Artificial intelligence is already involved in medical diagnosis, although you may never be told that it played a part. It can flag a suspicious area on a mammogram, examine a retinal image for signs of diabetic eye disease, identify patterns in an electrocardiogram, or help a clinician organize a complicated list of symptoms.

That sounds as though AI is becoming another diagnostician in the room. Usually, it is not. Some systems perform one narrowly defined task. Others rank possibilities, highlight abnormalities, or provide a second opinion for a clinician. General-purpose chatbots are another category altogether.

The important question is no longer simply, “Can AI detect disease?” It can—and in certain settings, it can do it very well. The more useful question is what happens between an AI finding and a diagnosis that should guide someone’s care.

Medical AI Is Not One Thing

The term medical AI covers tools with very different abilities, evidence, and responsibilities.

A retinal-screening system, for example, may analyze images from adults with diabetes and determine whether the findings warrant referral to an eye-care professional. The system has a defined purpose, accepts a particular kind of image, and looks for a particular level of disease.

Mammography software may work differently. Instead of issuing the final interpretation, it may mark areas that deserve attention or assign a suspicion score while a radiologist reads the examination. Other systems look for possible strokes on brain images, detect abnormal heart rhythms, or help pathologists examine tissue samples.

Then there are large language models—the technology behind general-purpose AI chatbots. These models can summarize medical information and generate plausible diagnostic possibilities from a written case. But fluency is not clinical validation. A chatbot that can discuss hundreds of diseases is not automatically a medical device proven safe for diagnosing any of them.

Putting all these tools under the single label “AI diagnosis” makes it almost impossible to judge what AI is actually doing.

What AI Can See

AI is especially good at finding statistical patterns in large amounts of structured data. An image-analysis system may compare pixels, shapes, textures, and spatial relationships at a level of detail that would be exhausting for a person to reproduce consistently. A prediction system may combine numerous variables and notice relationships that are easy to overlook when information is scattered across a medical record.

This can be useful in several ways:

• Drawing attention to a subtle finding. AI may highlight something that deserves a closer look.
• Providing a consistent second reading. Software does not become tired near the end of a long shift.
• Prioritizing urgent cases. A system may move a scan with a possible stroke or another critical finding higher in the work queue.
• Extending screening to new settings. A narrowly designed autonomous tool may allow screening to occur in a primary-care office where the relevant specialist is not present.
• Connecting information across time. AI may help identify changes across multiple images, laboratory results, or clinical visits.
  
  

These are meaningful capabilities. But recognizing a pattern is not the same as understanding a person.

What AI Can Miss

An AI system sees only what it has been designed and permitted to receive. A model examining a retinal photograph cannot know that the patient has recently developed weakness, is taking a new medication, or misunderstood a question on an intake form. A system analyzing medical-record data may inherit gaps and errors already present in that record.

AI can also perform impressively in the environment where it was developed and less reliably somewhere else. Patient populations differ. So do imaging machines, clinical practices, disease prevalence, data-entry habits, and the definitions used to label the original training cases. A model that learned from one population may not perform equally well for people of different ages, ethnicities, skin tones, or combinations of medical conditions.

This is why “trained on millions of cases” is not enough. More data do not automatically mean better data, and an enormous dataset can reproduce an enormous number of old assumptions.

AI may also be wrong in a way that looks convincing. A highlighted image, precise risk score, or polished paragraph can carry more authority than it deserves. People naturally tend to trust a confident recommendation, particularly when they are busy or uncertain. In medicine, that can create automation bias: a clinician may accept an incorrect AI suggestion or overlook evidence that points elsewhere.

The opposite can happen too. If a system produces too many false alarms, clinicians may begin ignoring it—including the warning that matters.

An AI Finding Is Not Yet a Diagnosis

A diagnosis is assembled from different kinds of evidence. Symptoms, medical history, examination findings, laboratory results, imaging, response to treatment, and clinical judgment may all contribute. AI can strengthen one part of that process without replacing the others.

Imagine that an AI system flags a suspicious area on a mammogram. That finding may prompt closer review, comparison with earlier images, additional imaging, or a biopsy. The AI has not established that the patient has cancer. It has changed what deserves attention next.

