Replace or Relieve? The Nurse and the Machine

The team decides how the machine works on the ward. One question remains, and for most people it is the first one. If a machine can do this much of what a nurse does, what happens to the nurse? This last chapter weighs what the machine does better, what it cannot do, and what researchers say.

The question is not “nurse or machine?” but “which tasks, decided by whom, for whose benefit?” The same technology can make a nurse’s job emptier or more human.

Two hospitals

Imagine two hospitals that buy the same AI system the same year.

In the first, managers look at the numbers. The machine answers the 2 a.m. helpline. It drafts discharge letters, writes up observations and flags deteriorating patients. So the hospital trims nursing posts. The nurses who remain cover more beds, check the machine’s work, and spend their shifts at keyboards approving alerts. The job becomes thinner, faster and lonelier.

In the second, managers ask a different question: what were our nurses doing that only a nurse can do, and what kept them from doing it? Studies of hospital work find that nurses spend roughly a quarter to two-fifths of a shift on documentation. The machine takes most of that. It handles routine questions overnight and passes anything uncertain to a human. The nurses spend the recovered hours at the bedside: explaining, noticing, comforting, catching the patient who “looks wrong”. Staff turnover falls.

Both hospitals are possible with the same technology. This chapter tries to say which tasks the machine is good at and which it is not. It then sets out what researchers are saying.

As throughout the book, the nurse stands for more than nursing. Teachers, social workers, lawyers, pharmacists and many others do jobs that mix information work, which machines are getting good at, with presence, judgment and responsibility, which they are not. The same two futures are open to all of them.

What the AI nurse does better, faster or cheaper

• It is always there. It does not tire at 4 a.m., does not get irritable, and answers in seconds. An ordinary chatbot prompt uses about nine seconds of television’s worth of energy (Chapter 13); a system that reasons and calls tools uses more.
• It does not get tired of checking. A model’s own recall is fallible (Chapter 14), but a database does not forget: a drug-interaction table, a protocol library, a patient record. The strength comes from connecting the model to those records (Chapter 12).
• It writes well, and kindly. In a 2023 study, health professionals preferred a chatbot’s replies to patients’ online questions over doctors’ replies in most cases. They also rated them empathetic far more often. But the doctors were answering quickly on a public forum, not in a clinic.
• It can hold a diagnostic conversation by text. In a blinded 2025 study with actors playing patients, specialists rated Google’s research system AMIE better than primary-care doctors on most measures. These were simulations, not real patients.
• It takes paperwork away. “Ambient scribes” listen to a consultation and write the notes. In a 2025 study, clinicians using them for a month reported markedly less burnout and more undivided attention for patients. (They were doctors and similar practitioners; tools for nursing notes are newer.)
• It spreads expertise. In a large study of customer-support agents, an AI assistant raised productivity most for the least experienced, in effect passing on the habits of the best workers to the newest. A similar effect for junior nurses is plausible but not yet measured.

In availability, retrieval, consistency, writing and paperwork, the machine is already strong and getting cheaper fast.

What it cannot do now, and what it may never do

Here precision matters. “Cannot do now” is different from “can never do”. The second is a much stronger claim.

Hard today, and maybe for a long time:

• Helping a real person reach the right decision. This is the surprise. In a 2025 Oxford study, 1,298 members of the public worked through ten medical scenarios. Tested on its own, the language model identified the relevant condition in 95% of cases. People using the same model identified it in fewer than 35% of cases and chose the right course of action in fewer than 45%. They did no better than a control group using whatever sources they liked. People left out key details; the model misread them; people misread its answers. Knowing the medicine is not the same as getting a frightened person to the right action. That takes the nurse’s skill. This result should temper everything this book has said about the AI nurse. Its abilities were shown in pieces, in separate systems and on tests, not in real conversations with frightened people.
• Being reliably right, and knowing when it is not (Chapters 7, 12, 14 and 16).
• Touch, lifting, physical care. Turning a patient, dressing a wound, helping someone to the toilet. Robots for these tasks lag far behind language models.
• Seeing and examining the patient. Today a text message cannot show anyone the child (Chapter 16); a photo or video call moves that limit. Whether future cameras and robot hands could examine a patient as a nurse does is a forecast, not a finding. Nothing today comes close.

Likely to remain human, even if the technology improves a great deal:

• Being responsible. Someone must answer for the decision: legally, professionally, morally. Machines can advise. Societies decide who is accountable, and so far that is people, through licensing and law.
• Owning the goal. Chapter 15 showed that the machine’s goals are delegated. A good nurse cares whether this child is all right: it is her goal, not an instruction. Patients feel that difference, and it shapes what the nurse notices and pursues.
• Being present. Much of nursing is being there: sitting with the dying, steadying the frightened, advocating for a patient who cannot speak for themselves. A machine can simulate warm words. Whether it can be with someone is the question the coda after Chapter 16 leaves open.

