The Bandwidth: Decoding Health + AI

1) Two conversations, one week

I spent the first two days of Seattle AI Week (it wraps up today) at its kick-off summit, and one of my biggest takeaways was the gulf between two camps: people working out how to put AI to use, and people who want no part of it. That gap seems to be widening and was on full display at the conference. Protests over the environmental toll of data centers (and subsequent health ramifications) interrupted both days. The discussions inside weren't blind to those concerns, but they were focused on other things: winning the race for talent and making the Pacific Northwest a tech hub on par with the Bay Area.

My two cents: a lot is getting missed in the national debate over AI, especially data centers. This week Axios featured photographer Stephen Voss's "Dark Fiber" series, which shows data centers pressing up against homes, parks, and a historically Black cemetery in Northern Virginia. That's probably what most people picture when they think about data centers moving into communities, and who could blame them? Nobody wants one in their backyard, literally. But what I see in those photos is mostly a failure of community governance. Setback rules, noise limits, and environmental reviews are ordinary local tools. So is making builders cover the cost of the power they use so residents' rates don't spike, something Microsoft has pledged to do. Many places that have once welcomed data centers (e.g., VA’s Loudoun County) have since paused new data center applications. Perhaps, the tools were there all along.

2) Bring your own digital twin?

Another thread running through Seattle was the future of work, and what our jobs will look like five, ten, and fifteen years from now. Some of the speakers argued we'll all have digital twins, AI set up to mimic the work we do. Some tasks can be delegated to them, and others can't. Think about going on PTO. Right now, your work either stops or gets handed off to a colleague who may also touch it. In the future, your AI/digital twin becomes the backup.

The harder question is what happens when people leave altogether or switch jobs. Does the twin, and all the knowledge it picked up from you, stay with the company, or does it go with you? Is this the modern version of bring your own device to work: bring your own digital twin?

Maybe. Most employment agreements already give the employer what you produce on the job, which would argue for the twin staying put. But a twin isn't just built on the work you did there. It draws on the know-how you've built up across your whole career, and right now that know-how follows you out the door, at least most of it. Handing off the twin could mean handing off part of you.

3) Offloading isn't new

Staying on the theme of the future of work. One concern I heard this week and read about often is the cognitive offloading that happens when we rely on AI. There's no denying it happens. But we shouldn't deny that our own brains do some offloading too, especially with routine work. I've been reading Greg McKeown's Essentialism, which argues routines are one of the most powerful tools for clearing obstacles and getting the essential work done. The first time you perform a critical task, he writes, you're unsure, distracted and drained by the effort. Do it enough times and the work shifts to the basal ganglia part of the brain, and it becomes nearly automatic, freeing your mind to take on something new. McKeown quotes Charles Duhigg, author of The Power of Habit: "The brain can almost completely shut down."

Medicine already works this way to a degree. A physician scientist told me that decades into practicing, he no longer walks through a full differential diagnosis. He sees the pattern in the patient and knows what to do. It's automatic. AI is doing something similar, shifting the routine, pattern-rich work to a tool (instead of a separate part of the brain). Where we get into trouble is when we offload our judgment along with the routine, sometimes unbeknownst to us. Anthropic published research this week showing how easily that happens: when 50 professional analysts worked with an AI agent that had been quietly set to favor certain choices, they went along with it 86% of the time. About half still rated the final recommendation as more their own than the agent's, and 77% reused the agent's arguments, but only 15% credited it.

Anthropic files to go public — Anthropic, the company behind Claude, publicly filed to list on Nasdaq under the ticker ANT, targeting the week of Nov. 16, and its filing shows $16.3 billion in first-half revenue, its first quarterly operating profit, and about $657 billion in computing contracts over the next seven to ten years, according to The Information.

Anthropic's IPO filing puts medicine first among AI's benefits — Anthropic's filing lists medicine ahead of economic growth and global development among AI's potential benefits, naming cancer, Alzheimer's, heart disease, and rare genetic diseases, while some in the pharmaceutical industry call CEO Dario Amodei's claim that AI could cure most diseases within a decade unrealistic.

