The Bandwidth: Decoding Health + AI

Most of the oxygen this week went to the standoff between the federal government and Anthropic over its sidelined Fable model, which we cover plenty of in the headlines below.

But another story grabbed my attention this week that I want to focus on: what AI is about to do to the cost of American health care.

The reflex is to assume AI will bring those costs down. Automate the back office, thin out the paperwork, catch disease earlier, and surely the bill shrinks. But mounting evidence points the other way.

PwC (my old firm) projected that medical cost trend, the annual growth in what it costs to treat patients, will climb about 9 percent in the employer market next year, and roughly 8.5 percent in the individual market. And when PwC lays out the forces behind that increase, the very technology that we hope will bend costs down does the opposite.

Five inflators, one of them AI

PwC points to five drivers lifting medical costs in 2027: AI-enabled tools that help providers capture more revenue; inflation and provider consolidation pushing up reimbursement rates; rising pharmacy spend; growing behavioral health utilization; and a No Surprises Act arbitration process that, in practice, is steering more money to out-of-network providers.

To PwC's credit, it doesn’t overstate the AI piece. They note that labor, supplies, and utilization remain bigger forces, and that heavier coding isn't always improper, since some systems have been under-documenting real complexity.

But the direction is unmistakable. AI scribes and ambient documentation tools capture more of what happens in a visit, which means more diagnoses, higher-severity codes, and bigger bills — even when the underlying care hasn't changed.

Just to put some numbers behind this. Recent data from Blue Cross Blue Shield show that at the hospitals adopting AI coding tools fastest, the share of maternity patients coded with serious post-delivery bleeding jumped from about 4 percent to more than 12 percent in three years, with no matching rise in transfusions, and an estimated national cost of $2.3 billion.

The biggest line items

So, AI may be adding to our national bill because of coding. Even if that weren’t an emerging issue, the other question is whether it can actually eat into the big cost drivers in healthcare.

Hospital care is about 30 percent of national health spending (and about 40% of the growth between 2022-2024) — the single largest category. And labor is 56 to 60 percent of what hospitals spend. The people caring for people are the cost. Unless we're prepared to replace doctors and nurses outright (which I don’t think we are based on emerging evidence), it's hard to see AI taking a real bite out of the largest driver of the bill.

The opposite effect

Just to add more complexity to the picture, even working exactly as intended, AI may push costs up. Better imaging finds more cancers earlier; better diagnostics surface conditions that would have gone unnoticed. That's good medicine. And more of it. Earlier detection widens the population in treatment.

The value is real, but so is the spending. Geoffrey Hinton made a version of this point a few months back, calling health care "a very elastic market": let each clinician do ten times as much, and we'll find ten times as much care to consume.

Avoiding that outcome is going to be tough work and require tough choices. It reminds me a lot of the conversation going on about AI adoption going on across sectors and companies, which necessitates reimaging how work gets done and organizational transformation.

The wrong way to adopt AI in healthcare is to bolt it onto existing workflows and payment models we already have, which only lets the system do what it was built to do: spend more, faster, and at scale. The right way is harder: use AI as the reason to redesign how care is delivered and how it's paid for. Lower costs won't come from layering intelligence on top of a system engineered to spend. They'll come from rebuilding what sits underneath — the care model and the payment model.

Improving health intelligence in ChatGPT — OpenAI announced that GPT-5.5 Instant, now available to all free ChatGPT users, matches its frontier "Thinking" models on the company's own health evaluations, and said more than 230 million people use ChatGPT for health questions each week. In OpenAI's own physician-led testing, it reports GPT-5.5 Instant responses were rated higher than physician-written ones across criteria like accuracy and completeness, and that responses flagged for factuality issues fell 71% over two months — all self-reported figures from internal benchmarks and self-monitored traffic.

In the Age of AI, Interoperability Isn't Enough — In a MedCity commentary, CodaMetrix CEO Hamid Tabatabaie argues that moving more data between systems won't fix healthcare AI's core problem: the same clinical story gets coded differently across systems, and documentation tools that generate codes without longitudinal context can produce conflicting or inaccurate outputs. He calls for an objective framework layered above interoperability to standardize how context and quality are interpreted across clinical, financial, and analytical uses.

Hyperscaler capex is on trend to outpace their cash inflows by the end of 2026 — An Epoch AI analysis of SEC filings finds that aggregate cash capital expenditures across the five largest hyperscalers — Microsoft, Amazon, Alphabet, Meta, and Oracle — are growing far faster than their operating cash flow (roughly 70% versus 23% a year) and are on track to overtake it around Q3 2026. With capex outpacing the cash their operations generate, most of these companies have already turned to debt or equity financing to fund their AI infrastructure buildout.

