ChatGPT Health Now Reads Your Medical Records — Jul 24, 2026

0:00300 million health questions a week, CHAT GPT just opened the doors. Welcome to MedAI Times podcast, your daily update on medical AI. Don't forget to like and subscribe. CHAT GPT Health opened to every American adult this week, and it can now read your medical records.

0:17The FDA gave one company a pass on part of pre-market review in exchange for Medicare data. The White House started drafting the test that clinical AI will have to pass. A randomized trial in nature medicine put an IAI beside doctors in three countries.

0:33Utah published six months of data on letting software renew prescriptions. Medtronic cleared AI that watches the operating field in real time. Tempest bought Personalis for one and a half billion dollars. Epic lost the executive who ran its AI program.

0:48And AMD put $5 billion into Anthropic. Yesterday, we counted how many FDA cleared AI tools reached the market with no clinical study at all. Today, the White House started writing the test they would have to pass.

1:03Start with OpenAI. CHAT GPT Health went live for every logged in adult in the United States on the free tier as well as Go Plus and Pro on the web and on iOS.

1:16It connects to Apple Health, to nutrition and fitness apps. And here is the part clinicians should notice, to medical records held in Epic and Oracle Health. OpenAI says people now ask it 300 million health questions a week.

1:29In January, that number was 230 million. 70% of those questions happen outside the dedicated health tab, which means people are asking about their bodies in the middle of ordinary conversations. The model underneath is GPT 5.6 Luna, which the company says beats its predecessor on HealthBench, OpenAI's own evaluation.

1:50What OpenAI does not have is an outcome trial. The terms still say the service is not intended for the diagnosis or treatment of any condition, a sentence written by lawyers for a product used by patients.

2:03And it went live one day after a Florida pastor sued the company, alleging the chatbot told him not to see a doctor. My read. Noise at the bedside, signal in the waiting room. The FDA did something this week it has never done.

2:17It named Dexcom the first participant in a pilot called Tempo, which lets selected digital health devices skip certain pre-market requirements while the manufacturer builds real-world evidence through Medicare.

2:31The paired Medicare model, run out of the Innovation Center, pays for those devices in four chronic disease areas, including diabetes and hypertension. The agency plans to take up to about 10 manufacturers per area.

2:44Dexcom's entry is an AI glucose program that folds continuous monitor data together with nutrition, activity, sleep, and stress. Read it plainly. This is the FDA trading review up front for evidence after the fact, and the evidence will be generated by the company that wants the payment.

3:00Which makes the second Washington story land harder. The White House Office of Science and Technology Policy, the FDA, and the Office of the National Coordinator are convening outside experts for a one-month sprint on how clinical AI should be benchmarked and evaluated.

3:17Written phase, then discussion phase, aiming at what the invitation calls a consensus set of principles. This was reported by a stat from an invitation they reviewed. And I could not find a matching press release from any of the three agencies, so treat the details as one sourced for now.

3:35Still, the direction is unmistakable. After years of clearing tools, the government is finally asking what counts as proof. The week's strongest evidence came out of Nature Medicine on Friday. A model called Retina4IRD reads retinal images and predicts which of 17 gene categories is causing an inherited retinal disease.

3:57In a multi-center randomized trial across China, South Korea, and Poland, 295 patients with genetic confirmation, clinicians using the AI landed the right gene in their top five guesses 88.5% of the time.

4:12The control clinicians managed 67.3%. Management decisions improved too. This is the shape of evidence the field keeps asking for, and it is worth noting that AI here assists a clinician rather than replacing one.

4:27Utah has spent six months letting an AI system autonomously handle prescription renewals for roughly 190 chronic condition drugs, and Forbes got the first data. In the first five months, the AI recommended approval 72% of the time, and physicians agreed with it in 91% of those cases.

4:47Of the 28% it escalated to a human, doctors judged 69% of the escalations warranted. Cost, about $4 per prescription. One detail sticks.

5:00The chair of Utah's medical licensing board said the board learned about the program from news coverage, not from state officials. Surgery got its own clearance. Medtronic unveiled Touch Surgery Aid at the Society of Robotic Surgery meeting, a computing platform built on NVIDIA hardware that runs multiple AI applications on live surgical video at once.

5:21Its first application, Instrument Exit Point, is FDA cleared for the Hugo robot and flags when an instrument drifts outside the surgeon's view. No performance data was released with the announcement, which is the usual pattern and the usual problem.

5:36But the platform framing matters more than the first feature. On the general side, AMD is investing up to $5 billion in Anthropic, which will deploy up to two gigawatts of AMD Instinct chips starting in the first half of 2027.

5:53AMD will run CLAWD across its own engineering teams in return. Anthropic also opened rare disease research grants this week, up to $50,000 in CLAWD credits per project, applications closing August 2nd.

6:06With the Monarch Initiative and Every Cure among the partners. And Fable 5 is now permanently included for Macs and Team Premium subscribers. Spotlight. The shortcut everyone in clinical AI has quietly adopted is using a large language model to grade other large language models because human expert review does not scale.

6:26A paper in NPJ Digital Medicine this week tested that directly on questions from Rwandan health workers. Five model judges were compared against six human clinicians and the best one matched human judgment on only four of 11 criteria.

6:41Building juries of models fixed exactly one more. Performance dropped further in Kinyarwanda than in English. Every source, every study, every link is in the description including the Nature Medicine trial and the Utah data.

6:55And here is what I want from the clinicians listening. Has a patient ever handed you a chat GPT summary of their own chart? And did it change what you did in that visit? Thanks for listening. Find us on YouTube and your favorite podcast app.

7:10See you tomorrow.