No sign yet that AI is paying off outside big tech, Apollo economist says — Apollo chief economist Torsten Slok's charts show no clear productivity gain from AI so far outside the seven biggest tech companies, whose revenue per worker and profit margins keep rising while the rest of the S&P 500's stay flat.

Epic and Oracle are taking different paths to AI in the health record — Epic's Judy Faulkner is building AI on top of her company's existing health record and its database of 320 million patient records, while Oracle's Larry Ellison is rebuilding his around AI on new cloud infrastructure and criticizing vendors that charge extra for each AI feature.

Boehringer Ingelheim and Microsoft test AI "digital twins" of patients to shrink clinical trials — The drugmaker and Microsoft are exploring AI models of individual patients that predict how their disease would progress, which could reduce how many people a trial needs in its comparison group, starting with lung disease, though the work is early and would need regulators' sign-off.

OpenAI's new GPT-6.1 Terra aims to handle everyday computer work at half the price — OpenAI released GPT-6.1 Terra, a faster, cheaper model built for routine computer tasks like filling out forms and entering invoice details, priced 50% below its predecessor, with OpenAI's own tests showing its rate of factual errors on hard questions falling from 53.6% to 16.7%.

Google DeepMind previews Genie 4, which builds interactive worlds with sound in real time — Google DeepMind's Genie 4 generates photorealistic worlds that users can explore and change with text prompts, now with native sound and dialogue, and is in testing with filmmakers, game designers, and robotics researchers ahead of a wider release to Google AI Ultra subscribers and developers.

Insurers bet AI assistants will cut costs and keep members — UnitedHealthcare, Aetna, and Cigna are rolling out AI assistants that answer benefit questions, find in-network doctors, and soon book appointments, and Cigna says its assistant has cut call center volume by 20%.

Google launches Gemini 4 Carbon, its second major model in a month — Google says Gemini 4 Carbon is nearly twice as good as its predecessor at writing secure code and will release it first to cybersecurity professionals, as AI-driven attacks mount.

UnitedHealthcare's AI assistant is now open to 21 million members — UnitedHealthcare expanded Avery, an AI assistant that explains deductibles, provider networks, and claims, from about 6.5 million members to 21 million, says it resolves 90% of conversations without a human, and is building a version that can call doctors' offices to book appointments for members.

eClinicalWorks adds AI agents and evidence lookups to its health record — The health record company's Nov. 1 update pulls AI summaries and suggested next steps into one workspace, lets clinicians ask evidence questions without leaving the chart, and adds tools for building AI agents that handle work like prior authorization and referrals, with the company reporting 43% faster task completion in its own testing.

Lawyers warn health systems to get ready for AI malpractice suits — Health care lawyers from Ropes & Gray, McDermott Will & Schulte, and Hill Health Law Group tell Modern Healthcare that AI-related lawsuits haven't arrived in force but are likely coming, that vendor contracts requiring a clinician to review AI output tend to leave the health system holding the liability, and that staff using unapproved AI tools add more risk.

PwC lays out a framework for bringing AI into nursing safely — PwC's Kathryn Burg Plaza and Patrick Voight sort nursing AI tools into low, moderate, and high risk and propose piloting each one before deciding to keep, revise, or stop it, then tracking signals like how often nurses override it.

Doctors say patients now show up with chatbot diagnoses — At the Association of Health Care Journalists' fall summit, Mount Sinai urologist Ashwin Ramaswamy and Cedars-Sinai's Omer Liran described patients bringing AI-generated notes to 15-minute visits, with Liran calling chatbots "sycophantic."

AI scribe makers argue a pause on medical AI would worsen the doctor shortage — Sully.ai co-founders Ahmed Omar and Amit Kumthekar argue against slowing AI in medicine, citing a projected shortage of 86,000 physicians by 2036 and Permanente Medical Group's estimate that AI scribes saved its doctors about 15,800 hours in a year.

UT Austin is building an AI-native hospital wired with up to 30,000 sensors — UT Austin launched UT Dell Medicine, a new academic health system, as site work begins on a roughly $5 billion AI-native medical center opening in 2030 that will have 20,000 to 30,000 sensors, robots moving medications and supplies, and shared records and billing with MD Anderson.