The distinction matters in the other direction as well. If AI does not flag an abnormality, that does not prove that disease is absent. No screening method detects every case, and a patient’s symptoms may still justify further investigation.

This is the part often lost when performance is reduced to a headline such as “95% accurate.” Accuracy can combine different kinds of successes and errors into one number. Patients may care much more about two separate questions:

• How often does the tool miss the disease it is supposed to find?
• When it reports a positive result, how often is that result actually correct?
  
  

Those answers depend not only on the tool but also on how common the disease is in the population being tested. Even a strong screening test can produce many false alarms when it is used to search for a rare condition.

FDA Authorization Is Important—but It Is Not a Universal Endorsement

The U.S. Food and Drug Administration maintains a list of AI-enabled medical devices authorized for marketing. Radiology accounts for a large portion of the list, but authorized devices also appear in cardiology, neurology, pathology, ophthalmology, and other specialties.

Authorization does not mean that the FDA has declared “AI” generally safe or superior to clinicians. It applies to a particular product and its intended use. A system authorized to examine one type of image for one finding should not be assumed to diagnose unrelated conditions or work equally well outside the setting for which it was evaluated.

AI also creates an unusual oversight problem: software may be updated, clinical populations may change, and performance can drift after deployment. For that reason, regulators and developers are increasingly concerned with the entire life cycle of a medical AI tool—not only its performance on the day it enters the market.

Where Does Responsibility Belong?

When AI contributes to an error, responsibility can become difficult to locate. Was the problem in the training data, the software, the way the hospital installed it, the information entered, or the clinician’s response? What if the clinician disagreed with the AI but followed it because hospital policy encouraged its use? What if the clinician rejected a correct recommendation?

“Keeping a human in the loop” sounds reassuring, but it does not solve everything. The human needs enough time, information, authority, and relevant expertise to question the system. Otherwise, human review can become little more than a signature placed after an automated recommendation.

Patients also need a way to raise concerns when the result does not fit what they are experiencing. A diagnostic process becomes less safe—not more—if an AI-generated conclusion is treated as a reason to stop listening.

Five Useful Questions When AI Affects Your Care

Patients may not always know when AI has been used, and clinicians may not have access to every detail about a proprietary system. Still, these questions can help clarify what role it played:

1. What information did the AI analyze? Was it reading an image, reviewing laboratory data, listening to speech, or summarizing the medical record?
2. Was it screening, assisting, or making an autonomous assessment? Those are different levels of responsibility.
3. What other evidence supports the result? An AI finding should be considered alongside symptoms, history, examination, and appropriate confirmatory testing.
4. What happens if the AI result and the clinical picture disagree? A mismatch may be a reason to investigate further, not automatically choose one over the other.
5. What is the next step? A useful result should lead to an appropriate action—additional testing, specialist review, monitoring, or reassurance.
   
   

These questions are not an argument against using AI. They are a way to understand what the tool contributed and what still needs human attention.

Are We Ready?

We are ready for some medical AI tools because they have a defined purpose, credible testing, appropriate oversight, and a clear place in patient care. We are not ready to assume that every AI system deserves the same trust—or that strong performance in a study guarantees the same result in every clinic and every patient population.

The greatest promise may not come from asking AI to replace the diagnostician. It may come from giving clinicians another way to notice what could have been missed, organize complicated evidence, and extend selected services where specialists are scarce.

But better pattern recognition does not eliminate the need for context, follow-up, judgment, or a patient who can say, “That answer does not fit what is happening to me.” AI can contribute valuable evidence. The diagnosis still has to make sense in the life and body of the person receiving it.

Note from Marsha

I use AI regularly and appreciate its ability to find connections, organize information, and lead me toward questions I might not have considered. I have also seen how smoothly and confidently it can present an answer that does not hold up when examined more closely.

That combination—real usefulness and convincing error—is why human review matters so much. The better AI becomes at sounding certain, the more important it becomes for us to keep asking whether the answer is actually supported and whether it fits the person in front of us.