The 2 a.m. message

It is two in the morning. A parent speaks or types into a phone:

“My daughter is 3. She’s had a fever all evening, 39.5. Now she has little purple spots on her legs, and they don’t go away when I press a glass on them. Should I wait until morning?”

A good nurse knows at once this cannot wait. A fever together with a rash that does not fade under pressure can be a sign of meningitis or sepsis, so the answer is: call an ambulance now. She knows how to say it, calmly and clearly, to a frightened person. She knows what she does not know: she cannot see the child, so she errs on the side of caution. Then she acts. She checks the child’s record, alerts the doctor on call, and phones back in the morning.

The answers, however, are not coming from a human this time. They are coming from a machine. Today, a parent may not realize they are talking to a machine. The machine can recognize the same warning signs in many phrasings. It can explain why the situation is urgent, ask a sensible follow-up question and use tools to take the next steps.

Now open the machine up with me. The machine holds only a clock and billions of switches, each either on or off. Every tick, each switch changes according to what a few of its neighbors are doing. That is all. No single switch knows anything about children or fevers.

So here is the question I want to answer:

How can switches following simple rules produce a voice so knowledgeable, reassuring, and helpful that it can do much of a nurse’s job?

One rung at a time

The answer is not a single big idea but a ladder of concepts. We will climb it together, one rung at a time. At each rung I add one ingredient and show what the machine could not do without it. Then I go back to the parent. How much of the nurse’s work can the machine, which I call the AI nurse, now do? The AI nurse combines abilities that have been shown in different systems; it is a way to explain the technology, not a tested clinical product. The real machines named at the end of each chapter show one ability at a time.

This ladder is not magic. It is science and engineering, and nothing on it is free: in Step 13 I stop to count what each rung has cost. Nor does climbing remove every limit. Some limits are mathematical and bound every computation, yours and mine included, as far as anyone knows; others depend on the evidence and resources available. Chapter 3 shows the first of them, and Chapter 16 looks at all of them from the top.

The climb has two stages. Steps 1 to 12 assemble the machine, and Step 13 counts its cost; these explain mechanisms that are well established. Steps 14 to 16 add nothing new but ask what has been built: what its words mean, whose goals drive it, and where it fails. These chapters examine meaning, goals, reliability and limits. Some of their questions are still disputed, and there I offer arguments, not settled results; others, such as how well a machine knows when it is wrong, can be tested. Steps 17 and 18 turn to the people who work with the machine: first the team on a hospital ward, then the jobs themselves.