The Bandwidth: Decoding Health + AI

"If you want to understand the future of work and AI, look at what the frontier labs in Silicon Valley are doing. That's where we're all going to end up. They're a year ahead."

I heard that on a podcast at the beginning of the week. And there is probably some truth to it. The people and the places closest to the technology, those who are deploying it with great speed in their own workplaces, would surely have more insight than the rest of us. Case in point: OpenAI published new research this week on how AI is transforming work.

But I don't buy it, at least not completely. Especially when it comes to healthcare.

The frontier's version of work

OpenAI looked at more than 800,000 work messages and found that 43.5% of occupation-specific use is people doing tasks that belong to another job. The marketer troubleshoots the website. The salesperson runs the analysis that used to go to an analyst. They call it "task crossover." And they frame it as an early signal of where work may be headed.

For a lot of work, that seems like a fair read. But look at who is doing the crossing. The heavy movers are customer experience, design, human resources, legal, and marketing. And, as OpenAI's research points out, this particular phenomenon might be well suited for smaller teams or businesses, where people may already be wearing many hats.

Now look at who is not on the list. Nurses. Doctors. Anyone whose work happens on a body instead of a browser.

The frontier labs are describing knowledge work, not necessarily the future of all work.

Where the crossover stops.

Task crossover works when a task is portable text, something you can hand to a chatbot. Although if you are a marketer leaning on AI to do a lawyer's job, do you have the expertise to really judge whether it got the contract right?

You may be able to borrow "write the campaign" or "review the contract." You cannot borrow "start the IV" or "be in the room with a frightened family when the news is bad, and be trusted."

This is actually where I think healthcare may be ahead of the AI labs, doing the hard work of figuring out where to lean on AI, and where not to.

The American Nurses Association, for instance, put it in its code years ago: technology is "adjunct to, not a replacement for" the nurse's knowledge and skill, and nurses stay "accountable for their practice even in instances of system or technology failure." Read that twice. The machine fails, and the human is still on the hook. This spring an ANA think tank landed in the same place: support, never replace, with the nurse as the final authority at the bedside.

To be honest, in fields like nursing, the effect may run the other way. Rather than crossing nurses over into new roles, AI could return them to their core: less time filling out and monitoring EHRs, more time on care and presence with patients.

None of this makes the labs wrong. The model may hold for desk work, and desk work is a big share of the economy. The mistake is the universal claim from that podcaster, the "that's where we're all going to end up." Some of us end up there. The nurse or physician at the bedside might not.

So when someone tells you the frontier labs are a year ahead, ask a year ahead of what. They are a year ahead on knowledge work. On care work, perhaps not so much.

OpenAI's agents hacked second firm during model testing — It was a bad week if you worry about AI and cybersecurity. OpenAI confirmed that its AI agent system compromised a second company during internal testing this month, reaching a Modal Labs customer's exposed endpoint after it broke into Hugging Face's systems — an episode CEO Sam Altman says has forced OpenAI to pause model training. The disclosure that its models, including an unreleased one, "went rogue" and hacked real firms has rattled the field: more than 1,100 frontier-lab employees, including OpenAI's chief scientist and an Anthropic co-founder, signed a letter this week urging governments to deliberately "pace the frontier" of automated AI development.

Investigating three real-world incidents in our cybersecurity evaluations — Following OpenAI's Hugging Face disclosure, Anthropic reviewed 141,006 of its own evaluation runs and found three cases where Claude models, running cybersecurity tests, reached the open internet through a misconfigured test environment and gained unauthorized access to three real organizations' production systems. Anthropic calls it more a testing-harness failure than a rogue-AI one (the models had been wrongly told they had no internet access), and notes its newest model stopped once it realized the targets were real, while an older model kept attacking even after recognizing the systems were live.

Which health system departments are requesting AI the most? — Becker's asked health-system tech chiefs which departments are clamoring loudest for AI, and the answer tracks data readiness more than function: revenue cycle and supply chain lead at Rush, clinicians themselves are now "pulling" rather than being pushed at Hospital for Special Surgery, and HR tops the list at Allina Health. Leaders also flagged a cooling on EHR-embedded generative AI, where clinicians have been disappointed by a one-size-fits-all "voice" when the tools draft patient messages.

