A radiologist in rural Ohio is reviewing a chest X-ray flagged by AI before she even opens the image. The AI has already spotted a 6mm nodule in the upper right lobe, ranked it by malignancy probability, and pulled three similar historical cases for comparison. She still makes the call — but she’s working faster, missing less, and spending more time with patients who need her. That’s not a concept paper. That’s what Google’s Med-PaLM 2 and systems like Nuance DAX Copilot are enabling in clinical workflows right now. The question isn’t whether AI is entering healthcare — it’s already through the door. The question is what’s actually changing, for whom, and how fast.
Diagnostics: Where AI Is Already Earning Its Stripes
The clearest, most documented wins for AI in healthcare are in medical imaging and pattern recognition. This isn’t surprising — these are exactly the kinds of high-volume, structured-data problems where large models have a proven edge.
Google’s DeepMind published results showing its AI matched or outperformed ophthalmologists in detecting over 50 eye diseases from retinal scans. Viz.ai has FDA clearance for stroke detection and is deployed in over 1,400 hospitals — it analyzes CT scans and pages the neurology team within minutes of a large vessel occlusion being detected. That speed matters: in stroke care, every minute of delay costs roughly 2 million neurons.
PathAI is doing similar work in pathology, using AI to analyze tissue slides for cancer markers with accuracy that’s pushing into specialist territory. And Aidoc has built an AI backbone that runs across radiology workflows — not replacing radiologists, but acting as a second set of eyes that never gets tired at the end of a night shift.
What’s important to note here is that these systems are narrow. They’re trained on specific tasks — detecting a particular anomaly, flagging a particular pattern — and they don’t generalize the way a human clinician does. A radiologist who’s been awake for 14 hours is still better at recognizing something they’ve never seen before. But for the high-volume, well-defined tasks that make up the majority of diagnostic work? AI is legitimately useful right now.
Clinical Documentation: The Unglamorous Win That Might Matter Most
If you ask doctors what’s burning them out, the answer isn’t usually the medicine — it’s the paperwork. The average physician spends nearly two hours on administrative tasks for every hour of direct patient care. That’s a system-design failure, and AI is quietly starting to fix it.
Nuance DAX Copilot (owned by Microsoft) records ambient conversations between doctors and patients, then automatically generates clinical notes in the physician’s style. It integrates directly with Epic, the dominant electronic health record system used by most major US health systems. Doctors using it report saving 3-5 minutes per patient encounter — which sounds small until you multiply it across 20 patients a day, 250 days a year.
Suki AI is doing similar work, targeting independent practices and smaller health systems that can’t afford Microsoft’s enterprise pricing. Abridge, backed by UPMC and others, is specifically focused on making notes more readable for patients — turning clinical jargon into plain-language summaries that get sent to the patient’s portal after the visit.
Demis Hassabis at DeepMind has talked about AI’s potential to handle the “cognitive load” of medicine — not to replace clinical judgment, but to free up the mental bandwidth that’s currently consumed by documentation and administrative overhead. The ambient documentation tools are the first real proof of that thesis at scale.
Drug Discovery: Long Time Horizons, Real Early Signals
This is the area where the potential is massive and the timelines are long. Be skeptical of any headline promising that AI has “solved” drug discovery — that’s not what’s happening. But there are genuine signals worth paying attention to.
AlphaFold 3, released by DeepMind in 2024, predicts protein structures and interactions with accuracy that was unimaginable five years ago. The original AlphaFold essentially solved a 50-year-old grand challenge in biology. AlphaFold 3 extends that to model how proteins interact with DNA, RNA, and small molecules — the kind of data that’s foundational to understanding how drugs work at a molecular level. DeepMind made the model weights available to researchers, which has accelerated work across dozens of labs worldwide.
Isomorphic Labs (a DeepMind spinout) is using this infrastructure to partner with Eli Lilly and Novartis on specific drug targets. These are real partnerships with real capital behind them, not academic exercises. But it’s worth being honest: we’re still years away from knowing whether AI-accelerated drug discovery translates into more drugs reaching patients faster. The bottleneck in pharma isn’t always target identification — it’s clinical trials, regulatory approval, and manufacturing. AI helps with the first part of the pipeline, not the whole thing.
Recursion Pharmaceuticals is taking a different approach — using AI to mine massive datasets of cellular images to find drug candidates, essentially treating biology as a search problem. They’ve partnered with NVIDIA and have a real pipeline, though most candidates are still in early-stage trials. Thinkers like Peter Diamandis have pointed to exactly this kind of AI-driven biological research as one of the most consequential near-term applications of the technology.
What AI Still Can’t Do in Medicine (And Probably Won’t Soon)
It would be dishonest to write about AI in healthcare without being clear about the limits. This is an area where overconfidence is genuinely dangerous.
- Clinical reasoning at the margin: Large language models like GPT-4 can pass the USMLE (medical licensing exam) and provide reasonable answers to many clinical questions. But medicine at the edges — rare presentations, atypical symptoms, patients with five comorbidities and complex social histories — still requires contextual judgment that current AI doesn’t reliably deliver. This gap is part of why the question of what it would actually take to reach AGI matters so much in a medical context.
- Real-time physical examination: AI can analyze data, but it can’t palpate an abdomen, observe how a patient walks into the room, or pick up on the non-verbal cues that often tell a clinician more than the labs do. Telemedicine is changing this partly, but the physical examination is still irreplaceable in many contexts.
- Liability and accountability: When an AI system contributes to a missed diagnosis, who is responsible? This is still legally murky in most jurisdictions. Clinicians often feel pressure to override AI recommendations not because the AI is wrong, but because the liability framework still puts the human in the seat of final accountability.
- Bias in training data: Most major medical AI systems were trained predominantly on data from large academic medical centers in wealthy countries. This creates real performance gaps when deployed in diverse populations. Yann LeCun and others have raised broader concerns about how AI systems encode the biases of their training data — in healthcare, those biases can translate directly into worse outcomes for underserved groups.