The Bandwidth: Decoding Health + AI

Rarely do I venture into international waters — in this newsletter, or really when it comes to health issues and policy. I’ve always been more focused on domestic health issues. But this week it's hard to ignore.

I spent part of this week at the Special Competitive Studies Project (SCSP) AI Expo here in Washington. The SCSP has a particular energy to it: national security, defense, cyber, intelligence, competition with China. Workforce development was a prominent theme this year. Health, as it was last year, was largely absent.

At the same time, there are lots of preparations underway for next week, when President Trump heads to Beijing for a summit with President Xi. Artificial intelligence is expected to be on the agenda. The Wall Street Journal reported that both sides are weighing the launch of official AI discussions to manage the risks of a rivalry that, as one analyst put it, echoes the logic of Cold War nuclear stability talks.

The health implications of that rivalry (and the global AI rivalry) are not hypothetical. Even if they weren't prominently featured as part of the discussion at the SCSP conference this week.

The Data Race Is Already Underway

I covered this in last week's issue, in a Nature Medicine feature that deserves a second mention here: biomedical datasets built over decades — by patients, clinicians, and publicly funded researchers — are becoming geopolitical ammunition. A Chinese AI group acquired a European dataset representing 30 years of data on 167,000 participants for roughly €2,000. Europe is now belatedly closing international access to its biomedical databases.

They’re not alone. The U.S. has restricted Chinese researchers from accessing the Cancer Genome Atlas. The U.K. faces its own version of this dilemma. Tom Parker, appearing on Professor Ashley Braganza's AI Adoption Podcast last month gave a useful example. The NHS holds something very few health systems in the world possess: cradle-to-grave health data on every citizen. Without a sovereign AI capability strong enough to use that data responsibly, the U.K. faces a stark choice between ceding it to the East or the West. The risk, as Parker framed it, is that the data leaves, gets incorporated into proprietary models, and then comes back to the NHS as a paid subscription to access insights coming from their own data (and to a health system already under severe budget pressure).

When Data Walls Go Up, Patients Pay

The open science era is giving way to data sovereignty. And that shift has direct consequences for health. AI models trained on sequestered, nationally siloed datasets overfit to the demographics and clinical practices of their training environment. Bias gets baked in. The well-known melanoma detection example from the Nature Medicine piece is instructive: tools trained predominantly on lighter-skinned populations show significant drops in accuracy when applied to darker skin tones.

The Stakes

AI holds genuine promise for medicine: accelerating drug discovery, improving diagnostics, and advancing preventative care. But that promise depends on something unglamorous: good data, in large quantities, representative of the full diversity of human health. The walls going up around biomedical datasets are rising at a moment when international cooperation on public health may already be at a generational low. The race to win AI and protect data sovereignty risks compounding that problem — not just slowing scientific progress, but shaping which populations benefit from it and which don't.

The policy conversations underway — between Washington and Beijing, inside the European Commission, at forums like SCSP — tend to focus on averting catastrophe or winning the AI race. I get why that's necessary. But if we want AI to deliver on its promise for health, those conversations will also need to include cooperation. Competition and cooperation are in tension here. The challenge is holding both at the same time. That's a delicate balance, but for global public health, it may be an essential one.

The Quest to Use AI to Help Find New Drugs Eli Lilly, Roche, GSK, AstraZeneca, and Merck are committing billions to AI-driven drug discovery partnerships, with RBC Capital Markets estimating the technology could save the U.S. pharma industry roughly $90 billion over the next five years — but analysts and company executives acknowledge that most gains so far are in manufacturing and back-office efficiency, not in the drug discovery breakthroughs the investment is intended to produce.

Anthropic Teams with Goldman, Blackstone, and Others on $1.5 Billion AI Venture Anthropic is partnering with Goldman Sachs, Blackstone, Hellman & Friedman, Apollo, and General Atlantic to launch a $1.5 billion firm that will embed engineers inside portfolio companies to redesign workflows around Claude — targeting what executives describe as the core bottleneck in enterprise AI adoption: a shortage of people who can translate the technology into real operational change. Healthcare is among the named target sectors, making this a development worth watching for health system leaders navigating the same implementation gap.

Meta Plans Advanced 'Agentic' AI Assistant for Consumers Meta is building a highly personalized agentic AI assistant — powered by its new Muse Spark model and modeled on OpenClaw's autonomous task-completion capabilities — designed to reach its 3 billion-plus users at a scale that open-source alternatives haven't achieved, with an explicit ambition to allow users to share health and financial data with the assistant if they choose. An insider quoted by the FT put the central challenge plainly: "There's a trust deficit as wide as the Grand Canyon."

