Why AI Isn’t Finding Cures (Yet)

If you’ve seen headlines lately, you’ve probably noticed a very specific kind of hype: AI is curing disease, AI is revolutionizing cancer care, AI found a life-saving drug, and so on.

It’s exciting. It’s dramatic. It’s also a little misleading.

The better question isn’t really “Is AI finding cures?” It’s more like:

Is AI helping humans discover treatments faster, especially for diseases we’ve struggled to treat for years?

And the answer to that is increasingly: yes.

Not in a sci-fi “robot doctor invents miracle pill overnight” way. More in an “AI can scan gigantic amounts of biomedical knowledge, spot overlooked connections, and help researchers test smarter ideas” way. Which, honestly, is still pretty incredible.

Let’s talk about what’s actually happening, where AI is already making a difference, and why one of the most promising frontiers isn’t always inventing brand-new drugs but finding new uses for the ones we already have.

First, let’s calm down the word “cure”

The word cure does a lot of heavy lifting in health headlines.

In medicine, a cure is rare, specific, and hard-won. Many diseases are not solved by a single breakthrough. Cancer alone isn’t one disease; it’s hundreds of diseases with different causes, mutations, behaviors, and responses to treatment. Rare immune disorders are similarly complicated. Alzheimer’s, autoimmune disease, infections, and genetic conditions all come with their own scientific mazes.

So when people say AI is “finding cures,” what they often mean is one of these:

• AI is helping identify a promising treatment candidate
• AI is improving diagnosis or early detection
• AI is helping repurpose an existing drug
• AI is helping personalize treatment
• AI is speeding up research

That may sound less dramatic than “AI cured disease,” but it’s actually more useful and more accurate.

The most practical AI win right now: drug repurposing

One of the biggest bottlenecks in medicine is drug development. Creating a new drug from scratch can take over a decade and cost enormous sums of money. It’s risky, slow, and failure is common.

That’s why drug repurposing is such a big deal.

Drug repurposing means taking a medicine that already exists, one that’s already been approved or at least studied, and asking: could this also work for a different disease?

This is where AI shines.

Modern biomedical AI systems can sift through huge “knowledge graphs” made up of connections among diseases, genes, proteins, pathways, symptoms, and drugs. Humans can study these relationships too, of course, but not at the scale of millions of possible drug-disease combinations.

A great recent example comes from Penn Medicine. Researchers used a machine-learning approach to analyze roughly 4,000 existing medications and identified adalimumab, a drug already approved for conditions like arthritis and Crohn’s disease, as the top predicted treatment for idiopathic multicentric Castleman’s disease (iMCD), a rare and potentially deadly disorder. The patient in the report had been nearing hospice care and later entered remission after treatment with the AI-identified drug. That is a remarkable story, and it highlights the real promise here: AI didn’t magically invent a new molecule. It helped uncover a life-saving option hidden in plain sight. Penn Medicine

And that may be the most underappreciated theme in this whole field: some answers may already exist. We just haven’t connected the dots yet.

Why rare diseases are such an important test case

AI may be especially useful in rare disease research, and that makes a lot of sense.

Rare diseases have a few major problems:

• They affect relatively small numbers of people
• Many have little funding
• There may be few or no approved treatments
• Traditional drug development often isn’t commercially attractive

That creates a tragic gap: patients can be desperately ill, but the economics of the system don’t naturally prioritize them.

AI can help by lowering the cost of hypothesis generation. Instead of starting from zero, researchers can use models to rank existing drugs that might plausibly work, then investigate the best candidates. This doesn’t eliminate the need for lab work, trials, or physician judgment, but it gives science a much better starting point.

David Fajgenbaum and the nonprofit Every Cure have become central figures in this space. Their goal is to use AI to scan the known universe of diseases, drugs, genes, and proteins to find promising matches for untreated or undertreated conditions. As The New Yorker described, this kind of system can generate ranked predictions for drug-disease pairs, especially for conditions where no one has had the resources or incentive to test old medicines in new ways. The New Yorker

That is not flashy in the Hollywood sense. It’s better. It’s practical.

AI and cancer: not one miracle, but lots of useful upgrades

Cancer is probably the area where public imagination runs wildest. People hear “AI and cancer” and immediately think: Did it finally crack it?

Not exactly. But it is changing the field in meaningful ways.

According to the Cancer Research Institute, AI is helping across the cancer pipeline: prevention, early detection, diagnosis, treatment planning, and drug discovery. Cancer Research Institute

That broad influence matters because cancer outcomes often depend on many small advantages adding up:

• catching disease earlier
• interpreting scans more accurately
• predicting who is high-risk
• matching therapies more precisely
• reducing time lost in the diagnostic process
• identifying biomarkers that humans might miss

For example, Harvard researchers reported AI models that used patterns from millions of patient records to predict future pancreatic cancer risk, important because pancreatic cancer is notoriously hard to catch early. Harvard Medical School

That kind of work is hugely promising because early detection is often the difference between a manageable disease and a devastating one.