Anthropic, Google DeepMind CEOs call for U.S.-led AI coalition at G7 — At a closed-door G7 lunch in France, Anthropic's Dario Amodei and Google DeepMind's Demis Hassabis reportedly urged world leaders to back a U.S.-led effort to govern AI through shared international rules, with Amodei pressing for coordinated limits on chip exports to China and a common framework for controlling access to frontier models. According to the reporting, nothing concrete emerged from the session, and China's foreign minister separately floated a competing AI cooperation body open to all countries.

Houston Methodist turns ambient AI into an enterprise tool, not just a pilot — Houston Methodist reports scaling its ambient AI documentation platform (from vendor Ambience) across ambulatory, emergency, and inpatient settings, with self-reported utilization above 85% of encounters, a 44% average drop in documentation time, and a 35% reduction in after-hours charting. CMIO Jordan Dale attributes the enterprise-wide adoption less to the technology itself than to tight EHR integration and clinician trust in note quality, with even niche specialties sustaining 70%+ use.

This State Is Testing Out AI Doctors—and Actual Doctors Aren't Happy About It — Through a regulatory sandbox, Utah has begun allowing an AI service from the startup Doctronic to renew patient prescriptions for drugs like Lipitor and Prozac — work normally done by physicians — though the current phase still requires a doctor to sign off on each AI-approved renewal before it's filled. Most of the state's medical licensing board has signed a letter seeking to suspend the project on safety grounds, citing liability questions, the absence of FDA review or published studies, and what one member called a "black box."

The real work for making dramatic gains against pancreatic cancer is just beginning — In a STAT First Opinion essay, Andrea Califano and Gideon Bosker — co-founders of the biotech Darwin Health — argue that even a real survival gain like daraxonrasib's recent pancreatic-cancer result is only a first step, because the deadliest tumors are "oncosystems" of multiple coexisting, shape-shifting cancer cell states that single drugs can't defeat. They contend AI and computational-biology methods can now map each state and match it to a targeted drug, pointing to Nature Genetics work on a pediatric brain cancer where seven cell states proved targetable with three drugs.

AI is about to escape human control — and nobody has a plan — In a Hill op-ed, contributor John Mac Ghlionn argues governments have no workable plan for the loss-of-control risks Anthropic raised this month when it floated a coordinated pause on frontier AI. He dismisses the administration's new pre-release model review and Europe's slower rulemaking as far too thin given the pace of advances.

Unpacking the Great American Artificial Intelligence Act of 2026 — Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) released a 269-page bipartisan discussion draft that would create the first comprehensive federal AI governance regime, imposing binding development, transparency, and auditing obligations on "large frontier developers" with over $500 million in revenue. Its preemption of state laws regulating model development — sunsetting in 2029, while preserving state authority over deployment and use — has drawn early opposition from groups including Public Citizen and the AFL-CIO.

Anthropic's Mythos Recall and the White House's Missing AI Safety Playbook — The Commerce Department, citing a reported jailbreak of Anthropic's Fable 5, issued an export-control directive on June 12 barring all foreign-national access to Fable 5 and Mythos 5, prompting Anthropic to disable both models for every customer. Editor Justin Hendrix argues the episode, landing days after the administration's "voluntary" oversight executive order, exposes the absence of a stable federal playbook for frontier-model risks.

An Open Letter on Transparent AI Cyber Protections — More than 100 cybersecurity executives and researchers, including Alex Stamos, Bruce Schneier, and Katie Moussouris, signed an open letter urging Commerce Secretary Lutnick and National Cyber Director Cairncross to lift the export-control directives on Anthropic's Fable and Mythos models. The signatories argue the models' cyber capabilities are not unique to Anthropic and are replicable on other tools, and that stripping defenders of the best models while adversaries advance creates risk without justification.

OpenAI Investigated by Coalition of State Attorneys General — A coalition of state attorneys general has opened an investigation into OpenAI, serving it with a subpoena from New York's AG seeking documents on its advertising, user engagement, handling of consumer and health data, and activities involving minors and seniors, the Wall Street Journal reported. The probe is the latest in a widening wave of state-level legal action against AI companies, following Florida's recent lawsuit against OpenAI and Sam Altman and earlier inquiries into xAI's Grok.

