The Bandwidth: Decoding Health + AI

Well, I'd planned to take this week off from the newsletter. But it's hard to sit out a week when this much keeps moving, so here's a shorter note instead.

With the Fourth of July on the horizon, I found myself thinking back to a piece I wrote in May on the geopolitics of health AI — the data walls rising between the U.S., China, and Europe, and what that fragmentation does to a technology that gets better with scale and worse in isolation. If you didn't get a chance to read it then, I'm re-sharing it here.

Back to a fuller, new Mission View next week. For now, here's what else caught my attention this week. And again, I hope everyone has a safe and happy holiday.

Claude Science, an AI Workbench for Scientists, Is Now Available — Anthropic launched Claude Science, an app that consolidates literature review, data analysis, and manuscript drafting into one environment with auditable, reproducible outputs and access to more than 60 domain-specific skills and connectors for genomics, proteomics, and cheminformatics. Early users cited in the announcement include Manifold Bio for target nomination, the Allen Institute for automating long-form literature reviews, and UCSF's Brain Tumor Center for germline variant analysis.

Insurers Bet Big on AI, But Wall Street Isn't Clear on the Payoff — UnitedHealth Group, CVS Health, and Elevance Health are collectively committing billions to AI for claims processing, customer service, and internal tools, but analysts say the companies have offered little specific guidance on when — or how much — that spending will pay off. TD Cowen projects AI could lift insurer operating margins by up to 1.6 points by 2030, but warns margins could flatline near zero if the investment doesn't outpace what it saves in headcount, and provider-side AI adoption could offset gains for insurers altogether.

Anthropic Rolls Out Claude Sonnet 5 as it Restores Fable 5 Access — Anthropic released Claude Sonnet 5, a more agentic model priced closer to its Opus tier, while also restoring access to Claude Fable 5 after a two-week suspension triggered by U.S. export controls tied to a reported cybersecurity jailbreak. The company is now proposing an industry-wide framework — with Amazon, Microsoft, and Google — for scoring jailbreak severity, alongside deeper pre-release testing coordination with the federal government.

UnitedHealthcare CEO: AI-Driven Cost Inflation Not 'Going to Be Persistent' — UnitedHealthcare CEO Tim Noel said AI-driven billing pressure from providers' coding and RCM tools "will get to saturation," even as PwC projects 2027 medical cost trends will hit a 17-year high, citing AI documentation tools as a top driver. Hospital executives have pushed back on payers' "aggressive coding" framing, arguing higher-acuity documentation reflects sicker patients and rigorous compliance standards rather than gaming — a standoff Cleveland Clinic's CFO called "unsustainable" for both sides.

How A.I. Might Change the Way Doctors Think — ER physician Helen Ouyang traces how AI scribes, now used in some form by nearly every surveyed nonprofit health system, have shifted her role from composing clinical notes to auditing machine-generated ones — a change she argues quietly erodes the deliberate reasoning that writing a note once forced. She points to an unstudied risk: not that AI gets facts wrong, but that its fluent, plausible drafts make it easier for doctors to skip the painstaking work of sorting through diagnostic uncertainty themselves.

OpenAI Proposes 5% Stake to Trump Administration to Ease Washington Pressure — OpenAI has proposed giving the U.S. government a roughly $42.6 billion, 5% stake in the company as part of a broader arrangement in which Washington would hold equity in leading AI developers via a sovereign wealth fund vehicle. The pitch comes amid mounting federal scrutiny of AI cybersecurity risk and competition from cheaper Chinese models — the same climate that led to Anthropic's brief, now-resolved suspension of Fable 5 and Mythos 5 under export controls last month.

A Former AI Regulator, Now in Industry, Says Biopharma Is Reading FDA's Guidance Wrong — Tala Fakhouri, who wrote FDA's AI policy before joining Parexel as chief AI and regulatory strategy officer, says the agency's intentionally flexible guidance is being read by industry in the most conservative way possible to avoid inspection risk. She points to outdated software validation rules, misalignment between FDA reviewers and inspectors, and the loss of experienced staff through recent layoffs as compounding factors.