Genuinely uncertain: how far the sense that “the whole picture is wrong” and the tacit “looks wrong” judgment (Chapter 4) can be learned from data, sensors and video. Honest researchers disagree.

What researchers are saying

Researchers disagree, but the debate has a clear shape.

Tasks, not jobs. In 2013, two Oxford researchers estimated that 47% of US jobs were at high risk of computerization. In 2016, an OECD study redid the calculation by looking at the tasks within each job and found about 9%. Most jobs mix automatable and non-automatable tasks. That shift, from asking “will this job go?” to asking “which tasks change?”, is now the standard approach. Nursing is a mixture par excellence.

Exposure is wide, effects so far small. Most US workers have some of their tasks exposed to language models, but exposure means the work could change, not that it has. A large Danish study linking surveys to earnings records found no measurable effect of chatbots on earnings or hours. Workers saved little time, and much of it went into checking the AI’s output.

But a warning signal for the young. Payroll data for millions of US workers show employment for 22-to-25-year-olds in the most AI-exposed occupations falling well behind expectations, by 19% by mid-2026. Older workers in the same jobs were largely unaffected. The data are observational: the authors call this an early descriptive indicator, not an estimate of how much AI caused. The decline is concentrated where AI substitutes for workers’ tasks. Entry-level work, where people learn their trade, may be the first to thin out.

The skeptic. MIT economist Daron Acemoglu estimates that AI will raise overall productivity by at most about two-thirds of a percent over ten years. That is real, but far from a revolution. He and others stress that the direction of technology is a choice: it can be aimed at replacing people or at making them more capable.

The hopeful case, made with nurses. MIT economist David Autor argues that AI’s best use is to extend expertise: letting more people do higher-value work that used to need an elite specialist. His leading example is the nurse practitioner. These are registered nurses with extra training who now perform tasks once reserved for doctors. Their numbers in the US nearly tripled between 2011 and 2022, helped partly by better information tools. On this view, AI could let more nurses do more of what doctors do, and do it well.

The alarmed case. Anthropic’s chief executive has predicted that half of entry-level white-collar jobs could be disrupted within one to five years, because AI is general and closes many escape routes at once. Comparative advantage, the economist’s usual answer, guarantees that some human work remains valuable, but not that it pays well.

And the demand side, which matters most for nursing. The World Health Organization counts about 30 million nurses worldwide. In 2023 it put the shortage at about 5.8 million and projected 4.1 million by 2030, concentrated in the poorest regions. Populations are aging. Few occupations are less likely to run out of work to do. The live question for nursing is less “will there be jobs?” than “what will the jobs be like?”

Who gets it. The shortage is worst where the AI nurse might seem most useful. But it needs power and connections (Chapter 13). It learned most from well-represented languages and hears some voices worse than others (Chapter 7). And its advice is only as good as the local guidance it can retrieve and the services it can send people to (Chapter 14). An instruction to call an ambulance helps only where an ambulance will come. Whether the AI nurse narrows gaps in care or widens them depends on choices about cost, language and local knowledge, and on who is present when those choices are made.

Nurses themselves. In 2024, hundreds of nurses protested outside a major US hospital system over AI tools introduced, they said, without proper testing or nurses’ involvement. Their demand was not “no AI” but “a say in how AI is used”.

The nurse, at the end of the book

So will the AI nurse replace the human nurse?

Present evidence supports assisting with a growing set of tasks, rather than replacing the full nursing role. The machine is likely to take on more first answers, lookups, documentation, monitoring and routine messages, and it may be better than tired humans at some of them. The rest of the job, examining, touching, judging in context and caring, depends on being present, owning the goal and being accountable. Earlier chapters found the machine lacks all three today. That is a reading of present evidence, not a forecast. Robots, sensors and law could all change, and this chapter has shown how much informed people disagree about speed.

That leaves two real risks and one real opportunity. Employers may use the machine to cut nursing where they should not, as in the first hospital. And the entry-level work where nurses learn their craft may thin out. The opportunity is the second hospital: a nurse relieved of the keyboard, supported by a tireless assistant, and free to spend more of her shift doing the parts of nursing that drew her to it.

Technology will not decide which future arrives. People will: hospitals, regulators, professions and nurses themselves. They do not hold equal power. As Chapter 15 showed, the developer, the hospital, the clinicians and the patients all shape the machine’s goals, and they do not weigh equally.

Food for thought

• If a hospital can choose between the two futures, what rules, incentives or votes would push it toward the second?
• Where will tomorrow’s experienced nurses learn their judgment if the machine handles the routine cases they used to learn on?
• If patients prefer the machine’s written answers, is that because the machine is better, or because the humans were never given time to write well?