The Bandwidth: Decoding Health + AI

Dario Amodei, Anthropic's CEO and a prolific author of these sweeping policy essays, published another this month. This one closer to a call to arms, arguing for more oversight and binding regulation of AI. That's not where industry usually plants its flag, but it's in character for Amodei and Anthropic.

His through-line is pace. Government moves slowly, and he grants that deliberate policymaking is often a feature rather than a flaw. The state holds grave powers we don't want used hastily. But AI, he argues, is the case where the speed mismatch turns dangerous, and where the machinery has to move faster than it was built to.

He covers several fronts. I'll stay focused on health.

He lands on two poles: a downside of biosecurity, with frontier models lowering the barrier to biological weapons. And an upside of discovery, with AI accelerating drug development in ways that could reshape medicine — paired with a worry that the FDA can't keep pace with how AI is changing the work.

The machinery is slow. That isn't the whole story.

Having spent time inside the government's health apparatus, I'll grant the premise. It moves slowly, sometimes painfully so. But slow is not the same as stuck, and there are ways to thread the needle.

Start by parsing two things that get run together. Improving the safety profile of frontier models in high-risk areas (e.g., cyber, biological threats, loss of control) is one set of problems. How the FDA adapts to AI in drug discovery, and review is another. Amodei himself treats them as distinct, and the health side is better handled on its own terms than folded into a frontier-model debate.

The experimentation is already underway.

While Amodei is urging the government machinery to move faster, that doesn’t mean that policy work isn’t underway.

The White House's AI strategy last July called for a "try-first" culture of regulatory sandboxes and AI Centers of Excellence, naming healthcare as a notably slow adopter.

And states have moved faster than Washington: Utah's Office of Artificial Intelligence Policy has used its sandbox to grant health-care relief agreements, including an AI prescription-renewal service for chronic conditions and a mental-health platform. Other states (e.g. Texas and Delaware) have followed suit. The logic is evidence-first: relax the rules in a confined environment, test under supervision, and bring the results back to legislators.

What's missing is a forcing function.

A sandbox tests ideas; it doesn't make them permanent, and it doesn't compel anyone to act on what's learned. For that you need a mechanism that brings Congress back to the table on a schedule to revisit how the agency applies AI and what needs adjusting.

We already do this elsewhere. The FDA's prescription-drug user fees are reauthorized every five years, a cycle that forces the agency, industry, and lawmakers to take stock. A similar reauthorization of the FDA's AI authorities would do the same. And if five years feels too slow for a field moving this fast, the National Defense Authorization Act shows we know how to reauthorize a sprawling enterprise every year. Tools that prove out in a state sandbox could then graduate to durable authority, written in during the next cycle.

Reauthorization is no cure-all, though, and it's worth being honest about how it fails. The same recurring deadlines that force action can become hostages to politics. This very week offers the cautionary tale: Section 702 of the Foreign Intelligence Surveillance Act (FISA), kept on a short reauthorization leash, sits days from lapsing as its renewal stalls. Unclear if a reauthorization vehicle for health AI could be built to resist that kind of fate.

Nonetheless, there are paths forward to address Amodei’s anxieties. States are testing, the White House has blessed concepts that allow regulatory experimentation. What's been missing is the mechanism to make the lessons stick: a way to let the FDA (or other health agencies) move faster on AI without asking it to move recklessly. Speeding up the machinery where we can and keeping the deliberation where it counts. 

The AI vendors health systems are actually buying — Becker's reviewed 2026 enterprise contracts, KLAS rankings, and outcome data to name 10 AI vendors moving health systems from pilots to systemwide infrastructure, spanning ambient documentation (Abridge, Microsoft's Dragon Copilot, Ambience), radiology triage (Aidoc), sepsis detection (Bayesian), capacity management (LeanTaaS), and clinical decision support (OpenEvidence). The throughline is that deals once structured as cautious pilots are now skipping straight to enterprise rollout, with Abridge reporting 250-plus health system customers and OpenEvidence reaching a $12 billion valuation.

OpenAI files confidentially for an IPO — OpenAI has confidentially filed to go public, joining rival Anthropic and SpaceX in a wave of AI mega-listings that could total hundreds of billions and test the market's appetite for AI firms, even after tech stocks sold off last week. The move carries a health angle: OpenAI has expanded well beyond ChatGPT into clinician-facing tools and health programs, and a public listing would expose the revenue pressures already reshaping the product, with Reuters, citing the Financial Times, reporting OpenAI is redesigning ChatGPT into an enterprise-focused "superapp" as business customers climb toward half of its revenue.

