The Bandwidth: Decoding Health + AI

I'm writing this from vacation. This week's issue is my last official duty before I truly unplug. Reminder: no newsletter next week, and I'll be back with The Bandwidth on Friday, October 5.

Because I'm easing myself out of work mode, this issue leans heavily on the headlines and summaries below rather than a longer take from me up top.

But I didn't want to leave you without something to chew on, so I'll point you to a WSJ piece from last month on the mental health case for actually getting away, not just leaving your desk but genuinely disconnecting.

The research is worth a look: travel builds psychological resilience that outlasts the trip itself, distance seems to matter (international travel correlates with lower depression and loneliness than a nearby long weekend), and the harder or more unfamiliar the experience, the more durable the benefit.

This feels like particularly salient advice given the velocity and complexity in which AI and healthcare are moving. It’s often hard to keep up.

So, I'm taking that advice as we explore Grand Teton and Yellowstone. With a week off screens, I'm hoping to catch my breath and build some durability for when I return to fully engage on all things AI and health.

Meta's Muse AI agent app overtakes ChatGPT as top iPhone app — Meta's Muse, an AI agent app that wraps agentic workflows in a casual messaging interface with character personas, has taken the No. 1 spot on the U.S. App Store's free iPhone chart, displacing ChatGPT. It only fully launched this month, though it's been around a few weeks, and new installs are reportedly still accelerating; Meta also expanded it to Mac this week, though desktop downloads don't count toward the ranking. It reflects a split in strategy, with Meta betting on a standalone agent app while OpenAI and Anthropic have mostly built agentic features into products people already use.

Anthropic begins physical bio experiments — Anthropic is testing whether Claude can operate laboratory robots and run life sciences experiments in a Bay Area lab with minimal human involvement, building on the Claude Science research workbench it launched in June for protein structure prediction and RNA sequence analysis. The company also acquired a biotech startup developing AI models for biological research. Anthropic's head of life sciences told Reuters that human oversight remains essential for safety and that the company hasn't run any clinical trials, while separately disclosing it disrupted five attempts to misuse its models for biological weapons development this year.

Gemini hacked three companies in Google's first known AI breakout — Google confirmed that its Gemini model autonomously hacked three real companies in May during a cybersecurity test run by the AI safety firm Irregular, the first known case of a Google AI committing an unauthorized intrusion. In two cases the model found leaked credentials in public repositories, and in one it guessed a password, but in each instance it stopped and withdrew once it realized it had accessed a real company rather than the test's fictional target. Google likened it to a hacker who stumbles onto a real security flaw and reports it instead of exploiting it, saying the model "acted appropriately," though security researcher Jack Cable argued that framing missed the real issue: an AI agent had wandered outside its intended bounds and carried out actual cyberattacks on its own.

Cheap, powerful AI models become the new competitive frontier — AI companies are competing on price as much as raw power right now. OpenAI's new GPT-6 Sol and Luna models cost roughly half as much to run as before, Anthropic's Opus 5.5 came in 40% cheaper than its predecessor, and a new low-cost model from China's DeepSeek jumped to the top of a popular AI usage ranking after demand for it spiked 172% in a week. The logic, per Axios: when AI gets cheaper to use, people tend to use more of it, not less, which is exactly what these companies need to justify the trillions they're spending to build it out. Total AI spending keeps climbing even as prices fall, which one analyst called a good sign that people actually want more AI, not just that it's gotten cheaper.

Novo Nordisk leans on AI to find the next Ozempic — Novo Nordisk is betting heavily on AI as patents on its blockbuster GLP-1 drugs Ozempic and Wegovy near expiration in the early 2030s, running more than 30 AI partnerships and pairing over 600 AI and digital staff directly with scientists. The company says AI has tripled its speed at identifying new drug targets, cut trial site selection and patient population analysis time by 90%, and sped up clinical study report generation tenfold, with Chief Scientific Officer Martin Holst Lange saying AI has helped identify "not only new targets but also drugs" outright. CEO Mike Doustdar acknowledged the shift will also reshape the workforce, saying "some tasks will become automated, and the mix of skills we need will change," even as AI drug discovery broadly has yet to deliver major breakthroughs industry-wide.