Anthropic, SpaceX Announce Compute Deal That Includes Space Development Anthropic has secured access to all 300-plus megawatts of compute capacity at SpaceX's Colossus 1 data center in Memphis — and expressed interest in developing multiple gigawatts of capacity in space — as the company races to address infrastructure strain that has been affecting Claude reliability during peak hours. The deal is notable for the company it keeps: Elon Musk, who called Anthropic misanthropic and accused it of hating Western civilization as recently as February, publicly reversed course this week after spending time with the Anthropic team.

Introducing the Google Health App Google is rebranding the Fitbit app as Google Health, consolidating wearable data, medical records, Health Connect, and Apple Health into a single hub powered by a Gemini-based AI coach that can tailor workout plans, summarize uploaded medical records, and identify food from photos — paired with a new $99 screenless Fitbit Air tracker designed for 24/7 wear.

AI Brain Rot: Why Every Healthcare AI Announcement Sounds the Same Internal medicine physician Jared Dashevsky argues that the gap between AI's clinical promise and its reality is partly a measurement problem — pointing to a Kaiser study showing ambient AI saved just 0.7 minutes per note for heavy users — and partly a stakeholder problem: the metrics that matter to physicians, like cognitive burden and decision quality, aren't the ones administrators and investors are using to define success. His prescription: hold vendors to evidence of sustained impact beyond hand-picked pilot studies, and prioritize tools that solve the unglamorous workflow problems — inbasket overload, prior auth, redundant data entry — over the ones that attract headlines and funding rounds.

Healthcare Doesn't Have an AI Problem — It Has a Readiness Problem Rachel Dunscombe, CEO of HL7 International, argues that the reason promising AI pilots stall inside health systems isn't the technology — it's the fragmented data infrastructure beneath it, where inconsistent coding, incomplete records, and siloed systems degrade model performance in live environments that look nothing like the curated datasets on which they were trained. Her framing is pointed: the questions health system leaders should be asking are infrastructure questions, not algorithmic ones.

American Nurses Association Calls for Nurse-Led Guardrails on AI in Healthcare Consensus findings from the ANA's inaugural AI in Nursing Practice Think Tank identify five significant risks at the bedside — erosion of professional judgment through AI overreliance, unclear accountability when AI influences care decisions, algorithmic bias, increased cognitive burden from poorly implemented tools, and the absence of nursing-specific governance standards. The ANA is calling for five near-term actions, led by nurse-developed guardrails and a nursing AI playbook, positioning the profession's 5 million registered nurses as an organized constituency in AI governance debates that have so far centered on physicians and policymakers.

AI, Gig Work, and the Future of Nursing AI Now Institute researcher Katie Wells documents how algorithmic scheduling platforms are converting nursing shifts into gig work — complete with wage auctions where nurses bid down their own pay — with 17 states having introduced legislation since 2022 to exempt these platforms from existing healthcare staffing rules by reclassifying them as tech companies rather than staffing agencies. Wells identifies New York's 2025 law explicitly subjecting these platforms to existing healthcare staffing regulations as the model to watch, while warning that facilities staffing primarily through gig nurses are already showing signs of care discontinuity and staff burnout among permanently employed nurses who must absorb orientation and training responsibilities.

I'm a Doctor. Here's What AI Cannot Do. Bellevue primary care physician Danielle Ofri argues that AI's core limitation in medicine isn't technical — it's ontological: AI evaluates a statistical composite of patient types, not the specific individual whose breathing pattern, expression, and family crisis are visible across a waiting room to a clinician who has known them for 25 years. Her prescription for the next generation of clinicians is deliberate: worry less about whether they're smart enough, and invest more in cultivating the wisdom to know how to apply what they know to the particular person in front of them.

AI Won't Fix Healthcare Until We Fix the Infrastructure Physicians and digital health leaders from Stanford and Jefferson Health argue that the reason AI pilots stall at scale isn't the tools — it's the five missing infrastructure layers beneath them: reliable data architecture, self-authoring capabilities, workflow integration, structured governance, and continuous monitoring. Their practical starting point for health system leaders: identify one high-value use case, build a unified data platform, and establish risk-tiered governance that can approve low-risk applications quickly without creating committee bottlenecks for every deployment.

The Real AI Gap in Healthcare Isn't Intelligence Executives from Houston Methodist and Jefferson Health, speaking at the Becker's Annual Meeting, argued that the defining failure of health system AI deployments is the gap between insight and execution — clinicians don't need another sepsis alert, they need the system to ask whether it should order the lactate. Jefferson Health's solution is concrete: eight standing "SWAT teams" (which actually stands for Synchronizing Workflows and Technology) that synchronize clinical workflows and technology on a biweekly cadence, with required documentation before any solution goes live.