FDA Workshop Highlights Promise, Risks Of AI In Generic Drug Reviews — At a two-day FDA workshop on generic drugs, officials and industry weighed using AI for tasks like updating the inactive ingredient database, calculating maximum daily dose values, and auditing applications, while repeatedly returning to questions of validation, traceability, and hallucinations. Much of the debate centered on how much human oversight AI-supported review requires — with FDA's Robert Lionberger asking whether systems could be made robust enough that reviewers needn't check every low-level extraction, and others insisting scientific interpretation must remain with human experts.

OpenAI, Google DeepMind, Anthropic request synthetic DNA screening legislation — A broad coalition including the CEOs of OpenAI, Google DeepMind, Anthropic, and Microsoft AI, alongside biotech executives, scientists, and national security officials, urged Congress to require synthetic DNA and RNA providers to screen every order and verify customers, rather than rely on the voluntary safeguards in place since 2009. The signatories warn that rapid AI advances could erode the knowledge barriers that have kept bad actors from designing dangerous pathogens, and they call for federal action this session to set a consistent national standard.

Seattle uses AI to help triage, divert 911 medical calls — A Seattle Times investigation found the Seattle Fire Department has used a Danish company's AI to listen to 911 medical calls for more than two years without notifying callers or conducting public review, using it to help divert some patients from ambulances to a nurse line. Experts flagged transparency, privacy, and bias concerns, noting the deployment was never assessed under the city's surveillance ordinance.

I'm a pediatrician. I want to prescribe the right AI to my patients — In a STAT First Opinion essay, Boston Children's physician Dua Hassan argues the real harm in children's screen time is passivity, not screens, and that well-designed AI could restore the "serve-and-return" interaction development depends on — by responding, asking questions, and modeling emotional language rather than optimizing for engagement. She calls for AI built for children to be designed with pediatricians and developmental scientists from the start and tested through randomized trials measuring real developmental outcomes, the way pediatric drugs are.

Who bears liability when AI gives bad prescribing advice — In an npj Digital Medicine comment, legal scholars Julia Etkin and Vincent Joralemon argue that the "learned intermediary" doctrine may shield AI developers when a clinician sits between the tool and the patient, but not when chatbots advise patients directly — leaving developer liability real but unsettled. Their more immediate claim is for clinicians: as patient use of AI grows, screening for AI-sourced prescribing advice may become part of the standard of care, and they recommend a routine intake question asking whether patients have used a chatbot for medication or symptom advice.

Sustaining High Reliability Amid Artificial Intelligence Adoption in Oncology — A Journal of Clinical Oncology commentary argues that as tools like ambient scribes and auto-segmentation spread through cancer care, AI reliability becomes a continuous sociotechnical problem rather than a one-time technical validation, because models can fail silently through data drift, plausible-looking hallucinations, automation bias, and clinician deskilling. The authors propose pairing high-reliability-organization principles with Reason's Swiss cheese model to cultivate what they call a capable human-in-the-loop — clinicians equipped to spot drift, calibrate trust, and override AI, rather than a nominal reviewer who approves plausible-but-wrong drafts under time pressure.

Study Finds Students with Highest Distress Use AI for Mental Health at Elevated Rates — A study co-led by Mass General Brigham investigators and published in the Journal of Affective Disorders found that 18% of surveyed college students reported using AI for mental health, with moderate-to-severe depression, severe anxiety, and suicidality each associated with roughly double the likelihood of such use. The authors caution that the students most drawn to AI for support may be the most vulnerable to its risks, and call for crisis-detection and referral mechanisms to be built into these tools.

General-purpose LLMs outperform specialized clinical AI tools on medical benchmarks — In a Nature Medicine study, NYU Langone researchers found that three frontier models — GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6 — outperformed the specialized clinical tools OpenEvidence and UpToDate Expert AI across medical-knowledge questions, clinician-alignment benchmarks, and 100 real physician queries reviewed blindly by 12 clinicians. On the real-query benchmark, the specialized tools performed no better than Google's auto-enabled Search AI Overview.

Agentic coding and persistent returns to expertise — In an analysis of roughly 400,000 Claude Code sessions, Anthropic — the maker of Claude — found that domain expertise, not coding proficiency, predicts success with AI coding agents, with users across major occupations succeeding at nearly the rate of software engineers and people generally deciding what to build while the agent decided how. The company cautions the findings are preliminary and rest on model-read transcripts rather than observed real-world outcomes.