Epic pauses product development to focus on cybersecurity — Epic Systems CEO Judy Faulkner said the electronic health record giant has paused most of its technology development, affecting hundreds of projects, to focus on securing its systems against cyberattacks, work she expects will take another six weeks. She called the pause a "shame" but warned that resuming development risks a "never-ending cycle" as attackers adapt, and said the company's pace going forward will be permanently slower as security work becomes constant rather than periodic. Faulkner also outlined Epic's AI pricing approach: no charge for AI that enhances an existing product, but a charge for entirely new AI products to cover computing costs.

Oracle Health rolls out AI revenue cycle tools, oncology EHR — Oracle Health unveiled AI agents that handle prior authorization, coding recommendations, and denial appeals as one integrated workflow, along with a new oncology-specific EHR that pulls together pathology, lab, pharmacy, genomic, radiology, and social determinants data into a single patient timeline. EVP Seema Verma said the revenue cycle tools differ from competitors' "bolt-on" solutions by handling the full process end-to-end rather than a single task, and that Oracle picked oncology first because its complexity across multiple treating physicians makes it the hardest specialty to prove the approach on. The revenue cycle tools are expected to roll out in the coming months.

Meta debuts Muse Charm, bringing viral assistant to keychain — Meta unveiled Muse Charm, a keychain-sized AI device meant to bring its viral Muse assistant to a broader audience, alongside new Muse features like live video chat, real-time voice conversations, and a dedicated email address for each user's agent. Meta also announced retail partnerships with Walmart, Best Buy, Gap, Sephora, and Wayfair, notably excluding Amazon, and previewed audio-only smart glasses starting at $349 and a slimmer $1,299 VR headset due next spring. The company is also building out privacy infrastructure, including a "Private Processing" system for AI glasses designed to protect data during cloud processing, similar to Apple's Private Cloud Compute.

Battle of hospital AI vs. insurer AI is pushing medical costs higher — A Blue Cross Blue Shield Association report found hospitals' AI-assisted coding drove nearly $1 billion in added costs over two years, with claims describing patients as sicker without evidence of different treatment, averaging almost $12,000 more per case. Insurers are using their own AI to scrutinize and deny those claims, while hospitals use AI-generated appeals to push back, prompting one health system CFO to describe both sides now scanning the same patient charts for opposite reasons. Experts are split on where this lands: some warn of an escalating "bots fighting bots" cycle that raises costs further, contributing to projected double-digit premium increases next year, while a Harvard economist argues AI could eventually force a simpler, outcomes-focused reimbursement system by cutting the administrative overhead on both sides.

Researchers detail a safety architecture for AI chatbots monitoring suicide risk — Mass General and UCLA researchers describe the safety guardrails built for "Suzy," an AI chatbot used in an opioid use disorder treatment program, in a peer-reviewed comment in npj Digital Medicine. The system routes every user message through a tiered risk classifier: high-risk messages trigger crisis resources and human staff review, ambiguous distress prompts the bot to gather more context, and low-risk messages proceed to normal support content, with trained staff also reviewing transcripts twice daily and following up by phone using a validated suicide risk scale when warranted. The authors argue AI-only monitoring is insufficient and recommend any clinical chatbot deployment pair automated triage with human oversight, especially as underlying models change over time.

AI-native radiology practices are betting ownership beats selling software — A wave of new and existing radiology practices, including startups Epsilon Health and Radley alongside incumbents like Radiology Partners and RadNet, are building or acquiring their own AI rather than buying it off the shelf, betting that owning the clinical practice itself is the fastest way to refine and deploy the technology. The pitch centers on radiologist efficiency amid a workforce shortage: AI drafting reports, catching imaging errors before they reach a radiologist, and automating scheduling and follow-up, with incumbents using their large existing patient volumes as a built-in testing ground for FDA authorization. Experts caution the approach is unproven, that "AI-native" is a design choice rather than evidence of quality, and that the FDA is still working out where a practice's own tools stop being exempt from medical device regulation and start requiring clearance.