A 'Historic' FDA Clearance Raises the Question: Is the LLM an Interface or the Decision-Maker? — UpDoc secured what it called the first FDA clearance for a diabetes app using "patient-facing large language models," but the company's CEO declined to say whether the LLM makes treatment decisions or merely serves as a conversational interface layered over a deterministic dosing algorithm. Regulatory experts told STAT the clearance pathway — secured via substantial equivalence to a 2018 device rather than the more rigorous De Novo route — suggests the LLM functionality doesn't yet trigger the oversight a true AI decision-maker would require.

Teens Are Turning to Chatbots for Mental Health Help. We Need Rules to Keep Them Safe — Citing new JAMA Pediatrics research showing chatbot use for mental health advice among adolescents jumped from roughly 1 in 8 to 1 in 5 in a year, Ryan McBain argues blanket bans are too blunt and calls instead for tiered rules that distinguish companionship, clinical guidance, and crisis response. He proposes treating all users as minors by default, banning companion-style bots for teens, and requiring independent audits — modeled on the U.K.'s Online Safety Act and FTC enforcement against Epic Games — rather than relying on self-reported age or company promises.

Google's AI Boom Sends Emissions, Power Use Soaring — Google's latest environmental report shows electricity demand jumped 37% and greenhouse gas emissions rose 18% — its largest annual increase yet — as AI infrastructure buildout outpaces the company's data center efficiency gains. Google held its share of carbon-free electricity roughly flat and signed a record 12 gigawatts of clean energy agreements, but electricity-related emissions fell just 3% from 2024, a slower decline than the prior year, underscoring how AI growth is straining even aggressive climate commitments.

Real-World Engagement With a Generative AI Conversational Agent for Mental Health Support — A retrospective study of 5,082 paid subscribers to the Mental app found a 92.6% session-to-session return rate, with most sessions occurring in the evening and outside traditional business hours. Ninety percent of users reported moderate-to-high distress at onboarding, and session satisfaction was a significant predictor of return — evidence, the authors say, that GenAI conversational agents can sustain engagement among people who face barriers to traditional mental health care, though whether that engagement translates into clinically meaningful outcomes remains an open question.

A New Look at AI's Impact on Jobs — A new working paper from Ramp Economics Lab, using Ramp spending data linked to Revelio Labs workforce records across more than 21,000 U.S. firms, finds that companies with high-intensity AI adoption grew headcount 10% over two years, with entry-level hiring growing even faster at 12%. The gains were concentrated entirely among high-intensity adopters — those in the top third of AI spend per employee — while low-intensity adopters saw no significant change, and the paper's author, Ramp's lead economist, notes AI adoption itself is unevenly distributed, correlating more with funding source and location than industry secto

Zelis Rolls Out AI Solution to Navigate No Surprises Act Disputes — Zelis launched an AI-native tool to help payers manage Independent Dispute Resolution (IDR) under the No Surprises Act, automating claim repricing, negotiation, and case management as dispute volume has ballooned to more than 100 times CMS's original projections. The tool comes as new CMS rules cut IDR administrative fees from $115 to $15 and add standardized reporting requirements, and as providers continue winning the large majority of arbitration cases.

Healthcare's AI Problem Isn't the Model — It's the Data — As health systems move from AI pilots to production, Dr. Jaime Bland, CEO of Aquila Health, argues the real barrier isn't model performance but fragmented, poorly governed data — pilots succeed on curated datasets, while live data is "messy, non-standard, and incomplete in ways the pilot never surfaced." She points to outcomes, social determinants, and race/ethnicity fields as the thinnest and most bias-prone, and argues AI governance is inseparable from data governance.

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

👉 July 16, 2026, 12:00–3:30 PM ET — Workshop: Simulation and Causal Models for AI Evaluation and Regulation

👉 Aug. 4–6 — Ai4 — Las Vegas

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

👉 Oct. 22–23, 2026 — HIMSS AI in Healthcare Forum San Diego + AI Leadership Summit — San Diego, CA

Till next time,

BC