A recent study explored the use of ChatGPT through clinical vignettes. The AI scored high, while the median physician score was lower. Physicians using AI scored slightly higher, and the AI outperformed both.
The difference was significant.
That’s the headline everyone quoted.
Here’s the part they buried.
Doctors with AI didn’t get better at diagnosing.
They achieved a slightly higher score compared to those without it. Statistically similar. What changed was clock time. They finished each case faster. So the pitch is really: AI makes the system faster. Not doctors. Not patients. The system.
AI algorithms are present in radiology departments across the country.
The American Hospital Association called medical imaging the biggest AI win in diagnostics right now. You walk into a hospital, get a chest X-ray, and something in the pipeline might flag it before any human lays eyes on the pixels. That’s not experimental anymore. That’s just how it works. Most patients never hear about it.
The PR Story vs. What the Data Actually Shows
Let me sit with that study for a second because the public version sounds way better than the data.
The story everyone tells: AI helps doctors. Better outcomes, smarter decisions, the whole augmentation narrative. What the data actually shows: doctors still had to verify everything. Still owned the final call. Just didn’t own it any more accurately than before. They got a speed bump. Useful if you’re running a jammed clinic.
Not obviously useful if you’re the person on table seven hoping someone caught what the last reader missed.
The tech underneath does real things. ML models scan imaging, genomes, EHRs, lab values. Looking for patterns that tired humans miss on hour nine of a shift. Deep learning identifies anomalies in CTs, MRIs, ECG waveforms.
The accuracy gains are real in specific tasks.
Here’s the weird part: an AI that always gets the same thing wrong is harder to catch than a doctor who screws up occasionally. Consistent wrong beats inconsistently right.
Nobody talks about that either.
Stroke Care Is Where This Gets Real
Lab benchmarks are one thing.
Clinical reality is another.
In acute stroke, AI pushes MRI and CT images to the stroke team’s phones instantly. Before the radiologist finishes their report. It predicts large vessel occlusion.
Estimates how much brain tissue is still salvageable.
Minutes saved in stroke assessment can lead to preserved brain tissue.
A system that gets the scan to the right people while the human reader is still in queue isn’t augmentation theater. It’s a different outcome for someone who might not be able to talk by the time the normal workflow catches up.
Radiology has gone furthest down this road. AI-enhanced imaging reduces noise, adjusts contrast on the fly, highlights relevant features. Makes radiologists faster in emergency settings. Same pattern recognition shows up in cancer detection, cardiac imaging, neurological screening.
Here’s the part vendors skip in their slide decks: clinicians don’t just accept AI labels.
They work backwards from the AI suggestion, stress-testing it against patient records, standards of care, other opinions. That’s more cognitive work, not less, when the AI is wrong. Distrusting a bad AI output takes longer than just reading the image yourself. The Nature paper on this is worth a read if you haven’t seen it. Link’s in the sources below.
The Consent Problem Nobody Is Solving
This is where I stop pretending to be neutral.
Most patients don’t know when AI was part of their diagnosis. Not disclosed. Not explained. Not offered as a second opinion. The technology is just there. In the workflow. Without the patient’s knowledge in any meaningful sense.
The odds are not trivial that something flagged your scan before a human saw it. That’s not necessarily bad. But it’s not neutral either.
The ethical frameworks for AI in clinical decision-making are lagging the tech itself. Professional bodies have published guidance. What clinicians owe patients in terms of disclosure? Not settled. Not consistently practiced.
Side note: their documentation about what the AI actually did is usually a mess. Buried in radiology reports nobody reads. Hard to find.
Hard to understand.
There’s a genuine equity angle though. AI can bridge diagnostic gaps in underserved regions. Remote diagnosis via AI-powered wearables in low-resource settings — that’s real. Rare disease diagnosis has improved in some cases since AI matches genomic patterns to known profiles. That used to require years of specialist referrals and repeat testing.
People who would’ve fallen through the cracks are getting answers.
But those same underserved regions and rural hospitals? Last places to get the validated, regulated, well-monitored versions of these tools. The AI algorithms are mostly in well-resourced systems. Promise: there. The deployment reality isn’t.
What You Can Actually Do
You’re not choosing whether AI reads your next scan.
The approvals exist, the systems are deployed, the efficiency incentives aren’t going anywhere.
What you can do: ask.
Got a scan and a diagnosis? Ask if AI was involved in the reading. Ask what the AI flagged and what the human confirmed. Ask for an explicitly human-to-human second opinion if that’s important to you. Most doctors won’t get defensive — especially once they realize you’re not questioning their judgment. Just asking about their workflow.
For healthcare operators and developers building in this space: the accuracy gains are real but deploying a model doesn’t solve the workflow integration problem. Speed without accuracy improvement means you’re building a better triage system, not a better diagnostic system. Those are separate goals with other metrics.
AI in healthcare diagnosis isn’t coming. It’s here. Radiology departments, stroke centers, genomics labs — already running. Whether it makes your care better depends on whether your doctors use it as a tool or a replacement for judgment. Ask the question. It’s your body.
Sources
– Stanford HAI: https://hai.stanford.edu/news/can-ai-improve-medical-diagnostic-accuracy
– Nature (AI in stroke care, clinician adaptation): https://www.nature.com/articles/s41746-025-01460-1
– American Hospital Association (AI in diagnostics): https://www.aha.org/aha-center-health-innovation-market-scan/2023-05-09-how-ai-improving-diagnostics-decision-making-and-care
– PMC (machine learning in diagnostics): https://pmc.ncbi.nlm.nih.gov/articles/PMC9955430/
– PMC (AI diagnostic accuracy): https://pmc.ncbi.nlm.nih.gov/articles/PMC11702416/
– ASLM (AlphaFold, rare disease genomics): https://aslm.org/artificial-intelligence-in-medical-diagnostics-revolutionizing-precision-and-speed-in-healthcare/
– Spectral AI (imaging enhancement): https://www.spectral-ai.com/blog/the-intelligent-revolution-ai-in-medical-imaging-and-diagnostics/