Apple rebuilds Siri around Apple Intelligence — Apple introduced "Siri AI," a ground-up rebuild of its assistant powered by next-generation Apple Intelligence, with personal-context understanding, onscreen awareness, and expanded Visual Intelligence, running on a mix of on-device models and server-side Private Cloud Compute that Apple says keeps personal data inaccessible even to Apple. For health, the architecture is the story: as a privacy-first assistant gains the ability to read across a user's messages, emails, and photos, the on-device boundary becomes what determines whether sensitive personal data stays contained.

Anthropic launches Fable 5, its most capable model, with biology held back — Anthropic released Claude Fable 5, a top-tier "Mythos-class" model the company says leads most capability benchmarks, alongside a restricted variant, Mythos 5, that it reports accelerated internal drug-design work roughly tenfold and generated novel molecular-biology hypotheses its scientists preferred about 80% of the time in blinded tests. Citing dual-use risk, Anthropic routes most biology and chemistry queries on the publicly available model to a less-capable fallback and is limiting full life-sciences access to a forthcoming trusted-access program for vetted researchers.

16 economists on AI and work: productivity yes, jobs uncertain — A WSJ survey of 16 economists found near-consensus that AI will lift productivity but a sharp split on jobs, with eight expecting no net change, five a net loss, and two net growth over the next five years. Several singled out healthcare as a prime beneficiary where AI could break expert-scarcity bottlenecks — David Autor's example was a nurse practitioner with diagnostic AI taking on cases that would once have needed a doctor.

A looming AI price war could cut what health systems pay — The WSJ reports OpenAI is weighing steep cuts to its token prices in anticipation of similar moves by Anthropic, as enterprise customers balk at AI costs that CEO Sam Altman called "a huge issue." For health systems budgeting to run AI at scale, a price war between the two dominant providers could ease the per-query economics — even as it underscores how interchangeable the underlying tools have become.

Nvidia teams with Abridge on a healthcare-tuned AI model — Nvidia is partnering with Abridge, the ambient-listening startup whose AI transcribes doctor-patient visits, to train a model purpose-built for clinical conversations on Nvidia's open Nemotron line, with Abridge feeding in its de-identified clinical data to sharpen documentation and, eventually, decision support. The deal is part of a broader land-grab — Microsoft and Mayo Clinic struck a similar healthcare-model partnership last week — and a bet that smaller, purpose-built open models can undercut proprietary ones on cost while running on a vendor's own hardware.

Doctors are quietly turning A.I. into a bedside consultant — More than half of U.S. physicians now regularly use OpenEvidence, the journal-trained clinical chatbot, which fielded 30 million questions last month, adoption that often spread as shadow A.I. before systems like Mount Sinai moved to formalize it through the EHR. The pull is strongest where specialist access is thin: a Fairbanks, Alaska physician told the Times the app is "like having a bunch of specialists in your pocket," and a self-described skeptic in rural Mississippi now treats it as a diagnostic thought partner.

Talkspace launches an AI mental-health tool with a therapist backstop — Talkspace rolled out Tee, a subscription AI tool the company says was built by mental health experts to help adults work through stress, anxiety, and relationship issues, with HIPAA-grade privacy and the ability to flag suicide, violence, and abuse risk. Its distinguishing design choice is human oversight: therapists can step in when the tool surfaces a concern, a hedge against the unsupervised-chatbot risks drawing scrutiny elsewhere in this space.

A doctor's stoplight guide to asking AI your medical questions — Writing for Harvard Health, Beth Israel Deaconess internist and AI researcher Adam Rodman sorts patient chatbot use into green (understanding general information, decoding lab results, prepping for a visit), yellow (exploring new symptoms — best done by having the bot interview you, since these tools tend to tell you what they sense you want to hear), and red (treatment decisions like whether to start a medication, which he says to simply avoid). The framework is a level-headed corrective to the "AI doctor" framing: a chatbot can inform the conversation with a clinician, not stand in for it.