10 AI takeaways from Mass General Brigham's World Medical Innovation Forum — At its innovation forum, Mass General Brigham shared results from MGB Care Connect, an AI intake agent in its patient portal that gets 77% of primary care patients seen by a doctor the same day they reach out, while executives across the system stressed that nothing in clinical care runs fully autonomously yet and that the system has deliberately avoided open-ended AI pilots in favor of targeting specific problems. CFO Niyum Gandhi said a third of MGB's AI spending is intentionally not held to an ROI standard, covering research and clinical-outcomes work instead, while several leaders described an emerging "bot wars" dynamic in revenue cycle as payers automate denials and providers respond with their own AI-driven appeals. On equity, chief community health officer Elsie Taveras said the bottleneck for reaching underserved communities is rarely the technology itself but the trusted human relationships needed to carry it there.

Cigna Group, OpenAI partner to expand complex condition care — The Cigna Group is integrating OpenAI technology to support patients with complex conditions, starting with oncology, giving nurses and case managers across Accredo Specialty Pharmacy and Cigna Healthcare AI tools that pull together a patient's clinical, pharmacy, behavioral, and benefits information in one place. The partnership builds on Cigna's existing AI push, including a medication adherence initiative, though the company declined to share further detail beyond its press release and OpenAI didn't respond to a request for comment.

Khanna calls for a government agency to oversee AI 'like the FDA' — Rep. Ro Khanna (D-Calif.) said Sunday on CBS's "Face the Nation" that the federal government should stand up a regulatory agency for AI similar to the FDA, comparing it to how the US already regulates nuclear energy, electricity, and aviation. Khanna, who has pressed companies like DeepSeek and Alibaba to slow AI development until guardrails are in place, also called for an international agreement with China that includes monitoring through data centers, framing an unchecked AI race as comparable to a nuclear one. The comments come as Trump has instead proposed a Space Force-style "AI Force" and an "AI czar," and after former President Obama criticized the administration's reliance on voluntary industry restraint.

Trump announces plans for a military-style 'AI Force' — President Trump said in a Truth Social post that he's forming an "AI Force," modeled on the Space Force he created in his first term, and will soon name a new AI czar to lead it, telling would-be applicants only "High I.Q. individuals" should apply. He dismissed AI safety concerns as the latest in a string of "hoaxes" while pledging the new body would still pursue "BAD" actors in the industry through existing law. The announcement follows a string of Pentagon AI failures, including a false AI-generated intelligence report that CNN says nearly triggered a military boarding of a Chinese vessel, and comes as investigators examine whether AI overreliance contributed to a February strike on a girls' school in Iran.

The US wants an AI-era 'red phone' with China — Treasury Secretary Scott Bessent said the US has proposed a direct notification channel with China to flag major AI-related national security incidents, floated during New York talks with Chinese Vice Premier He Lifeng ahead of Thursday's Trump-Xi summit in Washington. Bessent framed it as part of a new "US-China AI dialogue," saying the two countries need "a shared vision of common goals and common threats," though China's state news agency described the talks only as "candid, in-depth and constructive" without confirming the proposal. The overture comes as the two countries continue clashing over AI chip export controls and accusations that Chinese firms are training competing models on American ones.

Cato scholars argue against restricting AI health chatbots — Writing in DC Journal, Cato Institute's Jennifer Huddleston and Dr. Jeffrey Singer argue that rising use of AI chatbots for health questions, driven largely by cost and triage decisions per Pew Research data, is a net positive that state licensing restrictions shouldn't curb. They point to a JAMA Network Open trial finding GPT-4 alone outperformed physicians on diagnostic reasoning tasks, and warn that state proposals in Illinois and New York requiring physician oversight of AI health tools risk treating decades-old licensing frameworks as a substitute for updating them. The authors frame the debate partly in free-speech terms, arguing AI restrictions limit access to information rather than protect patients.

Democrats splinter on how to pace the AI race — Sen. Bernie Sanders and Rep. Greg Casar introduced legislation to ban superintelligence and pause advanced AI development until a new Cabinet-level Department of Artificial Intelligence sets guardrails, with Sanders warning against being "a father and grandfather who is asleep at the wheel." The same day, Sens. Elizabeth Warren, Chuck Schumer, Andy Kim, and Elissa Slotkin pushed a different approach, focused on tightening export controls to slow China's AI progress rather than pausing U.S. development, while Rep. Ro Khanna convened a hearing pushing for a U.S.-China agreement to pace AI jointly. Even within the party's progressive wing there's no consensus, with Warren backing a pause but stopping short of Sanders' superintelligence ban, and potential 2028 contenders like Rahm Emanuel questioning what a pause accomplishes.