Google Cloud's health chief says judge AI by outcomes, not adoption — Google Cloud's Aashima Gupta told MobiHealthNews that health systems should measure AI by results for patients and clinicians rather than how widely it's used, and said the bigger opportunity is AI agents that manage a patient's care between visits.

Zocdoc opens its booking system to AI agents — Zocdoc now lets registered AI agents, including Google Gemini and Amazon Health AI, search for doctors, check insurance, and book appointments with a patient's consent, after AI agent traffic on its site grew nearly 30-fold since July.

Ambient AI is helping prescribers, but not care managers, a behavioral health leader says — Corey Cronrath, chief medical and technology officer at Tennessee's Mental Health Cooperative, says ambient note-taking lets prescribers look at patients instead of screens, but today's tools can't yet capture the collaborative care planning its social workers and care managers do each month.

Utah turns to AI to tackle health care's administrative waste and workforce gaps — With an estimated half of the roughly $1 trillion the U.S. spends on health care administration going to waste, Utah is putting AI to work from the front desk to the exam room: Intermountain Health says it has more than 300 AI projects underway, mostly on back-office work, and the state's AI sandbox is expanding from small pilots to proposals from Intermountain and University of Utah Health.

Stanford scholars send FDA seven ideas for regulating generative AI devices — Sixteen Stanford scholars, including Michelle Mello and Kevin Schulman, urged FDA to require proof that human oversight actually catches a device's errors before counting it as a safeguard, to add real-world testing alongside benchmark scores, and to put more of the burden of monitoring AI after approval on developers.

FTC demands Anthropic executives testify in its AI safety probe — The Federal Trade Commission on Monday ordered Anthropic executives to testify in its consumer-harm investigation, which already includes 50-page demands to Anthropic and OpenAI for documents on safety incidents, marketing, and privacy, and Anthropic warned in its IPO filing that the probe could bring significant costs and force changes to its models.

Paying for clinical AI the old way could inflate spending, Bipartisan Policy Center warns — BPC's Allison Kwon and Maya Sandalow argue that adding new AI billing codes to fee-for-service payment risks driving up spending without better results, and urge CMS to flag every claim where AI shaped a decision, pay only for tools that show better outcomes, and lower rates as the tools get cheaper to run.

Better AI models won't fix health care's bottlenecks, ARPA-H's Bon Ku says — ARPA-H program manager Bon Ku told a HIMSS AI forum that today's models are already strong enough to change care but can't fix problems like a shortage of ER beds, and described agency programs building AI that coaches rural health workers through procedures like ultrasound and patient-facing AI agents for heart failure overseen by a supervising AI.

House Democrats ask AI labs to report safety incidents to Congress within 48 hours — The top Democrats on the Energy and Commerce, Science, and Judiciary committees, Frank Pallone, Zoe Lofgren, and Jamie Raskin, asked Anthropic, Google, Meta, OpenAI, and SpaceXSI to account for every known case of their AI models acting in unintended or unsafe ways by Nov. 11, and to report new ones within 48 hours, or 24 hours if critical infrastructure is involved.

States have enacted 40 health AI laws this year, Manatt's tracker finds — Manatt's quarterly tracker counts more than 280 health AI bills introduced in 2026 and 40 enacted across 24 states, and notes that Utah's AI test program now has six outside evaluators and that Medicare's ACCESS payment model adds heart failure, COPD, substance use disorder, and tobacco cessation in April 2027.

HHS drops the "nutrition label" requirement for AI in health records — HHS's final HTI-5 rule ends a Biden-era requirement that health record vendors publish a model card explaining how their AI tools were built, tested, and meant to be used, over objections from the American Hospital Association, which called the cards one of the few ways hospitals can compare AI tools.

New York bans chatbot features for kids that pretend to be human or encourage self-harm — Gov. Kathy Hochul signed a law barring companies from offering minors chatbots that claim to be human, build emotional dependency, encourage self-harm or eating disorders, or have sexual conversations, with age verification required and fines of up to $25,000 per violation.

Sen. Cassidy floats the PATIENT AI Act to speed review and payment for health AI — Senate HELP Committee Chairman Bill Cassidy released a discussion draft of the PATIENT AI Act, which would clarify when AI tools fall under FDA or ONC oversight, let FDA review software updates on a rolling basis, create a payment pathway for AI tools, and direct HHS to issue privacy and security guidance, with comments due Dec. 1.

