The Bandwidth: Decoding Health + AI

Last week, the pope weighed in on AI. This week, it was Senator Bernie Sanders who captured headlines by proposing in a New York Times essay that the federal government should take a 50-percent equity stake in the largest AI companies to create a sovereign wealth fund for the American public. And he wasn't the only government (or former government official) to make headlines this week. This week's section on government and policy developments is by far the longest.

Sanders' essay is worth taking seriously, whatever you make of the proposal. His core claim is that AI is built on collective human knowledge, and the public should share in the wealth it generates.

He's not the only one who thinks this. Sen. Elizabeth Warren urged in a Time magazine op-ed for an overhaul of the tax code, warning that the current payroll tax model incentivizes companies to drop employees and replace them with AI, especially as AI can increasingly replicate human brainpower. She proposed raising capital gains taxes and minimum taxes for major corporations to fund social safety programs for people whose jobs could be replaced by AI. She also called for an excise tax on data centers.

It’s worth noting that these policy proposals aren’t necessarily at odds with the frontier labs that would be subject to them. As the Sanders op-ed notes, and reported yesterday by NOTUS, the leaders of the labs themselves (and in some cases, the chipmakers) have expressed support for these concepts in one form or another.

Beneath the Headlines

Underneath the headline debate about ownership, wealth, and redistribution, a subtler set of economic signals crossed my radar this week about how AI may reshape jobs and the economy.

A Federal Reserve Bank of New York working paper found that remote work, not artificial intelligence, explains 64 percent of the rise in unemployment among young college graduates since the pandemic. AI has become the labor market's favored scapegoat. But some of that attribution may not be warranted. It’s too soon to tell. Indeed, there's an open question as to whether the layoffs rippling across the tech industry are due to over-hiring during the pandemic or to AI.

That doesn't mean AI has no economic effect. As one strategist put it this week, AI may be acting as a headwind to new hiring rather than a driver of mass layoffs – as employers consider whether a job (or job function) can be performed by AI instead of a human. Under this theory, the result may be the same: ultimately, payrolls stay soft, with fewer workers dismissed, but also fewer positions being opened.

A Familiar Pattern

History offers a useful frame for what structural AI-driven job displacement (or churn) could look like. Two recent NBER working papers (here and here) map what happens when major technologies enter the labor market. New types of work emerge that didn’t previously exist. Early in that cycle, that new work commands a substantial wage premium for those who can perform it. Over time, as skills diffuse through the workforce, the wage premium fades. The pattern has held across major technological transitions, and AI could follow the same arc.

There are early signals that the new-work cycle is already running in health care. Major health insurers — Elevance Health, UnitedHealth Group, MetLife — are now actively recruiting for a role that barely existed two years ago: the algorithmic auditor, variously titled AI governance manager, AI ethics officer, or model validation specialist, with salaries ranging from $115,000 to more than $270,000 (h/t to one of my reader’s for sharing the article by Ron Shinkman: A New Position At Insurance Companies: The AI Auditor). The job description is still being written; what's clear is that it demands a combination of technical fluency and regulatory knowledge that may be scarce and early in its diffusion.

Back to the Beginning

But the main question remains, who captures the early gains? The NBER research suggests that technology-related work follows a pattern when it first emerges. If so, that means the economic benefits will accrue first, and most heavily, to workers who may already be relatively advantaged - those more highly educated, and clustered in dense areas.

That pours fuel on demographic and political divides that are already inflamed. It also resurfaces the thrust of Sanders' proposal: a wealth fund isn't only about compensating society for the collective knowledge that AI is built upon. It’s also about how the gains are captured (or distributed) before that diffusion cycle and market forces run their course.

Enzo Health launches an "AI-native" EHR for home health — The startup's agentic system claims to automate the full patient episode from referral to billing, with company-reported figures of intake dropping from 70 minutes to five and charting time falling roughly 75 percent. The metrics come entirely from Enzo and a single early adopter, so the launch reads more as a marker of where agentic EHR ambitions are heading than as validated performance.

Anthropic files confidentially for an IPO — The company submitted draft S-1 papers to the SEC on Monday, days after a $65 billion Series H round valued it at $965 billion and pushed its valuation past OpenAI's. The filing also puts Anthropic's unusual governance before public investors for the first time: as a Delaware public-benefit corporation, its board must weigh shareholder returns against its stated mission, and its independent Long-Term Benefit Trust holds a special share class letting it elect a majority of the board through trustees barred from any financial stake in the company. Whether public-market investors accept that arrangement is an open question.

The AI "arms race" in medical billing — In a STAT First Opinion essay, Machinify CMO Darshak Sanghavi argues that the most consequential AI in U.S. health care isn't clinical but financial: hospital "revenue cycle" tools and insurer "program integrity" tools now fight over the same record, with patients absorbing the friction. He points to a Blue Cross Blue Shield analysis linking fast AI-coding adoption to a jump in maternal-hemorrhage coding unmatched by any rise in treatment, and proposes a unified adjudication engine.

