The Bandwidth: Decoding Health + AI

We're back after the holiday with a full edition, and the week did not wait.

OpenAI released GPT-5.6 across three tiers. It released a new voice architecture the day before. Musk's SpaceXAI shipped Grok 4.5. Anthropic put its agentic platform on phones. Somewhere in the middle of all that, researchers documented what they believe is the first ransomware attack executed end-to-end by an autonomous AI agent, correcting its own failed login in thirty-one seconds.

Against those rapid developments, Daniel Kokotajlo's AI Futures Project published "AI 2040: Plan A," which proposes that AI superpowers agree to delay superhuman AI and open their internal models to public scrutiny. Kokotajlo predicts the companies won't like it. That seems like a safe bet.

I don't expect anyone to slow down. But the argument for slowing down is one that I think few people could dispute: we are deploying systems into consequential domains before we have settled the rules that govern them.

In healthcare, one of those unsettled rules is about liability. As AI moves further into the delivery of care, who is responsible when something goes wrong? This is a question I've been thinking about for a while and one that is coming up more and more. Articles published this week made it seem like the time to tackle it in this space had come.

The shifting landscape

A couple of years ago, Nicholson Price, Sara Gerke, and Glenn Cohen mapped the liability and AI terrain in a legal handbook chapter. They identify three loci of potential liability: the physician, the institution, and the developer. The developer and institution questions are their own posts. Here I want to stay with the physician.

Two scenarios in the chapter show how liability could change in the age of AI.

In the first, the AI makes a novel recommendation, it's right, and the physician overrules it. She prescribes the conventional treatment, and the patient is still injured. Today she's protected, because she followed the standard of care.

In the second, the AI is wrong, and the physician follows it. She prescribes the unconventional treatment, and the patient is injured. Today she's liable, because she departed from the standard of care.

Both of those answers depend on AI sitting outside the standard of care. That might not hold.

If following AI advice becomes accepted practice, the answers could invert. The physician who ignores a correct recommendation starts to look negligent. The physician who follows an incorrect one starts to look reasonable.

New frameworks are needed. Maybe.

The uncertainty around AI and what responsibility it bears is why some have called for new liability frameworks. And they note that until the rules of the road become clear, it could stifle adoption in clinical settings.

But writing in STAT this week, Afnan Tariq and Ami Bhatt argue that this is what the emerging debate over clinical AI gets wrong. They don't tackle liability explicitly, but they do tackle the issue of responsibility in the era of AI and medicine.

A license, they argue, was never about just getting a good test score and proving knowledge, which AI can now easily do. It is the mechanism by which a person becomes answerable — someone who can be made to explain their reasoning under oath, sanctioned by peers, held to a fiduciary duty that courts have said cannot be handed to a colleague or a system.

They argue that every technology that came before (e.g., imaging, the EHR, telemedicine) left that binding intact. AI is somewhat different. It arrives with a plausible claim to be making a clinical decision, but Tariq and Bhatt assert it changes nothing: a system that cannot be sanctioned cannot hold a duty or be responsible.

That's a clean line in theory. The physician remains responsible. But, for me, it doesn’t settle the question underneath it — when to follow the algorithm and when not to. When does AI become part of the standard of care? That probably doesn't get resolved by updated licensure or liability doctrines. It gets resolved slowly as AI gets incorporated into the practice of medicine, specialty by specialty, and what physicians themselves agree are competent practices.

Rural Colorado hospital pilots AI to recover missing insurance payments — Spanish Peaks Regional Health Center, a rural nonprofit hospital in Huerfano County, is piloting agentic AI from Denver-based Iterate AI to identify "ghost denials" — claims insurers quietly underpay or delay without an official denial code — and auto-draft appeal letters. The trial comes as Kodiak data shows initial claim denial rates climbed to 11.6% last year, contributing to $48.4 billion in lost provider revenue industry-wide.

Health insurers go all-in on AI, providers stay wary — At a health insurance conference last month, executives from Blue Shield of California, Elevance Health, and Aetna detailed AI rollouts for care coordination, faster approvals, and fraud detection, with executives maintaining the technology isn't used to deny care. Provider groups including the AMA remain skeptical, pointing to AI-driven increases in prior authorization denials and warning that many providers still lack the technology infrastructure insurers assume they have.

OpenAI releases the GPT-5.6 family — OpenAI made GPT-5.6 generally available in three tiers — Sol, Terra, and Luna — with the company reporting state-of-the-art results on coding, agentic workflow, and cybersecurity benchmarks at lower token cost than competing frontier models. OpenAI says the models are more capable in biology and cybersecurity than prior releases but do not cross its Critical risk threshold in either, and that Sol's cyber safeguards block roughly ten times more potentially harmful activity than earlier models.