And on the research side, AI tools are increasingly useful for making sense of genomic data, selecting drug targets, and narrowing down which experiments are worth doing next. Not glamorous, maybe, but in science, shaving years off a process is a very big deal.

The less glamorous truth: AI is often better at finding patterns than proving answers

Here’s the important reality check.

AI is very good at:

• sorting huge datasets
• recognizing patterns
• identifying correlations
• generating ranked predictions
• suggesting hypotheses humans should test

AI is not, on its own, proof.

This distinction matters a lot in medicine. A model can suggest that Drug X might work for Disease Y. That does not mean doctors should immediately prescribe it everywhere. The idea still needs biological validation, safety review, clinical evidence, and ideally controlled trials.

In the Penn Medicine case, the AI prediction wasn’t used in isolation. Researchers also did laboratory work and found evidence that TNF signaling appeared elevated in severe iMCD, which supported the idea that a TNF-blocking drug like adalimumab might help. That combination, AI plus biology plus clinical judgment, is where the real power is. Penn Medicine

Sometimes the biggest breakthrough is simply asking the right question sooner

There’s a quiet beauty to this whole movement.

Medicine already contains an overwhelming amount of scattered knowledge: journal articles, trial data, protein interactions, clinical observations, patient records, pathway databases, safety reports. No human being can hold all of it in their head. No team can read everything fast enough.

That means some treatments may remain invisible not because they are impossible, but because modern science is fragmented. The signal is there. It’s just buried.

AI can help recover those buried possibilities.

That’s why stories like Fajgenbaum’s resonate so strongly. He survived Castleman disease in part because a repurposed drug, sirolimus, turned out to target a pathway involved in his illness. Later, AI-assisted efforts helped identify another repurposing opportunity for a different patient with iMCD. This is less about machine omniscience and more about computational memory at an absurd scale. The New Yorker

To put it casually: AI is becoming very good at saying, “Hey, this weird connection over here looks worth a closer look.”

And sometimes that is exactly what saves a life.

But yes, there are real risks and reasons for caution

Now for the part every responsible AI-in-health article needs to include: this can go wrong.

1. False hope

When patients hear “AI found a treatment,” it can sound far more certain than it is. Predictions are not guarantees.

2. Bad data in, bad outputs out

Medical AI models are only as good as the data they learn from. If the data is biased, incomplete, noisy, or unrepresentative, the recommendations can be flawed too.

3. Off-label doesn’t mean harmless

A drug already being approved for one condition does not automatically make it safe or effective for another population, dose, or disease context.

4. Transparency problems

Some AI systems are hard to interpret. If a model gives a recommendation but no one can explain why, clinicians may rightly hesitate.

5. Data privacy

Health data is among the most sensitive data we have. AI systems in medicine need rigorous governance, not a “move fast and break things” mindset.

6. Equity

If AI tools are trained mostly on data from specific populations or wealthy health systems, they may underperform for everyone else.

These aren’t side issues. They are central issues.

Medicine can’t afford technological arrogance. A wrong movie recommendation is annoying. A wrong treatment recommendation is dangerous.

So, is AI actually changing medicine already?

Yes, just not always in the headline-friendly way people expect.

The most realistic picture looks like this:

• AI helps identify high-risk patients earlier
• AI helps radiologists and pathologists spot subtle signals
• AI helps researchers prioritize which molecules and mechanisms to study
• AI helps repurpose existing drugs
• AI helps clinicians navigate expanding medical knowledge
• AI may reduce the time between insight and intervention

Those are not tiny improvements. In many diseases, they can be the difference between too late and just in time.

And sometimes “just in time” is everything.

The future probably belongs to hybrid intelligence

If there’s one theme emerging from all of this, it’s that the winning model is probably not AI alone or humans alone.

It’s humans with AI.

Researchers using AI to narrow the field. Doctors using AI-assisted tools but applying experience and skepticism. Biologists testing AI-generated ideas in the lab. Clinical teams validating what seems promising and rejecting what doesn’t hold up.

That’s a much more believable future than the simplistic either/or framing.

The best AI systems won’t replace medical expertise. They’ll function like an always-on, superhumanly well-read research assistant that never gets tired and has somehow skimmed millions of papers before breakfast.

You still need the doctor. You still need the scientist. You still need evidence.

But now those people may get better clues, faster.

Final thought: AI may not “find cures” all by itself—but it may help us stop overlooking them

That, to me, is the most compelling part of this story.

For years, medicine has had an information problem disguised as a discovery problem. Some solutions may already be scattered across journals, databases, and forgotten experiments. Some drugs may have second or third lives no one has systematically pursued. Some diseases may be more understandable than they appear once enough data is brought into one frame.

AI is not magic. It does not eliminate uncertainty. It does not repeal biology. And it absolutely does not guarantee a cure.

But it may help us ask better questions, connect neglected evidence, and make smarter bets.

And in medicine, smarter bets save lives.

That’s not a miracle.

It’s something more useful: progress.