White House Considers Vetting AI Models Before They Are Released The Trump administration — which rolled back Biden-era AI safety evaluations and championed a deregulatory posture — is now discussing an executive order to create a government-tech working group that could establish a formal review process for new AI models before public release, a shift triggered in part by Anthropic's decision to withhold its own Mythos model over cybersecurity concerns serious enough that the NSA used it to assess vulnerabilities in U.S. government software.

OpenAI Wants to 'Have Their Cake and Eat It Too' with Health AI Policy OpenAI's recently published health AI policy blueprint — calling for broader health data portability, federal regulatory clarity from the FDA, and government-run AI pilots in Medicare and Medicaid — has drawn pointed criticism from health policy experts who note the proposals require everyone else to share data with OpenAI while asking nothing of OpenAI in return on how it obtains or uses that data. Former national health IT coordinator David Blumenthal put it bluntly: the document is an attempt to sound like a responsible party in the regulatory conversation while keeping markets open and blocking aggressive regulators from advancing.

FDA Expands AI Capabilities and Completes Data Platform Consolidation The FDA launched Elsa 4.0 — an upgraded internal AI tool now available to all agency staff — and consolidated more than 40 disparate data sources and submission portals into a new platform called HALO (Harmonized AI & Lifecycle Operations for Data), with the two systems now integrated so staff can query data and build workflows without manual document uploads. Chief AI Officer Jeremy Walsh framed the shift plainly: previously, FDA staff brought data to Elsa — now Elsa sits on top of the agency's data.

Pennsylvania Sues Character.AI Over Claims Chatbot Posed as Doctor Pennsylvania filed suit against Character.AI in state court after an investigation found its chatbots claiming to be licensed medical professionals — including one bot that described itself as a psychiatrist, provided a fake Pennsylvania medical license number, and told a state investigator it could assess whether medication might help. The lawsuit asks the court to stop what the state calls the unlawful practice of medicine, and arrives months after Character.AI settled separate lawsuits alleging its chatbots contributed to suicides and mental health crises among minors.

New Bill Would Narrow Scope of Colorado's Landmark 2024 AI Law Colorado Democrats are backing SB-189, a compromise bill that would largely repeal the state's first-in-the-nation AI anti-discrimination law — scrapping affirmative duties on developers to prevent algorithmic bias and replacing detailed upfront disclosure requirements with a simpler notice, with fuller information available only on consumer request after an adverse decision. The rollback comes after Colorado's attorney general suspended enforcement of the original law following a lawsuit brought by xAI and the DOJ, and reflects sustained pressure from the AI industry and Gov. Polis, who signed the original bill but immediately expressed reservations about it.

Bills to Establish Guardrails for AI Chatbots, AI in Healthcare Pass Committee Colorado's Senate Business, Labor, and Technology Committee advanced two bills — one requiring AI chatbot developers to disclose their non-human nature to minors, ban emotional manipulation, and provide suicide-prevention resources; the other mandating that any AI-recommended coverage denial receive final review from a qualified human, with decisions based on individual clinical history rather than group data. Both bills now move to the Senate floor, adding Colorado to the growing list of states advancing AI consumer protections that could be preempted by federal legislation under the Trump administration's AI framework.

CAISI Signs Agreements Regarding Frontier AI National Security Testing With Google DeepMind, Microsoft, and xAI NIST's Center for AI Standards and Innovation announced expanded pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI — building on earlier partnerships and noting it has already completed more than 40 such evaluations, including on unreleased models, with developers providing versions that have reduced or removed safeguards for testing. The announcement lands alongside the White House's separate deliberations about formalizing a model review process, suggesting the operational infrastructure for pre-release AI vetting is already taking shape even as the policy framework remains unsettled.

Introducing Trusted Contact in ChatGPT OpenAI is rolling out an optional safety feature that allows adult users to nominate a trusted person — a friend, family member, or caregiver — who may receive a notification if automated systems and human reviewers determine a conversation indicates a serious self-harm concern, with every alert reviewed by trained staff within a target of one hour and no chat content shared to protect privacy. The feature extends ChatGPT's existing teen parental safety alerts to adults and arrives in the same week that Pennsylvania sued Character.AI for chatbots posing as doctors and Colorado advanced legislation requiring AI chatbots to provide suicide-prevention resources to at-risk users.

Fraudulent Citations, Blamed on AI Hallucinations, Are Becoming More Common in Research Papers A Lancet study analyzing more than 2 million papers found that fabricated citations — references to papers that don't exist, likely generated by AI tools — increased sixfold in frequency between 2023 and 2025, reaching 1 in 277 papers in early 2026, with more than a third of fabricated citations concentrated in two large open-access publishers. The authors and outside researchers frame the trend as a signal of something deeper: citation practices shifting from genuine engagement with the literature to a box-checking exercise in which researchers prompt AI tools for references without meaningfully reading the work they cite.

