AI May Make Your Doctor Better. Will It Make Your Healthcare Better?

Reading Rachael Bedard’s New York Times essay, “Being a Doctor Will Never Be the Same After A.I.,” I found myself thinking about the people on the other side of the examination table.

Bedard, a geriatrician and palliative care physician, describes how AI helps her consider possibilities, find information and challenge her thinking. She also worries about what happens to doctors who begin using these tools before they develop the judgment to evaluate the answers. An experienced physician brings something to that conversation that a resident is still learning.

That is an important discussion. I want my doctors to have both the best tools and the knowledge to use them. I also want the next generation to develop the skills needed to recognize when a convincing answer is wrong.

But what AI does to doctors is one question. What it does to healthcare for the rest of us is another.

When someone in your family needs care, the questions become practical. Can you get an appointment? Can you afford it? Do you understand what the doctor told you? Can you reach someone when the treatment is not working?

An AI system can be impressive and still leave those problems untouched. Depending on how it is deployed and who benefits financially, it could make some of them worse.

We Have Already Met Dr. Google

We have seen an earlier version of this with WebMD and the rise of “Dr. Google.” Long before chatbots, people typed symptoms into search engines, read about possible conditions and brought their findings into medical appointments. Sometimes they decided they did not need an appointment at all.

This was widespread years ago. In its Health Online 2013 research, Pew reported that 35% of American adults had gone online to try to identify a medical condition affecting themselves or someone else.

AI did not invent patients looking things up. Nor did it invent the possibility that they would misunderstand what they found.

What changes is the experience. A search gives you pages to read. A chatbot can ask follow-up questions, incorporate your answers and explain its response in language you understand. You can ask it to try again. You can tell it you are confused. You can keep talking without feeling that you are taking up someone’s time.

That is an appealing experience, particularly for someone who feels rushed, embarrassed or unheard in a medical setting. It can also feel like a consultation, even when the system has not demonstrated that it can safely perform that role.

We should take both parts seriously. Dismissing patients for using these tools does nothing to address why they find them useful.

Patients Could Gain a Great Deal

There is real value in helping people understand their care. Medical terminology can be difficult. Instructions can be confusing. Questions occur to us after we have left the office, when the person who could answer them is no longer available.

AI could help patients organize those questions, understand terminology and prepare for a more productive conversation. It could also help a family member make sense of information while supporting someone through an illness. These are substantial uses even without assigning the system responsibility for diagnosis or treatment.

There is evidence for this more focused approach. A trial published in Nature Medicine, involving 2,069 patients, tested a purpose-built chatbot used before specialist consultations. Researchers reported improvements in communication and consultation experience. The tool was integrated into a process that still included seeing a specialist.

That finding does not establish better long-term health outcomes or lower bills. It does show why the role we give AI matters. Helping someone prepare for care is a specific job whose benefits can be evaluated.

We also need to be honest about the alternative. Some patients already face significant barriers to getting human help. It is easy to compare a chatbot with a knowledgeable physician who is immediately available and has unlimited time. That physician is not the alternative available to everyone.

The opportunity to expand access deserves serious attention. So does the distinction between access to information and access to care.

Having Answers is Only Part of the Job

Bedard brings years of training and experience to her interactions with AI. She can test a suggestion against what she knows, what she observes and what she has learned about a particular patient.

Most of us bring a different set of skills. We may describe what concerns us while leaving out something we do not realize matters. We may accept an explanation because it fits what we hoped to hear. A fluent response can make it difficult to recognize how much remains uncertain.

Physicians are not immune to overconfidence or overreliance on technology. But making medical information available does not automatically give the rest of us the ability to evaluate it.

An Oxford-led study published in Nature Medicine makes that distinction concrete. Researchers tested 1,298 people in the United Kingdom using simulated medical scenarios. Participants assisted by the tested chatbots did no better at identifying relevant conditions and choosing an appropriate course of action than people using their usual resources.