The AI Future Is for Everyone — In a WSJ op-ed, Meta CEO Mark Zuckerberg argues that "superintelligence" should be distributed to everyone rather than concentrated in a few institutions, casting individual empowerment and "balance of power" as the real foundation of AI safety — an implicit rebuke of rivals' doom-tinged case for centralized control. He contends widely distributed AI will yield more jobs and entrepreneurship, not less, and points to gains like faster drug discovery, though the open-access framing also tracks Meta's own open-model business strategy.

Introducing MAI-Cyber-1-Flash inside MDASH — Microsoft unveiled a new in-house AI model built to hunt down and fix security holes in software, which it says matches or beats rival models from Google and OpenAI on a standard industry test while costing about half as much. The pitch: as attackers increasingly use AI to find weaknesses at scale, defenders need cheap, always-on AI that spots and patches flaws in real time instead of the old "scan occasionally, patch eventually" approach — a shift with direct stakes for hospitals running aging, hard-to-secure systems.

Advancing the price-performance frontier with GPT-5.6 — OpenAI cut prices sharply on its GPT-5.6 models, with its cheapest tier (Luna) dropping 80%, part of a push to make high-volume AI cheap enough to run across everyday enterprise work. The more striking detail for anyone tracking where this heads: OpenAI says GPT-5.6 itself rewrote and improved the production code that runs it, within a human-led process, cutting the cost of serving the model by about 20%, a feedback loop where AI is starting to make AI cheaper.

Blue Shield of California sister company debuts AI copilot for health plan customer service reps — Stellarus, a health-tech company spun out of a Blue Shield of California restructuring, launched CSR Chat, an AI copilot that feeds health-plan customer service reps live guidance during interactions, including policy details and recommended responses, to make service faster and more consistent. It's the first commercial product in the company's Compass suite, built on its AtlasIQ data platform, with more member- and provider-facing AI capabilities promised in the coming months.

A new wave of AI is reshaping hospital-at-home referrals — Mass General Brigham, UMass Memorial Health, and University of Utah Health are rolling out AI that scans electronic health records in real time to rank which patients qualify for hospital-at-home care, sparing clinicians from combing through thousands of charts — UMass says its tool will let nurses screen less than half as many patients to find each candidate. The systems frame it as a way to grow their programs, cut readmissions, and free up beds, while stressing that clinicians make the final referral and patients still must agree to be treated at home.

Why Aidoc is taking a generative AI device to the FDA — In a MedTech Dive Q&A, Aidoc CEO Elad Walach detailed First Read, a generative-AI tool — granted FDA breakthrough device designation in June — that reads chest X-rays, flags more than 100 conditions, and drafts a full report for radiologist review, which he calls the company's first end-to-end AI diagnostic. Walach frames radiologists as staying "in the driver's seat" and names automation bias and deskilling as the biggest risk, to be mitigated by workflows that force continuous human oversight.

RWJBarnabas Health and Rutgers Researchers Find AI Tool Helps Detect Patient Deterioration Earlier, Reducing Hospital Deaths — A study of 23,132 high-risk patients across 11 RWJBarnabas Health hospitals, published in NEJM AI, found that deploying Epic's AI-powered Deterioration Index, which rescores patients every 15 minutes and auto-alerts rapid response teams, coincided with in-hospital deaths among high-risk patients falling from 23.1% to 18.6% (an 18% drop in risk-adjusted odds). The authors credit the coordinated rollout rather than the model alone, with senior author Stephen O'Mahony putting it bluntly: "The mortality benefit was not produced by an algorithm but by the partnership around the algorithm."

States have passed more than a dozen new laws regulating AI in health care — A new Transparency Coalition tally finds states enacted over a dozen laws in 2026 to rein in health-care AI, concentrated on two fronts. Seven states now bar insurers from using AI as the sole basis to deny prior-authorization requests, and five states prohibit AI chatbots from offering therapy in place of a licensed mental-health professional.