Alphabet seeks up to $80 billion for its AI buildout — The company plans to raise the equity — including a $10 billion private investment from Berkshire Hathaway — to scale AI infrastructure and compute, on top of its recent record 100-year bond issuance. The move underscores how the major hyperscalers, on track to spend more than $750 billion on AI this year, are now tapping outside capital despite historically strong cash flows.

Mayo Clinic and Microsoft to build a healthcare-specific frontier model — The two organizations announced a collaboration to develop a frontier AI model trained on Mayo's de-identified clinical data and longitudinal insights, with Mayo owning the model and Microsoft distributing it through Azure Foundry APIs. It's an announcement rather than a shipped product — the model will first be tested inside Mayo's own clinical environment.

Microsoft's Build 2026 leans hard into AI agents — At its developer conference, Microsoft unveiled Scout, an always-on "Autopilot" agent that acts across Teams, Outlook, and files without being prompted; a new generation of in-house MAI models spanning reasoning, code, image, voice, and transcription; and Majorana 2, a next-gen quantum chip the company says is 1,000x more reliable than its predecessor. The announcements signal how aggressively it is rebuilding Windows and Microsoft 365 around AI agents, with the figures and capabilities all company-stated.

Meta enters the enterprise AI race with a "Business Agent" — At its Conversations conference, Meta unveiled an agent that can act on businesses' behalf across WhatsApp, Messenger, and Instagram — booking appointments, processing payments, and qualifying leads — plus a broader platform connecting to outside systems like Shopify and Zendesk. The launch lands days after hackers tricked Meta's AI support chatbot into surrendering access to high-profile Instagram accounts, a reminder of the security exposure that comes with letting agents take real actions.

Hospitalists are already using AI at the bedside — mostly without support — A JMIR commentary from University of Colorado hospitalists notes that roughly two-thirds of hospitalists now use AI in clinical work, mostly LLM tools and largely without health-system integration, often to generate differential diagnoses and management options. The authors argue adoption is outpacing training and governance, and that — as with the EHR — impact will turn on how well the tools are embedded in clinical workflows, not on use rates alone.

Patients are building their own AI record of doctor visits — and health systems are uneasy — As ambient scribes become standard on the clinical side, a wave of consumer apps — including VisitRecall, Advoca Health, and Kin Health — are giving patients their own AI-generated summaries and action items from encounters, but when health data moves outside provider systems, HIPAA protections no longer apply and the fine print governs. Health system leaders are increasingly uncomfortable with two AI systems interpreting the same visit, and some hospitals have moved to forbid patient recording altogether.

Florida sues OpenAI and Sam Altman over AI harms — Attorney General James Uthmeier's 83-page complaint, the first state lawsuit against OpenAI, alleges the company knowingly shipped an unsafe product and seeks to hold Altman personally liable under a public-nuisance theory. The suit follows an April criminal investigation tied to a campus shooting and lands as states move to fill the AI-oversight gap.

Trump signs AI cybersecurity and "covered frontier model" executive order — The June 2 order directs federal agencies to harden critical infrastructure against AI-enabled threats, stands up a voluntary AI cybersecurity clearinghouse with industry, and lets developers give the government up to 30 days of early access to designated "covered frontier models" before broader release — while explicitly ruling out any mandatory licensing of AI models. It names rural hospitals among the critical-infrastructure operators slated for federal cyber tools and frontier-model access, placing under-resourced health systems inside the order's security perimeter.

Former CISA director: The Trump AI order is a first step — Congress needs to finish the job — Jen Easterly, former CISA director and current RSAC CEO, argues in a New York Times guest essay that the administration's new AI cybersecurity executive order represents a meaningful shift from its hands-off stance, but a voluntary 30-day review framework isn't durable enough — and Congress needs to impose binding obligations on the most advanced models. She frames frontier AI cybercapabilities as dual-use technologies analogous to pathogens or nuclear materials, noting that hospitals, financial systems, and power grids are among the institutions least equipped to defend themselves if those models are deployed carelessly.

House lawmakers unveil bipartisan AI bill that would preempt state laws — Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a 269-page discussion draft that would require top AI developers to disclose and address the safety and security risks of their most advanced models, fund a NIST AI standards office at $300 million over three years, and mandate third-party audits — while preempting state AI laws for an initial three-year period. The preemption provision has drawn immediate criticism from AI safety advocates and lawmakers in both parties who warn it would eliminate a key accountability backstop, and Politico describes the draft as Congress' last realistic shot at federal AI legislation before the midterm elections.

Utah's AI prescription pilot triggers a regulatory standoff — and a transparency red flag — Utah's Medical Licensing Board was reprimanded by the state attorney general for process violations after publicly challenging the state's autonomous prescription-renewal pilot without holding a formal meeting, while state officials maintain the Doctronic pilot is fully authorized. A Nature Medicine transparency audit found that Doctronic declined to share data from its own study — the one that persuaded Utah regulators to authorize the pilot — despite a prior commitment to make it available upon request.

