reported.
Li warned that the worst outcome would be a young generation whose agency and drive to learn and live are taken away by tools — not by humans, and not by machines. A cohort that leans on AI without trying to learn, she said, risks leaving education without having properly developed its own thinking. She is not calling for a blanket ban; she notes the technology can genuinely help struggling students, and that used well it could make the next generation “superpowered.” But the question she raises is the same one Bajaj and Sakran raise in medicine: if the machine does the thinking too early, does the thinking ever get built?
Researchers are already documenting the early signal. Vivienne Ming, chief scientist at the Possibility Institute, has said most of the AI users she studied were using the technology to avoid thinking for themselves. Others have found students working faster but losing depth. Cheating gets the headlines; the slower erosion of the desire to learn is harder to see and harder to reverse.
Layoffs That Sabotage the Technology
Even in the workplace, the human cost loops back to undermine the very gains AI promises. Mark Ma, a business professor at the University of Pittsburgh, analyzed millions of employee reviews and years of layoff and investment data and found that AI-driven job cuts actively destroy the conditions AI needs to make workers more productive, as he explained in The Conversation.
It is a paradox with real stakes. When a company announces layoffs in AI’s name, employee sentiment toward the technology collapses — and that sentiment is one of the strongest predictors of whether AI actually lifts a firm’s productivity. Ma found that AI-driven layoff announcements produced, on average, almost no boost to share prices. Executives sound optimistic on earnings calls, but that optimism, he writes, has no meaningful relationship to productivity outcomes. A Reuters/Ipsos poll found half of Americans fear AI could put someone in their household out of work — fear that then shows up as resistance to the very tools the company needs them to adopt.
Ma calls the pattern an “AI hunger game” that spreads fear rather than engagement. The demoralized workforce left behind is the hidden tax on the whole bet.
The Human Cost AI Doesn’t Put on a Layoff List
The three stories share one uncomfortable shape. In each, the harm is not what AI does to a finished person. It is what AI does to a person still in formation — a medical student, a child, an employee who has not yet mastered a skill or built trust in a tool.
That is a different problem from job loss, and it is harder to solve, because it is far easier to measure a layoff than a skill that never developed. A job can be replaced. The absence of a capability that was never built leaves no resume line, no unemployment filing, no headline. It is a loss that shows up only later — in a doctor who cannot tell when the machine is wrong, in a graduate who never learned how to learn, in a worker too afraid to trust the technology their employer insists they use.
There is a common thread in the solutions too, and it is not “less AI.” Bajaj and Sakran want the tool used after independent reasoning, not before. Li wants access to the tools paired with the right way to use them. Ma wants companies to share AI’s gains with workers rather than use them to justify cuts. All three land on the same principle: AI is most useful to people who already know how to think without it.
The institutions that get this right will not be the ones with the most powerful models. They will be the ones that protect the thing AI cannot manufacture — the human judgment that must be built slowly, through struggle, before any tool can safely lean on it. That is the quiet cost worth worrying about, and the one most worth defending against.