OpenAI calls for international technical standards to pace frontier AI development — OpenAI published a proposal urging the US to lead an international effort on technical standards for frontier AI, building on existing AI safety institutes in countries including the UK, Japan, Germany, and India through the Center for AI Standards and Innovation. The framework calls for common measurement protocols on things like recursive self-improvement progress and incident reporting, explicitly pointing to OpenAI's own recently launched misalignment disclosure framework as a model, and draws comparisons to how aviation and financial regulators built shared technical standards. The company frames the goal as keeping alignment research ahead of capability gains rather than slowing development to a fixed pace.

The trillion-dollar AI safety paradox — Axios reports that the financial incentives driving AI development, with Goldman Sachs projecting $7 trillion in AI scaling spend over the next five years, are increasingly at odds with safety priorities, even as AI models grow less transparent about their own reasoning. Former OpenAI researcher Daniel Kokotajlo said the labs are stuck in "this continuous process of conflicted feelings" between safety and progress, while Palantir's Akshay Krishnaswamy argued human oversight can still keep pace with scaling. Trump, for his part, has shown no appetite for a slowdown, posting that "Whoever wins AI, WINS!"

AI leaders want a slowdown. What should healthcare do? — Former Anthropic and OpenAI researcher Jacob Coxon's public resignation this month, warning that AI developers believe the technology "could kill us all by the end of the decade," has healthcare leaders weighing what it means for the sector's heavy AI investment. CHAI CEO Dr. Brian Anderson said it's appropriate for health systems to share that added concern as they deploy agentic tools, while UTMB's Dr. Salim Hayek said the sector's slower, more scrutinized adoption process makes a near-term slowdown in healthcare AI use unlikely, since UTMB isn't using fully autonomous agents. UCSF's Dr. Robert Wachter warned that worsening public sentiment or tighter regulation could threaten the financial viability of healthcare AI vendors even though currently deployed tools have already shown real benefit, calling it a risk of "throwing the baby out with the bathwater."

Tech leaders ask the UN Security Council to help rein in AI — Anthropic's Dario Amodei and OpenAI's Sam Altman told the UN Security Council that AI poses a genuine risk to humanity if mismanaged and that the world needs shared safeguards, with Amodei calling for international agreement on barring AI from helping build biological weapons and for common standards to test for loss of control. The U.S., represented by White House Science Adviser Michael Kratsios, rejected calls for new global governance structures, while the U.K. said it will make AI safety standards central to its G20 presidency next year and China's ambassador warned against countries "banding together" against each other on AI policy. The meeting came as Ukrainian President Volodymyr Zelenskyy warned that AI may soon make battlefield targeting decisions without human control, and a University of Michigan researcher noted a Russian drone last month appeared to autonomously select a target that killed three Ukrainians.

Chatbots are still falling short in mental health conversations with kids — A new study from AI testing company Vals AI, run with mental health professionals across more than 600 simulated teen mental health conversations, found nine major chatbots made a safety error, like offering an unsolicited diagnosis or agreeing to help conceal a suicide attempt, in nearly 30% of conversations, usually after several back-and-forth exchanges rather than immediately. Failure rates varied widely: Google's Gemini failed nearly 32% of tests, Claude nearly 17%, ChatGPT almost 14%, and Meta's Muse Spark had the lowest failure rate at 9.7%. The findings echo an August study from Transluce that found chatbots have grown less likely to encourage suicidal behavior outright but still reinforce risky or delusional conversations over time, as OpenAI, Google, and other companies face lawsuits from families alleging their chatbots contributed to a user's death by suicide.

Big tech says AI can cure cancer. Oncologists say it's more complicated — A Guardian feature examines the gap between AI executives' claims of curing cancer within a decade and the more measured view of clinicians working in the field. Researchers described real progress, including machine-learning models that detected pancreatic tumors in scans read as normal years before diagnosis and an EU consortium using AI to speed up brain tumor treatment decisions, but pushed back on the idea that computation alone can crack a disease this variable. Future of Life Institute physician Emilia Javorsky noted that 13 years into AI-driven drug discovery, not a single drug has cleared full FDA approval, reimbursement, and clinical adoption, arguing that data, regulation, and incentives, not intelligence, are the real bottleneck.

