·6:29

OpenAI Just Signed 8 Major Hospital Systems — Jul 10, 2026

Show notes

OpenAI just signed eight hospital systems the same week two studies said not so fast.

Run time: 6:29

In today's episode:

  1. OpenAI signs eight major hospital systems
  2. NHS deploys AI scribes to every A&E
  3. Nature Medicine finds top medical models brittle
  4. Kenya trial: AI helped notes, not patients
  5. Weill Cornell agents design clinical trials
  6. Cardiac digital twin flags non-responders before implant
  7. Philips gets FDA nod for AI ultrasound
  8. Medicare launches new health-tech and AI office
  9. Claude Cowork expands to web and mobile

TL;DR:

  • OpenAI is buying its way into hospitals — eight health systems signed, 230M weekly health users — but it's distribution, not clinical evidence.
  • Two peer-reviewed papers this week undercut the hype: Nature Medicine shows flagship models are brittle under simple adversarial tweaks, and a Kenyan RCT shows an AI helper improved notes but not patient outcomes.
  • The at-scale wins remain operational, not diagnostic: NHS ambient scribes save ~47 min/shift; Philips' new FDA-cleared ultrasound automates measurement, not diagnosis.

Sources cited:

Subscribe: YouTube

medAI Times is for educational and informational purposes only. The content does not constitute medical advice, diagnosis, treatment recommendation, or professional clinical guidance. Consult qualified healthcare professionals and refer to official sources before making clinical, research, regulatory, or business decisions.

Transcript

Auto-generated from the episode audio. Click any timestamp to jump the player there.

OpenAI just signed eight hospital systems the same week. Two studies said not so fast. Welcome to MedAI Times podcast, your daily update on medical AI. Don't forget to like and subscribe. From MedAI Times, here's the week in medical AI.

OpenAI signs eight major health systems and pitches hospital CEOs directly. The NHS puts AI scribes into every emergency department in England. Nature Medicine stress tests the top medical models and finds them brittle.

A large trial in Kenya says an AI helper made doctor's notes better, but left patients no better off. Whale Cornell builds a team of agents that design clinical trials. Philips gets an FDA nod for an AI ultrasound.

Medicare stands up a brand new office just for health tech. And Claude Cowork lands on the web in your phone. Last Friday, it was Anthropic putting Claude into the research lab. This week, OpenAI walked straight into the hospital.

I'll start with OpenAI. A Forbes feature this week laid out how the company is selling healthcare. And the short version is Sam Altman is personally pitching hospital chief executives. Eight health systems are now enterprise customers.

And these are not small names. Cedars-Sinai, HCA Healthcare, Memorial Sloan-Kettering, NYU, Langone, AdventHealth, Albany Med. OpenAI says 230 million people already ask ChatGPT a health question every week.

So it hired more than 260 physicians and reviewed 700,000 sample answers to sand down the rough edges. There are now three products in six months, ChatGPT Health for patients, a clinician version, and enterprise deals.

Here's the part worth holding on to. None of this is a clinical trial. It's reach, revenue, and record integrations. OpenAI lost $39 billion last year, which tells you why healthcare suddenly matters so much.

So the verdict, signal for hospital boardrooms, noise for the exam room. This is a land grab, not new evidence. And the deployment story with actual numbers came out of Britain. NHS England is rolling ambient AI scribes across accident and emergency and outpatient care nationwide.

The pitch, the AI listens to the visit and drafts the notes so the clinician isn't typing. St. George's Hospital in London clocked 47 minutes saved per shift. NHS England projects capacity for over 9,000 extra A&E consultations a day once it scales to 11,000 clinicians, real payoff.

But notice the ambition. The AI is taking notes, not making diagnoses. The first at-scale win for medical AI is deleting paperwork. Now the counterweight, and it's peer-reviewed.

Nature Medicine published a robustness study from Microsoft Research and Scripps with Eric Topol advising. They ran adversarial stress tests on the flagship models, GPT-5, Gemini, Claude, and specialized medical models too. The findings are humbling.

The systems could often guess the right answer, even when the images or key clinical inputs were removed. Which means they're leaning on shortcuts, not reasoning. Change a few words in the prompt, and they stumble, then produce confident but flawed reasoning to justify a wrong answer.

Pair that with a big real-world trial from the University of Birmingham. Run across 16 clinics in Kenya with more than 9,000 patients. The AI helper improved documentation and decision quality, but 14-day treatment failure was 2.2% with the tool versus 2% without.

Statistically, no difference. Better notes, no measurable benefit to the patient. That's the whole benchmark to bedside gap in one week, which makes the next item timely. A team at Weill Cornell, published in Nature Communications, built Emulet RX, five specialized agents that design a clinical trial together.

A supervisor, a trialist, an informatician, a clinician, and a statistician. They run what's called target trial emulation over de-identified records, essentially simulating the trial before anyone enrolls a single patient.

And they reproduce known treatment effects in heart failure, septic shock, and Alzheimer's. The authors are clear it needs multi-site validation before anyone leans on it. But attacking the design bottleneck before a dollar is spent is a smart place to point agents.

One more from the research bench. A group in Ekaternburg published a cardiac digital twin for resynchronization pacing. About 30% of patients who get these pacemakers don't respond, and nobody wants to implant hardware that won't help.

The twin builds a patient-specific 3D heart from MRI and CT, then virtually paces candidate sites before the procedure. Accuracy was 0.78 against 0.58 for the standard calculator.

The striking finding, in the pilot, eight of 13 non-responders had no viable pacing site at all. The model didn't just aim the lead, it flagged when not to operate. On the general side, an anthropic note.

Claude Cowork, its agent that runs longer tasks on your behalf, expanded this week from a desktop app to the web and mobile for max subscribers. Start a task at your desk, check it from your phone, pick up the finished output later, even with the laptop closed.

It's a small thing that signals where all of this is heading. Assistance that keep working when you've walked away. Tag that new. Spotlight. Adversarial stress testing of medical models is the technique to watch.

Instead of scoring a model on a clean benchmark, you strip out the image, shuffle the answer choices, or reword the question, then see if it still holds up. That nature medicine paper shows why it matters. A model can ace the test and still be guessing.

Expect regulators to start asking for this kind of robustness evidence, not just accuracy numbers. That's the week. Every link and source is in the description, including both peer-reviewed papers if you want to read past the headlines.

And the question I'll leave you with, the one I genuinely like your answer to, would you let a general purpose model like CHAT GPT draft any part of your patient's care? And if so, which part would you trust it with? Thanks for listening.

Find us on YouTube and your favorite podcast app. See you tomorrow.