The White House revives its push to preempt state AI laws — Axios reports the administration is negotiating with Sen. Marsha Blackburn to pair federal preemption of some state AI laws with measures like the Kids Online Safety Act and the NO FAKES Act, a narrower package than the blanket preemption that drew heavy pushback last time. Since most AI rulemaking that touches health has so far happened at the state level, where preemption lands would reset the regulatory baseline health systems are planning around.

HHS wants AI to help evaluate addiction and mental health programs — A new HHS request for information asks patients, providers, and policymakers how the department could use AI and advanced analytics to measure the effectiveness of substance use and mental health programs on a real-time or continuing basis. The RFI lands amid a broader federal AI push: HHS led the government in AI use in FY2025, CMS leadership has said it wants to become an "AI-first agency," and regulators are still wrestling with how to oversee mental health chatbots and Utah's experiment with AI prescribers.

A doctor argues teen AI dependency is outpacing policy — Writing in Tech Policy Press, physician-researcher Caroline Figueroa argues that lawmakers fixated on acute harms — the Senate Judiciary Committee just advanced the GUARD Act to verify ages and bar AI companions for minors — are missing the slower risk of teens substituting chatbots for human emotional support, citing research that nearly one in five U.S. young people now turn to AI for mental health advice. Her prescription isn't a ban but design mandates: chatbots built to avoid overconfident social advice and to actively steer minors toward parents, teachers, and clinicians.

House appropriators hand HHS a long AI to-do list — A House Appropriations Committee report urges HHS to study non-clinical AI, lean on accreditation programs like the Joint Commission's and URAC's new AI certifications to keep tools responsible, and finally reform how Medicare pays providers who use AI, while pressing CMS to deploy AI against fraud. Notably, the same lawmakers drew a line against CMS's controversial push to use AI-driven prior authorization to curb overutilization in traditional Medicare.

Congress moves to kill CMS's AI prior-authorization pilot — The House Appropriations Committee also unanimously added an amendment to its 2027 HHS spending bill barring CMS from funding the WISeR model — the AI-driven prior-authorization pilot for traditional Medicare — or any program that would extend prior auth to it. In parallel, Sen. Ron Wyden introduced a Congressional Review Act resolution to overturn the underlying CMS rule, arguing it should have gone to Congress before rollout, as provider groups including the AHA and AMA keep up the pushback.

The AMA draws its line: AI assists, doesn't decide — At its annual House of Delegates meeting, the AMA adopted policies declaring that AI should support rather than replace physician judgment and pledging to push for rules that keep humans in the loop wherever AI touches care, explicitly opposing autonomous or semiautonomous systems as substitutes for physician review in insurance coverage decisions. It also wants transparency mandates for AI-driven prior authorization — disclosure of the clinical logic and data behind a denial — plus regular audits of the underlying tools.

AI use cases are surging across HHS, led by a 148% jump at the FDA — A Bipartisan Policy Center analysis of HHS's own AI inventory found use cases climbing across every agency between FY2024 and FY2025, with the FDA up 148%, the CDC 87%, CMS 78%, and NIH 51% — though the report cautions some of that jump may reflect newly required reporting rather than newly deployed tools. With most use cases still in pre-deployment, the report reads the trend as a leading indicator of more to come.

Patient records are flowing more freely, and leaking to lawyers — Epic and three health systems are suing data-sharing platform Health Gorilla for allegedly letting third parties posing as providers pull more than 300,000 medical files, with one company admitting it took records under false pretenses and sold them to class-action lawyers. The case exposes a governance gap at the heart of interoperability: as AI makes mining personal medical data easier, the networks built to share records still can't reliably verify who is requesting them.

A framework for judging AI by mission, not just cost — Writing in npj Digital Medicine, two physician-researchers propose "Total Mission Value," a governance framework that pushes health systems to weigh an AI tool not only on cost but across five ethically grounded domains — patient care, staff experience, operations, economics, and education and research — with equity running through all of them. Their sharper claim: an organization that can't even build a mission-aligned scorecard for a given AI tool may not be ready to deploy it.

People are using AI chatbots to fill the gaps health care leaves open — A Nature commentary unpacks a new Nature Health analysis of roughly 618,000 health conversations with Microsoft's Copilot, which found that about a quarter were personal questions, with queries about symptoms and emotional well-being spiking at night and on mobile devices, when conventional care is hardest to reach. The piece notes the structural gap: these general-purpose tools face none of the accuracy, oversight, or accountability requirements that govern medical devices, even as more people lean on them for health decisions.