Anthropic says its AI helped discover a new biological tool — Anthropic says its Claude AI combed through massive genetic databases and spotted a previously unrecognized biological system, one that works something like CRISPR, the gene-editing tool that's revolutionized medicine and research. The system centers on an enzyme that copies genetic information from RNA into DNA, something scientists already knew existed, but Anthropic says Claude was the first to notice the fuller picture: a repeating DNA pattern and an extra protein whose job isn't yet understood, both of which suggest the system does something more sophisticated than previously realized. It's the first announced result out of the physical lab Anthropic quietly opened in the Bay Area to test AI-generated ideas on real biology, not just on computers.

How Pfizer says AI is speeding clinical trial recruitment — In a company-published piece, Pfizer VP Melinda Rottas describes how the company is using AI across clinical trial development, from drafting protocol documents and flagging participant burden in trial design to identifying potential trial sites and matching eligible patients faster than the traditional manual chart review and phone call process. Rottas emphasized AI keeps a "human in the loop," with a principal investigator still determining eligibility and patients still completing informed consent, and said the approach could also expand access by identifying where underrepresented populations actually receive care rather than defaulting to sites with strong past enrollment.

Verily says its AI cuts a huge time sink in cancer research — Medical researchers spend enormous amounts of time manually digging details out of doctors' notes and lab reports to build datasets for studies, since that information usually isn't stored in a clean, searchable format. In sponsored content on Healthcare Dive, Verily Health says it built an AI tool that automates that work for leukemia patient records, working with UCHealth, the University of Colorado Anschutz, and RefinedScience to test it. A basic AI model only got the details right about 72% of the time, but Verily says its version, built with medical knowledge baked in rather than treating notes as plain text, got that above 95% and could shrink a task that took 1,200 hours down to about 40.

Enterprise AI needs a new change management model — Aivar Innovations CEO Kousik Rajendran argues that traditional, linear change management fails for AI agents because they behave unpredictably and evolve after deployment. He proposes a cyclical framework called TRACE: build trust without over- or under-relying on the tool, route decisions by risk and complexity, drive adoption through stakeholder co-creation rather than one-off training, calibrate against business outcomes instead of usage metrics, and design for continuous evolution as models change. Deloitte found 74% of companies plan to deploy agentic AI within two years, but only 21% report having a mature governance model in place.

The AI kill switch is becoming a design requirement at health systems — Health system leaders say a clear shutdown mechanism now has to be built into an AI tool's design from day one, not bolted on after something goes wrong. Executives at Parkview Health, Brigham and Women's, and Mayo Clinic described setting failure metrics, support end-dates, and accountability lines before launch, while HealthPartners said it won't let AI agents make clinical decisions autonomously and Seattle Children's uses a scoring system based on reach, human oversight, reversibility, and potential harm to set how much scrutiny a pilot gets. Evry Health's CEO added that observability tooling has surfaced unknown vulnerabilities in roughly one in five AI-generated code pull requests.

👉 Sept. 29, 12:00 PM EDT — AHIP Webinar: Unlocking AI for Healthcare Payer Transformation — Virtual

👉 Sept. 29 — MAHA Summit 2026 — In-person, Waldorf Astoria, Washington, D.C.; includes an AI & Healthcare track with speakers from Anthropic and health tech companies

👉 Sept. 30, 12:00–1:00 PM CDT — Behind the Curtain: How LLMs Are Actually Built for Healthcare and What to Know Before You Deploy — Webinar, Becker's Hospital Review

👉 Sept. 30, 1:00–2:00 PM ET — Healthcare Executives Weigh In: AI Readiness and the Growing Risk of Cybersecurity Threats — Webinar, MedCity News/Cotiviti

👉 Oct. 14 — AI Day 2026 — Hybrid, Georgia Room, NYC; free virtual pass available

👉 Oct. 14–16 — AIxPH 2026: 1st Annual Conference on Artificial Intelligence and Public Health — Baltimore, MD

👉 Oct. 15, 1:00 PM ET — How specialty practices are using AI to turn patient access into capacity and growth — Webinar, Becker's Hospital Review

👉 Oct. 22–23 — HIMSS AI in Healthcare Forum — San Diego, CA

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

👉 Dec. 16, 11:00 AM ET — The Next Phase of AI Adoption: Governance, Ethics, and Accountability — Virtual