The models performed much better at identifying conditions when they received the scenarios directly. Something was lost in the interaction between the person and the system.

This was a study of particular models and hypothetical cases. It was not a verdict on every medical AI product, and it did not measure actual patient injuries. But it illustrates why medical knowledge inside a model is insufficient evidence that people can use it successfully.

Curiosity and skepticism are useful habits. They cannot be the entire safety strategy. A patient should not have to understand medicine well enough to catch every consequential omission. The service itself needs to handle uncertainty and recognize when human help is needed.

Follow the Cost of Care

I suspect the economics will do a great deal to determine how this develops.

For a person struggling to afford healthcare, the appeal of an inexpensive answer is obvious. For an organization paying to deliver care, the appeal of automation is equally obvious. Those interests may align, but we should examine whether they actually do.

West Health-Gallup research published in April found that, among recent users of AI for health information, 14% said they used it because they could not afford a doctor visit. Sixteen percent said they could not access a provider.

In a separate question, 14% of recent users reported skipping a provider visit after receiving AI information or advice. The survey does not establish whether those visits were unnecessary or whether people missed care they needed.

It does establish that this discussion extends beyond future possibilities. People are already incorporating AI into decisions about whether to seek care, including when money and access are obstacles.

We need to distinguish three different measures: What it costs to provide care, what the patient or purchaser pays, and what the complete course of care costs over time.

An organization could reduce its operating expenses without reducing anyone’s bill. A less expensive encounter could be a genuine improvement if it addresses the problem effectively. It could also produce additional expense if a patient has to come back repeatedly or a condition goes unrecognized. Those outcomes require measurement.

The same applies to time. If AI reduces administrative work, an organization could use that capacity to improve access or give clinicians more time with patients. It could also increase the number of patients each clinician is expected to handle. The technology does not make that management decision.

And AI’s financial role extends well beyond the examination room. A September 24 KFF Health News briefing highlighted Times reporting on a dispute over hospitals’ use of AI in billing. The Blue Cross Blue Shield Association claimed that more complex coding was increasing expenses without evidence of different treatment.

That is an insurer association’s claim in a payment dispute, not an independent finding of fraud or proof that AI caused every increase. Still, it identifies a use of the technology that patients should pay attention to. Systems can be deployed to improve care, strengthen reimbursement claims or scrutinize payment requests. Success for the organization deploying them does not automatically establish a benefit for the patient.

A promise of efficiency needs a second sentence explaining who receives the savings and what happens to the quality of care.

Who Gets a Doctor, and Who Gets a Chatbot?

One possible outcome concerns me: People with resources get clinicians equipped with increasingly capable AI, while people without resources get AI as their principal available source of help.

That is a risk to examine, not a claim that every healthcare organization is building such a system. There will be tasks that technology can handle effectively, and many patients may prefer that convenience. We should welcome uses that earn their place through evidence.

But choosing a digital service because it works well is different from accepting one because the human alternative is unaffordable or unavailable. A system that expands our choices is different from one that quietly narrows them according to what we can pay.

For any delegated task, we need to know whether the tool works for the people expected to use it. We also need a practical route to qualified human care when the situation exceeds its capabilities. Telling someone to consult a physician means little if the service leaves them unable to reach one.

This is where Bedard’s concern about training returns to the patient’s experience. If human judgment remains part of the plan for handling difficult cases, developing that judgment must remain part of the investment. It cannot simply be assumed to exist whenever an automated service needs backup.

I want doctors to benefit from AI. I want patients to benefit from it, too. There is no reason those ambitions cannot advance together, but we should judge the results separately.

For the rest of us, the evidence will be in appointments we can obtain, bills we can afford, explanations we understand and care that produces better outcomes. It will also be in whether someone qualified is reachable and accountable when things go wrong.

Making healthcare less expensive to deliver would be an achievement. Making good healthcare easier to obtain and afford is the achievement patients need.