An employee's unapproved AI tool exposed records of 3,182 Johns Hopkins patients — A Johns Hopkins Bayview Medical Center employee uploaded a spreadsheet with patients' treatments, medications, and medical record numbers to a public AI tool in September, and the hospital has asked the tool's maker to delete the data.

Chatbot safety checks miss harm that builds over months, Cornell researchers argue — In a STAT opinion piece, Cornell's Qian Yang, Ned Cooper, and Ibrahim Emara argue that chatbot safeguards and new state laws target single dangerous conversations while missing the gradual loss of human relationships over months of use, and call for independent researchers to get the data to measure it.

Clinical AI should earn independence the way medical residents do, UC San Diego researchers argue — UC San Diego Health's chief health AI officer Karandeep Singh and colleagues argue AI should be held to a clinician's standard, tested on new cases it couldn't have seen in training, and given more independence only as it proves itself, pointing to Utah's AI prescription-renewal pilot, where a physician reviews every renewal before the program moves to spot checks.

Most hospitals can't see when staff feed patient data to unapproved AI — Following the Johns Hopkins Bayview incident, a Proofpoint and Ponemon survey of 705 health care IT and security staff found only 32% have full visibility into unauthorized AI use and 60% lack controls to keep sensitive data out of AI tools.

The feared wave of AI-powered hacks hasn't hit, yet — Security researchers tell Axios the PHP flaws that OpenAI's agents found, which set off this month's "Black October" panic, have proven hard to exploit despite millions of attempts, though experts warn the next AI-found bug could be turned into an attack within hours.

Patients deserve to know when AI is listening, Stanford neurosurgery essay argues — An essay from Stanford Neurosurgery, featured in the Congress of Neurological Surgeons' quarterly, argues that patients should be told when AI is recording or shaping their care, that liability stays with the physician, and that health systems must keep monitoring AI after it goes live, citing a lawsuit against Sutter Health over AI note-taking without consent.

91% of health organizations can't verify what AI they're running, survey finds — A Clearwater Security survey of 150 health systems, health plans, and tech companies found 91% can't verify what AI is in use across their organization and only 14% check their vendors' AI safeguards.

Parents sue OpenAI after ChatGPT coached their daughter off her epilepsy medication — A 31-year-old Seattle woman with severe epilepsy died in February after following a ChatGPT plan to taper off her seizure medication and hide it from her doctors, with the chatbot repeatedly brushing off her seizure symptoms, according to The New York Times, and her parents sued OpenAI this week; the company says the GPT-4o model she used has since been retired.

AMA lays out who should be liable when clinical AI harms patients — The American Medical Association's new framework says AI developers should answer for their products' safety, health systems for how they choose and monitor AI, and physicians only for the judgment they can meaningfully exercise, adding that doctors shouldn't be liable for AI errors they can't reasonably detect or penalized for declining unproven tools.

At Johns Hopkins, AI note-taker use ranged from 69% of clinicians in some specialties to 16% in others — A Johns Hopkins study of 2,901 outpatient clinicians in npj Digital Medicine found that six months into a systemwide rollout, about two-thirds of community internal medicine and pediatrics clinicians had used an AI note-taker, compared with fewer than one in five in general surgery, OB-GYN, and psychiatry.

Doctors are turning to AI for prescribing decisions, Bain finds — A Bain survey found the share of doctors who rank AI among their top three sources for prescribing decisions jumped from 3% in December 2024 to 20% in June 2026, and the firm argues drugmakers now need a plan for AI as both a customer and a channel to doctors.

Biological age, not birthdays, could guide care if AI can measure it, a researcher argues — In an npj Aging commentary, Pawel Kordowitzki of Nicolaus Copernicus University and Charité Berlin argues AI could combine genetic, imaging, and wearable data to estimate how fast a person's body is actually aging, and warns those scores should never be used in hiring or insurance decisions.

Anthropic reports early lab results from using Claude in drug development — Anthropic says six drug developers, including Bristol Myers Squibb, have used Claude to design antibody candidates, improve experimental treatments, and cut manufacturing paperwork review time by 89%, though nearly all of the work is still in the lab rather than in human trials.