FDA accepts an AI liver-toxicity prediction tool into its drug development pilot program — The FDA's Center for Drug Evaluation and Research has accepted a letter of intent for an AI-driven digital liver model that assesses drug-induced liver injury risk in small-molecule compounds by comparing their chemical structures to drugs with known safety profiles. The tool enters the agency's ISTAND pilot program — the first stage of a multi-step qualification process that could allow drugmakers to use it in regulatory submissions, with the potential to reduce animal testing and improve early safety decisions before human trials begin.

Inside the administration's push for autonomous AI doctors — A Washington Post investigation details how Trump officials — including DOGE leader and HHS adviser Amy Gleason and CMS administrator Mehmet Oz — are quietly building regulatory pathways for AI chatbots to diagnose illness and prescribe medications with limited physician oversight, backed by $50 million in cardiovascular-care AI grants and a new Medicaid reimbursement track for AI wellness apps. The medical establishment is pushing back: a Nature Medicine study found AI chatbots correctly diagnosed conditions just 34 percent of the time, physicians warn the technology is far from ready to practice independently, and Pennsylvania has filed suit against Character.AI for presenting its chatbot as a licensed medical professional.

What the pope's AI encyclical means for Catholic hospitals — STAT's Brittany Trang reports that health systems are weighing how the new papal encyclical on AI translates into practice, with Providence's ethics lead expecting it to sharpen Catholic and broader thinking on human dignity and data stewardship. Legal analysts say it reframes AI governance as a duty rather than an obstacle and cautions that a chatbot feeling human and friendly doesn't make it the kind of trusted relationship medicine depends on.

Patient-facing AI cancer information is scarce and low quality — A Penn/Abramson team screening 320 Google and YouTube results found only a third of the relevant webpages and under a quarter of videos were high quality, with median readability at college level against the 6–8th grade standard for consumer health. Just 15 percent of webpages mentioned the risk of AI hallucinations.

Distressed students are turning to unregulated AI for mental health — In a survey of 675 college students from the 2024–2025 Healthy Minds Study, about 18 percent reported using AI for mental health, and severe depression, severe anxiety, and suicidality each roughly doubled the odds of doing so. The cross-sectional design can't establish cause, but the pattern points the highest-risk users toward the tools with the fewest safeguards.

A fifth of US youth have used AI chatbots for mental health advice — A nationally representative JAMA Pediatrics survey found 19.2 percent of adolescents and young adults reported using AI chatbots for mental health advice in 2025 — up from one in eight a year earlier — and nearly two-thirds told no one they were doing so. Most users rated the advice helpful, though the authors caution that perceived helpfulness may reflect chatbots' tendency toward sycophancy rather than the quality of guidance.

Clinician AI use jumps, but providers fear deskilling — A Wolters Kluwer Health survey of more than 350 clinicians found roughly three-quarters of doctors and 70 percent of nurses now use AI at least weekly, up sharply from a year ago, even as 74 percent named loss of critical-thinking and decision-making skills among AI's greatest risks. Just 27 percent said they knew how their organization governs AI use.

A taxonomy for cutting through "world model" hype — Stanford AI pioneer and World Labs founder Fei-Fei Li distinguishes three types of AI systems now marketed as "world models": renderers (visually plausible outputs for human eyes), simulators (physically accurate representations that hold up under interaction), and planners (systems that take actions toward a goal). Published on Li's Substack and reflecting her lab's research agenda, the essay offers a useful analytical frame — including a candid acknowledgment that the gap between robotic demo reels and real-world deployment in settings like operating rooms remains substantial.

Your AI assistant's inbox is an attack surface — SafeBreach Labs researchers recently demonstrated how hidden malicious instructions embedded in a WhatsApp notification could direct Google Gemini to silently exfiltrate data — bypassing Google's own layered defenses, which the company documents in its Workspace admin guidance. The attack surface is architectural, not app-specific: any notification an AI assistant reads is a potential delivery channel, and the practical countermeasure is permission hygiene — audit what your AI assistant can access and disable anything not actively used.

👉 Jun. 8–10 — Fortune Brainstorm Tech, Aspen, CO

👉 Jun. 9–18, 2026 — Google Cloud-Hosted Startup School: Agentic AI — Virtual

👉 Jun. 11, — CHAI Webinar: From Pilot to Production: How Healthcare Leaders Are Measuring AI ROI — Virtual

👉 Jun. 15–18 — Databricks Data + AI Summit 2026, San Francisco + Virtual

👉 June 24 — CTRC 14th Annual Digital Health Summit: Bridging the Gap — Digital Health and AI for Smarter Care, Stronger Teams, and Sustainable Healthcare — Virtual

👉 Jul. 7–10 — AI for Good Global Summit 2026, Geneva, Switzerland

👉 Aug. 4–6 — Ai4, Las Vegas

👉 Oct. 13–16, 2026 — AIxPH 2026: 1st Annual Conference on Artificial Intelligence and Public Health, Baltimore, MD

And finally, if you like what you are reading, please share this newsletter with your networks and encourage them to sign up. ✍️ 🆙 And/or, give me a shout out on LinkedIn.

Till next time,

BC