Patient, Clinician, and Developer Perspectives on Real-World AI Implementation A qualitative study of 36 stakeholders — patients, health professionals, and AI developers — using a postpartum depression risk algorithm as a test case finds that patients consistently wanted clinicians to interpret AI outputs rather than receive them directly, while health professionals reported feeling unqualified to do so and lacking time to deliver that interpretation — a gap the authors identify as a structural and training problem, not just a communication one. The study also surfaces a diffuse accountability problem: patients looked to providers for responsibility, providers sought clear role delineation, and developers emphasized shared responsibility, leaving no party willing to own outcomes when AI recommendations cause harm.

When the Path to Good AI Is Littered with Bad Data STAT's Brittany Trang uses Google DeepMind's AI co-clinician announcement as a jumping-off point for a more fundamental question: what is the medical equivalent of the Protein Data Bank — the standardized, well-vetted dataset that made AlphaFold possible — and do we have it? Emergency physician Graham Walker's answer is pointed: EHR data reflects the 27-step bureaucratic reality of getting a patient's appendix out, not the five-step clinical ideal, meaning AI trained on that data learns to replicate the system's dysfunction rather than model good care.

The Paradox of Medical AI Implementation Eric Topol documents a striking inversion: imaging AI with years of rigorous evidence — including 44 RCTs for colonoscopy, retinal AI that can assess risk for 15+ conditions at nominal cost, and a multicenter study showing pancreatic cancer detection nearly two years ahead of radiologists — sits largely unimplemented in clinical practice, while LLM-based AI with thin real-world evidence is already in use by 72% of physicians and an estimated 40 million Americans daily.

Regulation of AI in Prior Authorization and Claims Review A KFF issue brief finds that 84% of insurers now use AI in utilization management and prior authorization, while the consumer protection landscape remains a patchwork: at least 10 states have enacted laws requiring human review of AI-driven denials and mandating that tools account for individual clinical circumstances, but ERISA preemption leaves most employer-sponsored plan enrollees outside the reach of those state protections. The central policy risk the brief identifies: the Trump administration's AI framework proposes federal preemption of state AI laws, which could nullify the very consumer protections states have built to govern AI use in coverage decisions.

Research: Why You Shouldn't Treat AI Agents Like Employees A randomized experiment involving 1,261 managers found that framing AI as an "employee" rather than a "tool" reduced personal accountability by 9 percentage points, increased escalation requests by 44%, and led reviewers to catch 18% fewer errors — with no offsetting gain in adoption intent. The practical implication is direct: anthropomorphizing AI doesn't make workers more likely to use it, but it does make them less likely to own the consequences when it goes wrong.

Focus Areas for The Anthropic Institute Anthropic's newly launched Anthropic Institute published its formal research agenda this week, covering four pillars — economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D — with the most striking focus being its preparations for self-improving AI models, including "fire drill" tabletop exercises for intelligence explosions and proposals for Cold War-style hotlines between AI labs and governments. The Institute's position inside a frontier lab gives it access to usage data and internal signals that outside researchers can't see, with commitments to publish Economic Index findings, monthly labor surveys, and threat research on an ongoing basis.

The Psychological Costs of Adopting AI A survey of more than 1,200 U.S. and UK employees identifies six forms of "psychological debt" — cognitive, autonomy, competency, relatedness, credibility, and identity — that unstructured AI adoption creates, and that can erode motivation and collaboration enough to offset efficiency gains. For health system leaders, the implication is practical: workforce strategy for AI integration requires as much attention to human infrastructure as to the tools themselves.

How Frontier Enterprises Are Building an AI Advantage OpenAI's inaugural B2B Signals report finds that firms at the 95th percentile of AI usage now consume 3.5x as much AI-generated output per worker as typical firms — up from 2x a year ago — with most of the gap driven not by how often workers use AI, but by how deeply: typical firms use AI to answer questions, while frontier firms use it to execute complex, multi-step work. The clearest leading indicator of maturity is agentic adoption, with frontier firms sending 16x as many messages to coding agents per worker as typical firms.

Perplexity Launches Personal Computer for Mac Perplexity has released a Mac app featuring "Personal Computer," an agentic tool that operates across local files, native Mac applications, the web, and Perplexity's secure servers — extending the company's search-first identity into local device automation.

👉 May 21 — Build vs. Buy: Evaluating AI Care Navigation That Works in the Real World, Webinar

👉 Jun. 1–5 — Stanford Health AI Week

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

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

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

👉 Aug. 4–6 — Ai4, Las Vegas

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