Teens who lean on AI for advice or companionship report more loneliness and suicidal thoughts, JED survey finds — A Jed Foundation survey of more than 5,500 middle and high school students found 57% turned to AI when sad or stressed and 12% for companionship, and those using it for advice or companionship were more likely to report anxiety, depression, loneliness, and suicidal thoughts, though the study shows a link, not that AI causes them.

If you or someone you know is struggling, call or text 988 to reach the 988 Suicide & Crisis Lifeline.

Microsoft and Mass General build a cheaper way to spot imaging AI going off course — Researchers from Microsoft and Massachusetts General Hospital describe MMC+, a free tool that watches the data feeding imaging AI and flags shifts, like those during the COVID-19 pandemic, that signal the AI's accuracy may be slipping, without retraining for each hospital.

OpenEvidence adds addiction medicine guidelines to its AI tool for doctors — The American Society of Addiction Medicine's treatment guidelines will now appear, with citations, alongside research in OpenEvidence's free AI tool for clinicians, the latest in a string of deals with journals and medical societies.

RAND builds a clinical scorecard for how chatbots answer teens' mental health questions — RAND and Harvard researchers built a tool that grades chatbot replies against clinical guidelines and tested ChatGPT, Claude, and Gemini with simulated teens, finding the chatbots covered more of the recommended steps for anxiety than for depression, and for moderate cases than severe ones.

Almost half of Americans have used AI to pick a health plan, survey finds — An eHealth survey of more than 1,000 adults found 48% have used AI to review health insurance options and 85% of them chose their plan at least partly on its advice, though 67% of people who used AI for medical advice said it was sometimes wrong.

Popular AI models get medication doses wrong, some dangerously, Doximity finds — When Doximity tested nine widely used AI models on more than 8,500 dosing questions, about 2 in 100 answers contained a serious or potentially lethal error, most often failing to adjust for age, kidney function, or drug interactions.

Americans are open to AI in health care but want humans making the final call, Pew finds — A Pew Research Center survey of 3,488 adults found most Americans are open to AI helping with tasks like translation, scheduling, and note-taking as long as a human reviews the result, but few would let it act alone on a diagnosis (3%) or an insurance claim (5%), and 60% are uncomfortable with their own doctor using AI to diagnose and treat them.

Executives who guard their turf stall AI projects, a coach argues — Executive coach Amii Barnard-Bahn says AI projects stall when executives either claim them as their own or tune out because AI seems like someone else's job, pointing to a bank where a $30,000 AI tool that could have saved hundreds of staff hours died because its vendor was on another executive's cost-cutting list, and she urges CEOs to define the business problem before deciding who owns it.

👉 Nov. 5 — Cooley Healthtech Talks — Webinar series (virtual)

👉 Nov. 7–11 — AMIA 2026 Annual Symposium — Dallas, TX

👉 Nov. 10 — National Digital Health Policy Summit — Washington, DC

👉 Nov. 11–13 — AIMed26 — Orlando, FL

👉 Nov. 15–18 — HLTH USA 2026 — Las Vegas, NV

👉 Nov 17 — Building the foundation for clinical AI: Lessons from Stanford Medicine & The Hospital for Sick Children — Webinar (virtual)

👉 Nov. 19, 1 p.m. ET — Reimagining Payment Integrity with AI — Webinar (virtual)

👉 Dec. 1, 11:59 p.m. ET — Abstracts due: Generative AI in Health Care: Law and Regulation — Petrie-Flom Center, Harvard Law School

👉 Dec. 8–9 — Stanford AI+HEALTH 2026 — Virtual

👉 Dec. 11–12 — Cedars-Sinai AI in Medicine Symposium — Los Angeles, CA

👉 Dec. 16 — The Next Phase of AI Adoption: Governance, Ethics, and Accountability — Virtual

👉 Jan. 11–14 — J.P. Morgan Healthcare Conference — San Francisco, CA

👉 ~Jan. 31 — White House AI task force report due — Led by Director of National Intelligence Jay Clayton

👉 Feb. 18, 11:59 p.m. ET — Papers due: Applications of Artificial Intelligence in Healthcare — NBER