AI in Healthcare in 2026: What Is Real in Diagnosis, Drug Discovery, and Clinical Care

Updated September 2026: rewritten around new trial results, including the largest randomised trial of AI in breast cancer screening.

AI in healthcare gets two kinds of headlines. One says AI will replace doctors.

The other says it’s all hype.

Neither is right. The honest picture is narrower, and more interesting.

AI is already doing real work in medicine, mostly in focused tasks like reading scans. Elsewhere, the promise is still ahead of the evidence.

Here’s what’s real now, what’s early, and how to tell the difference.

How is AI used in healthcare today?

Most medical AI today is narrow. It does one job, on one kind of data, inside a clear workflow.

The pattern is simple. AI works best when the task is narrow, the data is clean and a human checks the result.

Screening: where the strongest evidence is

The biggest trial so far

The clearest evidence comes from breast cancer screening in Sweden. The MASAI trial randomly assigned 105,934 women to one of two kinds of screening.

In standard screening, two radiologists read every mammogram. In the AI group, software sorted scans by risk.

Low-risk scans got one reader, and high-risk scans got two, with AI marks to help.

What it found

That means about 29% more cancers found, with 44% fewer readings. False alarms didn’t rise significantly.

Crucially, cancers that appeared between screens, known as interval cancers, were not more common with AI. That was the key safety question, reported in The Lancet in 2026.

Why this matters

Many countries are short of radiologists. A tool that saves reading time without missing more cancers could make screening programmes more sustainable.

The AI didn’t work alone, though. At least one radiologist still read every scan, and humans made the final calls.

It’s worth being clear about what the trial shows, though. It found more small, early cancers.

Whether that saves lives needs longer follow-up.

Getting more people tested

Screening only works if people actually get screened. That’s where one of the most practical uses of AI shows up.

Children and young adults with diabetes need regular eye checks for diabetic eye disease. Many never go, because it usually means a separate appointment.

A randomised trial at Johns Hopkins, published in 2024, tried an AI eye exam during the usual diabetes visit. The result came back on the spot.

The AI was autonomous. It gave a result without a doctor needing to read the images first.

Among those with an abnormal AI result, 64% then saw an eye specialist. In the referral group, only 22% completed the follow-through.

One caveat: an author had financial ties to the company that makes the AI system, which the paper disclosed.

Regulated tools are multiplying

AI in medicine isn’t only experimental. In the US, the Food and Drug Administration keeps a public list of AI-enabled medical devices authorised for sale.

Many are imaging tools. The FDA notes the list isn’t comprehensive, so more AI is in use than it shows.

Authorisation isn’t the same as proof that patients do better. For most tools, trials like MASAI simply don’t exist yet.

The bias problem

AI learns from data. If that data under-represents some groups, a tool may work less well for them.

The World Health Organization’s guidance on AI models in health lists bias among its main concerns, alongside accuracy and accountability.

Even strong trials can leave gaps here. The first MASAI results noted that race and ethnicity data weren’t collected.

That matters before a tool spreads to more diverse populations. A screening tool that works well on average can still miss more in some groups.

The fix is not to avoid AI. It’s to test it on the people it will actually serve, and to keep checking once it’s in use.

Drug discovery: better maps, not instant medicines

Drug discovery is where AI’s long-term promise looks biggest. It’s also where the hype runs furthest ahead.

The best example is protein structure prediction. The AlphaFold Protein Structure Database now offers open access to over 260 million predicted structures.

Knowing a protein’s shape helps scientists find drug targets and design molecules. That speeds up the early, exploratory stage.

But the slow parts of medicine remain slow. Lab testing, safety studies, manufacturing and clinical trials still take years.

AI improves the starting map, but it doesn’t skip the journey.

Chatbots and generative AI: useful, but supervised

Large language models, the technology behind chatbots, are spreading fast in healthcare.

They’re being used to draft clinical notes, summarise records and write patient letters. Those are real time savers for busy clinicians.

The risk is that they can sound confident while being wrong. The World Health Organization has published guidance on these models in health, warning about accuracy, bias and accountability.

My rule of thumb: use a chatbot to understand a diagnosis or prepare questions. Don’t use it to replace one.

Ask it to explain terms and options, then take those questions to someone who can examine you.

Surgery: better tools, not robot surgeons

AI in surgery mostly means better support for the surgeon. Think clearer imaging, instrument tracking and guidance during robotic procedures.

Fully autonomous surgery in everyday hospitals isn’t here. The surgeon remains in charge, and accountable.

That may change one day, but only after careful trials.

That’s the right order of events. Surgery is a poor place to test full autonomy before safety and responsibility are well understood.

What earlier research found

The first MASAI results

The first MASAI results, published in The Lancet Oncology in 2023, covered 80,033 women.

AI-supported screening then found cancers at a similar rate to standard reading, with 44% less reading work. It was judged safe enough to continue, setting up the later results.

Detection gains confirmed

A 2025 analysis in The Lancet Digital Health covered the full group of 105,934 women.

It confirmed the 29% higher detection rate, mainly of small invasive cancers and high-grade early cancers.

How it fits together

The screening evidence has grown step by step: first safety, then detection, then interval cancers. That’s how medical AI should earn trust.

For most other uses of AI in healthcare, that chain of evidence doesn’t exist yet.

How much should you trust this?

Promising. Screening AI now has strong trial evidence. Most other uses are earlier.

What makes it convincing

• MASAI is a large randomised trial, with results published step by step in top journals.
• The ACCESS eye trial was also randomised, with a very large difference between groups.
• The FDA list shows AI tools passing regulatory review, not just lab demos.

What makes me cautious

• MASAI took place in one country’s screening programme, so results may differ elsewhere.
• Finding more cancers doesn’t automatically mean fewer deaths.
• The eye trial was small, at one centre, and an author had ties to the maker.
• Many approved AI tools lack trials showing better patient outcomes.
• Evidence for chatbots in clinical care is still thin.

What this means for you

AI is increasingly part of your care, often without you noticing. A few questions help you stay in charge.

• Ask whether AI was used. For scans and screening, it’s reasonable to ask how results were checked.
• Keep up with screening. If AI makes screening quicker or easier, that’s a genuine benefit.
• Use chatbots to prepare, not to diagnose. They’re good for questions to ask your doctor.
• Check the source of health apps. Look for regulatory approval and published evidence.
• Remember a human is accountable. Your clinician is still responsible for decisions about your care.

In this AP report, Swedish researchers describe the results of AI-assisted breast cancer screening:

What we still don’t know

1. Does AI screening save lives? Longer follow-up of trials like MASAI is needed.
2. Does it work everywhere? Results from one country’s programme may not transfer.
3. How safe are chatbots in real clinics? Good trials are still scarce.
4. Will AI speed up approved medicines? Faster discovery hasn’t yet clearly shortened drug development.
5. Who is responsible when AI errs? Rules on liability are still developing.

My take: trust the trials, not the demos

What I like about the screening story is its patience. MASAI tested safety first, then detection, then the harder question of missed cancers.

I’m less convinced by sweeping claims about AI transforming all of medicine. Most tools haven’t faced that kind of test.

So my advice is simple. When you hear about a new medical AI, ask one question: was it tested in a trial on real patients?

If the answer is yes, it’s worth your attention. If not, it’s still a promise.