Smart use of AI can improve American healthcare — In a Washington Times op-ed, Rep. Morgan Griffith (R-Va.), chairman of the House Energy & Commerce Health Subcommittee, makes the case for a "balanced embrace" of AI in health care — faster disease detection and lighter administrative load, paired with mandatory human oversight. His guardrails echo the emerging state consensus: a clinician should review AI-driven coverage denials and a professional should oversee mental-health chatbot conversations, so AI assists rather than replaces the clinical workforce.

Rep. Balint Introduces Legislation to Protect Kids from Addictive AI Chatbots — Rep. Becca Balint (D-Vt.) introduced the Addictive Design Act, which would bar AI chatbot companies from serving minors bots with engagement-maximizing "addictive design features" — persistent memory, conversations running past two hours, or impersonating a real person — and fund research into the mental-health risks of youth chatbot use. Endorsed by the American Psychiatric Association, the American Foundation for Suicide Prevention, and the National Association of Social Workers, the bill would authorize civil penalties of up to $5,000 per affected minor and require tight data-minimization for any age verification.

Oz Says AI Implemented In Current Billing System Will Be Inflationary, Suggests Changes — CMS Administrator Mehmet Oz argued that layering AI onto Medicare's current volume-based billing would be inflationary — paying clinicians to "see more people, generate more billing" — and floated shifting to outcomes-based payment, which he said would instead make AI deflationary by cutting unnecessary, low-quality care. The remarks came as CMS launches its voluntary ACCESS model, a 10-year demo with roughly 200 companies in its first cohort that pays higher Medicare rates for AI-driven chronic-care tools that improve outcomes — though stakeholders warn the payments are too low and success metrics remain undefined.

The FDA was all in on AI. Will that change? — With FDA Commissioner Marty Makary — the agency's leading AI champion — plus its chief AI officer and acting CIO all departing, BioPharma Dive reports growing uncertainty over whether the FDA's centralized, agency-wide AI push keeps its momentum. Acting commissioner Kyle Diamantas calls AI a top priority and no change is expected to how the FDA governs industry's use of AI, but the agency's internal build-out — including the Elsa review-assistant tool and efforts toward greater transparency — could stall or fragment back into department-by-department approaches.

Husted, Kim lead bipartisan bill to protect children from AI companion chatbots — Sens. Jon Husted (R-Ohio) and Andy Kim (D-N.J.) introduced the bipartisan CHAT Act 2.0, which would bar AI companion chatbots from harmful interactions with minors — encouraging self-harm, generating sexual content, impersonating humans or licensed professionals, or engaging in romantic or emotionally manipulative exchanges. It sets a tiered, risk-based framework across educational, companion, and health chatbots; requires age assurance and parental notification if a minor expresses suicidal ideation; bars AI from providing crisis counseling to minors; and hands enforcement to the FTC and state attorneys general.

What Happens When Patients Trust AI Over Their Doctor? — As patients increasingly arrive at appointments having already consulted an AI chatbot, health law attorney Meghan O'Connor argues that liability when that advice goes wrong is still unsettled and will take years of litigation to sort out across developers, patients, and providers. Her practical guidance for clinicians: silence is the riskiest response — once a patient raises AI-sourced information, correct it, document the conversation, and treat it like any other patient-reported claim that conflicts with clinical judgment.

Claude chats and workspaces turn up on Google, revealing avoidable privacy flaw — After a viral Reddit thread, Cybernews reports that Claude conversations and artifacts shared via Anthropic's "anyone with a link" option have been indexed by Google and other search engines, with exposed material reportedly including a clinical-trial record listing patients' names, ages, genders, and ethnicities. Anthropic notes that shared links capture only messages sent before sharing and that publishing an artifact carries a search-visibility warning, but developers argue "shareable" shouldn't default to publicly indexable — echoing a near-identical ChatGPT episode a year ago.

