AI in Healthcare: What's Actually Happening

Healthcare is one of those areas where AI hype meets genuine life-saving potential. Every week there's a new headline about AI "revolutionizing" medicine, "cracking" disease diagnosis, or "transforming" drug discovery.

I've been curious about what's actually real versus what's still science fiction. After digging into studies, talking to people in the field, and tracking actual deployments, here's an honest assessment.

Where AI Is Genuinely Working

The clearest wins in healthcare AI are in areas where pattern recognition meets high-volume repetitive tasks. Not coincidentally, these match the general pattern of where AI excels across industries.

Medical Imaging: The Star Use Case

AI systems have gotten remarkably good at spotting patterns in X-rays, CT scans, MRIs, and other medical images. This is probably the most mature healthcare AI application.

What's actually deployed:

The crucial detail: These systems don't replace doctors. They act as a "second pair of eyes," flagging potential issues for human review. A radiologist still makes the final call.

Why this works well:

• Medical images are standardized (unlike, say, patient histories)
• There's clear ground truth (either cancer or not cancer)
• The task is pattern-matching, where AI excels
• Doctors can easily verify or override AI suggestions

Workflow and Administrative Tasks

Less glamorous but highly impactful: AI reducing paperwork burden on healthcare workers.

Examples in use:

• Automated transcription: AI converts doctor-patient conversations to structured notes
• Prior authorization: AI predicts which treatments will be approved, streamlining paperwork
• Scheduling optimization: AI helps manage appointment booking and resource allocation
• Clinical documentation: AI drafts notes from voice recordings that doctors review and approve

Healthcare workers spend an enormous amount of time on documentation. Tools that reduce this burden—even by 20-30%—translate to more time with patients.

Patient Monitoring and Early Warning

AI systems that continuously monitor patients and flag deterioration before it becomes critical:

• ICU monitoring: AI analyzes vital signs to predict complications hours before they become obvious
• Sepsis detection: Early warning systems that catch blood infections before they become life-threatening
• Post-operative monitoring: AI tracks recovery patterns and flags patients at risk of complications

These systems don't make diagnoses—they raise alarms. They're most valuable in high-volume settings where human attention is stretched thin.

Drug Discovery: Promising but Slower Than Headlines Suggest

You've probably seen headlines about AI "revolutionizing" drug discovery. Companies like Insilico Medicine, Recursion, and Exscientia have raised billions on this promise.

Here's the honest reality: AI is genuinely useful in early-stage drug discovery, but doesn't magically speed up the overall process.

What AI Actually Does in Drug Discovery

Target identification: AI helps identify which proteins or pathways might be involved in a disease. Previously, this took years of experimentation. AI can suggest candidates much faster.

Molecule generation: AI can design new molecules predicted to bind to specific targets. This accelerates the "throw things at the wall" phase of finding drug candidates.

Property prediction: AI predicts whether a molecule will be toxic, absorbable, stable, etc. This helps filter candidates before expensive lab testing.

Why It Doesn't Speed Everything Up

The challenge is that drug development has multiple phases:

AI can compress the first box. But clinical trials take as long as they take—you can't rush biological responses in human bodies. And regulators aren't going to relax safety requirements because a computer said the drug would work.

Bottom line: AI might shave 1-2 years off a 10-15 year process. Valuable, but not the revolution some headlines suggest.

Mental Health: Careful Optimism

AI-powered mental health tools are proliferating, from chatbots like Woebot to diagnostic assistance tools. The potential is significant—mental health has massive unmet demand and long wait times.

What's working:

• Chatbots for guided cognitive behavioral therapy exercises
• Mood tracking apps that identify patterns
• Tools helping therapists between sessions

What's concerning:

• AI can't handle crisis situations safely
• Risk of people substituting AI for needed human care
• Limited evidence for long-term effectiveness
• Privacy concerns with intimate personal data

This is an area where I'm genuinely uncertain whether the benefits outweigh the risks. Proceed with appropriate caution.

The Stuff That Worries Me

For all the promise, there are real concerns worth understanding:

Bias and Equity

AI trained on historical medical data can inherit historical biases:

• Dermatology AI often performs worse on darker skin (training data skewed white)
• Heart disease prediction models trained mostly on male patients
• Socioeconomic factors embedded in training data

If an AI was trained on data from well-resourced hospitals, it might not work well in under-resourced settings. This could widen, not narrow, healthcare disparities.

The Black Box Problem

Many powerful AI systems can't explain their reasoning. A radiologist looking at a scan can say "I see an irregular border here that concerns me." An AI just outputs a probability score.

This makes it hard to:

• Know when to trust AI recommendations
• Learn from AI insights
• Satisfy regulatory requirements for explanation
• Debug errors when they occur

There's active research on "explainable AI," but it remains challenging for the most powerful models.

Data Quality and Privacy

Medical AI is only as good as its training data. Problems include:

• Incomplete or inconsistent records
• Data from one population may not generalize
• Obvious labels (like diagnosis codes) may not capture clinical reality
• Privacy regulations limit what data can be collected and used

And of course, medical data is among the most sensitive. The privacy implications of aggregating it for AI training are significant.

Regulatory Lag

Healthcare is (rightly) heavily regulated. But AI doesn't fit neatly into existing frameworks:

• Is a diagnostic AI a medical device?
• Who's liable when AI makes a wrong recommendation?
• How do you validate a system that changes over time?

Regulators are working on this, but there's inevitable lag between technology capabilities and appropriate oversight.

What's Coming Next

Areas to watch over the next few years:

Personalized treatment: AI analyzing genetic, imaging, and clinical data to recommend individualized treatments.

Preventive health: AI identifying people at risk before they get sick, enabling early intervention.

Clinical decision support: More sophisticated tools that help doctors navigate treatment options.

Surgical assistance: AI guiding surgical robots or providing real-time feedback during procedures.

Administrative automation: More paperwork handled by AI, freeing clinical time.

The theme: AI augmenting human healthcare workers, not replacing them.

The Takeaway

AI in healthcare is real and genuinely helping, but mostly in supporting roles. The dream of AI as a doctor—making diagnoses and treatment decisions autonomously—is still very far away and may never be fully appropriate.

For now, think of healthcare AI as better tools for the humans doing the work. Faster image analysis. Less paperwork. Earlier warnings. Smarter research.

The revolution is more evolutionary than headlines suggest, but that doesn't make it less valuable. Incremental improvements in a system that affects everyone's health add up to significant impact.

Related reading:

• MedGemma: Google's New Medical AI
• AI Ethics and Governance
• Healthcare AI Tools