Clinicians say AI is buying them time and capacity, Philips survey finds — Philips's Future Health Index 2026, drawn from more than 2,000 clinicians across 10 countries, reports that over a third say AI lets them see more patients (a median of five more per week) and nearly half claim annual time savings of at least 132 hours. The same survey flags an adoption gap that tempers the optimism: 77% report inadequate or inconsistent AI training, and 72% reach for personal AI tools when their organization's options fall short.

An AI reads brain-tumor slides in minutes instead of weeks — Researchers built Hetairos, an AI that identifies 102 brain-tumor subtypes from the cheap, standard stained-tissue slides used in hospitals everywhere, matching diagnoses that normally require a specialized molecular test costing around €400 and taking two weeks — Hetairos did it in about 12 minutes for €1–2 on ordinary computers. Tested on more than 11,000 slides across four continents, it was confident and accurate on most cases and beat five expert neuropathologists working from slides alone, though its authors frame it as a triage tool (its name means "companion" in Greek) that flags which cases still need the fuller molecular workup rather than replacing it.

Perplexity says its AI agent turns workers from operators into supervisors — In a study of its own "Computer" agent run with Harvard Business School researchers, Perplexity reports that delegating multi-step tasks to the agent cut estimated task time by roughly 87% and cost by 94% on matched work, while pushing users to attempt more complex jobs and cross into fields outside their own expertise, healthcare among them.

Health-policy experts are wary of letting AI prescribe — A Cornell Health Policy Center survey of 63 health-policy scholars found broad skepticism toward AI medical licensure: 83% were uncertain about or opposed to a bill, the Healthy Technology Act, that would let FDA-authorized AI prescribe medications, citing thin clinical evidence. On regulation, the panel leaned toward an approach emphasizing post-market surveillance over premarket clearance (46% in favor), though several warned regulators may be ill-equipped for the volume and pace of algorithms entering clinical use.

A rigorous trial finds AI chatbots no better than a good brochure for vaccine hesitancy — In a randomized trial of 1,297 parents of unvaccinated children, brief conversations with a GPT-4o chatbot raised HPV-vaccination intent immediately, but the effect faded by 45 days and never exceeded that of standard government public health materials, whose more modest gains were the ones that persisted. And no intervention — chatbot or brochure — moved actual self-reported vaccination uptake, a reminder that shifting stated intent isn't the same as changing behavior.

Why drugs and AI chatbots won't cure health care's cost disease — In a Health Affairs Forefront analysis, Soleil Shah, Suhas Gondi, and Niyum Gandhi argue that the popular "goods will save us" thesis — novel drugs averting care, AI bots replacing clinicians — misses how cost actually works: neither reduces the underlying labor, time, and supplies, and an autonomous AI doctor would only displace the cheapest encounters (primary care is just 5–7% of spending) while leaving surgery and other big-ticket care untouched. Their alternative is augmentation over substitution — deploy AI to lift clinician and administrative productivity, from documentation to prior auth to throughput, and reinvest the reclaimed hours rather than trying to productize care away.

CIOs are on the hook for AI they don't control — IBM's new survey of 2,000 technology executives found two-thirds say they're held accountable for AI systems they don't fully control, and 77% report adoption already outpacing their governance capabilities. Organizations that built control directly into their AI systems rather than relying on manual oversight reported 25% fewer incidents — a case for treating governance as architecture, not an afterthought.

The unglamorous prerequisite for AI: clean data — In a Healthcare IT News interview, Verato's Joe Hickey argues health systems are racing to deploy AI without fixing the patient-identity data underneath it, citing an S&P Global report that found 84% of organizations believe data mismatches already cost them revenue. His "garbage in, garbage out" warning is self-interested, since Verato sells identity-resolution tools, but the underlying point lands: AI trained on duplicate or mismatched records propagates those errors invisibly across every workflow it touches.

ChatGPT now turns your data into charts inline — OpenAI rolled out native chart creation in ChatGPT on web and mobile, letting users drop in data or a comparison and get a chart back without leaving the chat, with bar, line, pie, and scatter types available as interactive visuals rather than just static images. For leaders who live in spreadsheets, it's a quick way to prototype a comparison before formalizing it, with the usual caveat: check the underlying numbers before the chart leaves your screen.

👉 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

👉 July 5, 2026 — HHS comment deadline: RFI on addiction and mental health

👉 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