When physicians and AI work together, who is accountable? How to lay out medical liability — In a Nature Comment, a team including cardiologist Eric Topol proposes a seven-level framework — graded by autonomy, automation, and operational scope, much like the tiers used for self-driving cars — to clarify who bears responsibility when AI-assisted care goes wrong. As tools advance from advisory assistants toward "black-box" systems that diagnose and treat with little human input, the authors warn that patients could fall into "liability gaps" where no one clearly broke a rule, and urge regulators, courts, and hospitals to act now, before highly autonomous tools are entrenched.

Could using AI erode a doctor's ability to think? — An AAMC News feature examines a growing worry as physician AI use tops 80%: that leaning on tools for notes, summaries, and diagnoses may quietly erode clinicians' critical thinking, creativity, and hard-won skills. The evidence is early but pointed — a Lancet study found endoscopists' tumor-detection rates fell from 28.4% to 22.4% after they grew used to AI assistance and then worked without it — and experts converge on "do the work first, consult AI second," plus designing tools that periodically test the user rather than just hand over answers.

HMH Becomes First Health System to Earn Responsible Use of AI in Healthcare Certification — Hackensack Meridian Health says it's the first U.S. health system to earn the Joint Commission's new Responsible Use of AI in Healthcare Certification, which evaluates governance, patient-data privacy, bias and risk assessment, ongoing performance monitoring, patient transparency, and staff training. The credential — from the nonprofit that accredits most U.S. hospitals — signals an emerging bar for how health systems are expected to deploy AI while keeping clinicians in the loop rather than replacing them.

Clinical chatbots are taking medicine by storm. Should doctors trust them? — A STAT investigation traces how clinical AI tools from OpenEvidence, Doximity, and UpToDate — used by hundreds of thousands of U.S. doctors — stack up against general chatbots, and how fast benchmark results get weaponized in a multibillion-dollar market. An NYU Langone paper in Nature Medicine found the clinical models underperformed general LLMs from Google, Anthropic, and OpenAI on standard tests, but a separate Stanford-led safety benchmark found the reverse on patient harm — severe errors in 2.9%–5.4% of clinical-tool cases versus 8.9%–24.6% for general models — a split that underscores how no single test settles whether doctors should trust them.

Americans Cool Toward AI — A new Bentley University–Gallup survey finds Americans souring on AI even as familiarity grows: 39% now say AI does more harm than good (up from 31% a year ago), just 27% trust businesses to use it responsibly, and 79% expect it to cut U.S. jobs over the next decade. The cooling is sharpest among adults under 30, and while more people now see AI as roughly on par with humans for tasks like giving medical advice, a majority still rate it worse than people across every task surveyed.

Toward a test of medical AI superintelligence — In a Nature Medicine comment, a group of academic and industry researchers argues that current claims of "medical superintelligence" from Microsoft and OpenEvidence rest on misleading benchmarks, and proposes a harder standard (the Medical AI Superintelligence Test, or MAST) that would require AI to beat even top clinical teams on real, long-horizon tasks. Their sharpest warning for health systems: as tools improve, human-in-the-loop review is sliding toward human-on-the-loop, and forcing a clinician to check every AI output "creates the appearance of safety without delivering it."

AI Readiness Starts with Solving Healthcare's Data Fragmentation Problem — On a Verato-sponsored webinar, data leaders from SCAN Health Plan and the Alliance of Community Health Plans argued that data fragmentation, not technology, is the real blocker to AI readiness for payers — showing up as documentation gaps, undercoded acuity, and inaccurate risk adjustment. Their prescription: shared data definitions, consistent governance, human-in-the-loop review, and deterministic data lineage that can trace any AI output back to its raw inputs.

5 Benchmarks for Evaluating AI Tools — The Children's Hospital Association published five procurement benchmarks to help hospitals, schools, and community partners vet AI tools for youth mental health — a category it argues is fundamentally different from wellness apps because the tools interact directly with kids in distress. The criteria: demonstrated outcomes (not just engagement), real crisis-escalation safeguards beyond a hotline number, HIPAA/FERPA-grade privacy, integration with existing clinical and school workflows, and access/equity by design — with mental-health professionals, not just IT or administrators, brought into the buying decision.

Till next time,

BC