|
3, 2, 1: Health AI Brief
Every Friday
May 8, 2026
|
|
|
AI is reshaping healthcare fast. Below are 3 key AI developments, 2 studies, and 1 takeaway for this week to help you better lead with AI. Target read time: 5 minutes. |
|
On May 5, UnitedHealthcare announced it will eliminate prior authorization requirements for 30% of services by year-end 2026, including select outpatient surgeries, echocardiograms, and some outpatient therapies and chiropractic care. CEO Tim Noel framed the shift: "Prior authorization is an essential safeguard but should only be used when it truly protects patients and improves care." UHC noted it currently requires PA on only 2% of medical services, with 92% of submitted requests approved within 24 hours. So what?
The largest US insurer just walked back about a third of its prior auth rules, voluntarily, after years of class-action pressure on AI-driven denial tools and bipartisan scrutiny in Congress. The 2% / 92% framing is a defense of the current scope; the 30% cut is an admission that that scope was still too wide. I expect others to follow. On May 5, Pennsylvania's Shapiro administration filed suit in Commonwealth Court against the company behind Character.AI, alleging its chatbots engaged in the unlawful practice of medicine. According to the state, a chatbot called "Emilie" presented itself to a state investigator as a licensed psychiatrist and fabricated a Pennsylvania medical license number. Governor Shapiro's office calls it the first enforcement action of its kind by a governor in the United States. Character.AI responded that its bots are user-created roleplay characters with prominent disclaimers reminding users that "everything a Character says should be treated as fiction." So what?
Pennsylvania just brought the unlicensed-practice-of-medicine framework to AI companions. That framework predates the internet, and it's being applied to AI products now: if a chatbot says it has a license, the state can ask for it. How courts decide this will set precedent well beyond Pennsylvania. Read the Pennsylvania announcement → | Read the TechCrunch coverage → OpenAI recently published a policy document titled "Keeping Patients First: A Blueprint for AI in U.S. Healthcare," released alongside the company's ChatGPT for Clinicians launch. The blueprint centers on three themes: patient-directed data portability (extending information-blocking rules to labs and pharmacies), clinician-supervised AI deployment, and modernization of FDA pathways including new "regulatory sandboxes" for AI-enabled medical software. David Blumenthal, the former national coordinator for health IT under President Obama, told STAT that OpenAI is "trying to have their cake and eat it too": the proposals sound responsible while keeping regulatory paths clear for OpenAI's own products. So what?
Both blueprint and critique have a point. That doesn't make Blumenthal's "cake and eat it too" framing wrong. It just means the proposals should be scrutinized based on their merits, not discredited outright based on who wrote them. The substance covers real gaps our system needs to fix. Read the OpenAI blueprint (PDF) → | Read the STAT critique → |
|
A Mayo Clinic study published in Gut validated REDMOD (Radiomics-based Early Detection Model), an AI tool that analyzes routine abdominal CT scans for textural patterns of early pancreatic cancer that radiologists can't see on their own. Across nearly 2,000 CT scans, REDMOD identified 73% of pre-diagnostic cancers at a median of about 16 months before clinical diagnosis, nearly doubling specialist-only detection rates. For scans taken more than 2 years before diagnosis, the AI flagged nearly 3 times as many early cancers as specialists reviewing the same images. The model now moves into a prospective trial called AI-PACED, which will evaluate clinical integration in patients at elevated risk. Why it matters
Pancreatic cancer is the cancer where almost everything depends on stage at diagnosis, and almost everyone is diagnosed late. A model that pulls signal out of CTs already being done for other reasons doesn't ask the system to add a screening step; it asks it to read what's already there. That's a much smaller operational lift than building a new pathway, and a useful pattern I'm expecting/hoping will emerge elsewhere in oncology, too. The Massachusetts Nurses Association's 2026 State of Nursing survey of 485 active registered nurses found that 38% reported using AI in their work, more than double the 18% who said the same a year earlier. The same survey found that 80% had received no AI training from their employer, 50% said they're uncomfortable using AI tools, and 81% expressed concern about being held liable for AI-generated harm. Why it matters
AI use among nurses doubled in a year, from 18% to 38%, yet that's still well below the 81% figure for doctors from the AMA earlier this year. Per the nurses themselves, the big gap is in training, comfort, and clarity around liability. |
|
Friction favors scale.
This week, the deployment side of healthcare AI got harder to operate in; the production side consolidated into platforms with the scale to operate through it. On the deployment side: UnitedHealthcare, the largest US insurer, said it will eliminate prior authorization for 30% of services by year-end. Pennsylvania filed what its governor called a first-of-its-kind enforcement action against an AI company whose chatbots claimed medical licenses. And the Massachusetts Nurses Association reported that 4 in 5 nurses had received no AI training from their employer, even as AI use among nurses more than doubled in a year. On the production side: Roche announced a definitive agreement to acquire PathAI for up to $1.05 billion. It's a pharma-side echo of Microsoft's 2022 $19.7 billion Nuance acquisition (voice and AI scribe) and Oracle's 2022 $28.3 billion Cerner purchase (EHR). Takeaway
What I find striking is who benefits. Friction comes in many forms: legal restrictions on AI products, voluntary retreats from AI's higher-stakes use cases, workforce gaps that force AI platforms to compensate for absent training. All of them favor scale. Only consolidated platforms have the deep pockets to make upfront investments, absorb compliance + engineering costs, and delay revenue long enough to build the legal, regulatory, and clinical scaffolding required. Ultimately, each tightening increases lock-in, making incumbents harder to displace. |
|
|
Know someone who'd find this useful? Share |
Restrictions favor scale.
3 signals, 2 studies, 1 insight for health leaders
May 8, 2026
•
11 min read