Towards autonomous medical artificial intelligence agents — In a Nature study, German-led researchers tested MIRA, an autonomous AI agent operating in a sandboxed EHR, on more than 500 retrospective emergency-department cases, where it matched or exceeded board-certified and mixed-seniority physicians on diagnostic accuracy and made guideline-concordant, medication-safe decisions under identical conditions. The authors stress the results come from simulation on past records rather than live care, flag possible benchmark contamination, and argue such agents should augment clinicians under supervision rather than replace them, pending prospective real-world validation.

Americans and AI 2026: Chatbots, Smart Devices and Views on Impact — A Pew Research Center survey of 5,119 U.S. adults found that about half now use AI chatbots — roughly a quarter daily — though only a minority turn to them for medical advice or emotional support, and views of AI tilt notably negative even among younger adults. Roughly two-thirds say AI is advancing too quickly, about seven in ten expect it to make their personal information less secure, and 67% have little or no confidence in the government to regulate it effectively.

AI Is Splitting the Job Market in Two, PwC Study Shows — A PwC analysis of more than a billion job postings across 27 countries found that roles requiring AI skills grew nearly eight times faster than the overall market in 2025 and carried a widening wage premium, with the fastest growth in jobs where AI amplifies human judgment — PwC cites radiologists and recruiters — rather than ones where it simply de-skills a task. Greater AI exposure correlated with faster headcount growth rather than job losses, though healthcare lagged most sectors in AI-driven job growth at under 1%. The widening premium tracks the pattern in the NBER research we covered two weeks ago, where new work commands an early wage premium that fades as the skills diffuse.

Clinical AI—Whom Does It Serve and Who Pays? — A JAMA Health Forum Viewpoint argues that the main barrier to clinical AI in routine care is increasingly economic rather than technical: the organizations paying to buy, integrate, and monitor a tool are often not the ones that capture its downstream savings, a misalignment that strands technically sound products before they reach routine use. The authors call for four things to be settled before any pilot — who pays, who primarily benefits, what defines value, and who holds accountability over the tool's life cycle — arguing that clinical studies become interpretable only once that economic clarity exists.

Is AI ruining our skills? Early results are in — and they're not good — A Nature news feature reports mounting evidence of AI-driven "deskilling," led by a study of experienced Polish endoscopists whose adenoma detection rate on unassisted colonoscopies fell from 28.4% to 22.4% after they grew accustomed to an AI detection tool. It lands alongside a survey this month finding that 77% of physicians and 70% of nurses worry about losing skills to over-reliance on AI — and researchers quoted in the piece caution there is no established fix yet.

Using AI to help physicians diagnose rare genetic diseases affecting children — In a study published in NEJM AI, researchers from Boston Children's Hospital, Harvard, and OpenAI used OpenAI's o3 Deep Research model to reanalyze 376 rare-disease cases that had remained unsolved after prior expert review, surfacing evidence-linked hypotheses that — after specialist review, further testing, and lab confirmation — led physicians to establish 18 new diagnoses, a 4.8% added yield. The authors stress the model diagnosed no one and made no clinical decisions, and note the study was retrospective, with reviewers unblinded to the model's confidence scores and no measurement of time saved, cost, or false-positive burden

A frontier without an ecosystem is not stable — In a post on X, Microsoft CEO Satya Nadella argues the durable AI advantage for an organization isn't picking the best model but building a "learning loop" on top of models that compounds its own institutional knowledge — letting it swap out a generalist model without losing the "company veteran" expertise encoded in its systems. He frames owning that loop, via private evals and reinforcement learning on internal data, as the real test of a firm's control and IP in the AI era.

How modulAIre and IBM are lowering the barrier to AI adoption — An IBM product blog details how its "AI-in-a-Box" platform — built on IBM Fusion and watsonx with partner modulAIre — packages LLMs, analytics, and automation so resource-strapped nonprofits can stand up governed AI in weeks instead of months. For lean teams weighing build-versus-buy, it's a case for managed, prebuilt AI stacks as a faster on-ramp than assembling infrastructure from scratch.

To Thrive Alongside AI, Focus on Mindset—Not Skillset — In HBR, Goldman Sachs CIO Marco Argenti argues that workers shouldn't cling to the sliver of their job AI can't yet do, but instead reimagine their entire role — shifting from hands-on operator to supervisor of AI agents while holding onto durable human judgment. For leaders, he frames the real work as change management: setting radical productivity targets, obsessing over evaluations that define "what good looks like," and getting organizational data in order before scaling AI.

👉 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