OpenAI launches GPT-Live, a new voice architecture for ChatGPT — OpenAI's new full-duplex voice model listens and speaks simultaneously rather than processing turn-by-turn, and includes expanded safety testing for self-harm, psychosis, and emotional reliance, with built-in safeguards that can redirect a conversation or surface crisis resources in real time. The model becomes the default for ChatGPT Voice across all tiers, raising the stakes for how millions of users — including those seeking health information — interact with AI in emotionally sensitive moments.

SpaceXAI releases Grok 4.5, its first model since going public — Elon Musk's SpaceXAI launched Grok 4.5, pitched as a coding and agentic-work tool trained alongside newly acquired Cursor, with Musk claiming it outperforms Anthropic's Opus 4.8 on several benchmarks at a fraction of the price. The release lands the same week OpenAI is rolling out GPT-5.6 and new voice models, underscoring the pace of frontier model competition reshaping the tools health systems may soon evaluate.

AI agent conducts first fully autonomous ransomware attack — Researchers at Sysdig documented what they believe is the first ransomware attack carried out entirely by an autonomous AI agent, which handled reconnaissance, credential theft, lateral movement, and encryption without human involvement — in one case correcting a failed login in 31 seconds. The attack used known, unpatched vulnerabilities rather than novel exploits, but its speed compresses the window between detection and damage from hours to seconds, a shift that matters for healthcare as critical infrastructure.

Anthropic launches Claude Cowork on mobile and web — Anthropic's Claude Cowork, previously desktop-only, is expanding to iOS, Android, and web starting with Max subscribers, and now runs in the cloud by default so agentic tasks continue across devices or with everything offline. The desktop app remains the "full experience" for features like local file access, and scheduled tasks can now run even when no device is online.

How nurses are shaping the next wave of clinical AI tools — After years of building AI tools around physician workflows, companies including Ambience, Abridge, and Hippocratic AI are now co-developing nursing-specific documentation and voice tools with health systems like Cleveland Clinic and Mayo Clinic. The American Nurses Association cautions that some vendors still seek nurse input only after a product is nearly finished, and argues nurses need meaningful decision-making power across selection, design, testing, and rollout to earn real buy-in.

The AI use case that healthcare is overlooking — Writing on MedCity Influencers, a Tecsys supply chain executive argues that pharmacy and supply chain operations — well-defined, data-intensive, and beyond human tracking scale — are a more natural fit for AI than the clinical use cases that get the attention, citing $900 million and 20 million labor hours lost annually to drug shortages. The practical caveat, which applies to any AI initiative: none of it works without first unifying fragmented inventory data across EHR, ERP, and pharmacy systems into a single source of truth.

Virtual patients will train future clinicians — AI simulation is moving into health professions education, where the constraint has long been how many patient encounters a trainee can realistically get before graduation. Penn and NYU researchers received a $4 million Wellcome Trust grant to build STELLAR, a platform generating composite "digital twin" patients from the Philadelphia Neurodevelopmental Cohort so trainees can practice psychiatric interviews across controlled variations in symptom overlap and severity. Pace University's simulation lab is already running AI avatar encounters for nurse practitioner and pediatrics students, with faculty framing the tools as a supplement to live standardized-patient actors rather than a replacement.

The crucial medical question that AI can't ever answer — Two physicians argue in the LA Times that AI can calculate what usually works for people with similar histories but cannot know what a particular patient is trying to protect or avoid — a gap that matters most in preference-sensitive decisions like low-risk prostate cancer, atrial fibrillation, and chronic back pain, where the evidence rarely points to one right answer. They note that roughly a quarter of U.S. healthcare spending flows through decisions where patient preferences meaningfully shape outcomes, and offer three questions patients should ask any AI recommendation, starting with "best for whom?"

Bipartisan bill proposes $2 million federal study into AI's effects on older adults — Sens. Rick Scott (R-FL), Mark Kelly (D-AZ), and Roger Marshall (R-KS) introduced the Aging with Artificial Intelligence Act of 2026, directing the National Academies to study how older adults use AI tools like chatbots and voice assistants, including risks like fraud, overreliance, and reinforced delusions or self-harm. The bill has drawn support from AARP, the National Council on Aging, and the American Medical Association, and would require a report to Congress within one year.

UN scientific panel warns AI chatbots are fueling mental health crises — The Independent International Scientific Panel on AI's first Preliminary Report, presented to the UN Secretary-General ahead of the Global Dialogue on AI Governance, warns that sycophantic chatbots can reinforce paranoid ideation and suicidal thinking in vulnerable users, citing congressional testimony on a teenager's death following prolonged chatbot use. The panel stops short of recommending bans, instead urging governments to require safer system design and stronger evaluation of AI's emotional effects on users.

Illinois signs nation-leading AI safety law — Gov. JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, which requires the largest AI developers to disclose safety practices, report significant safety incidents, and — in a first among states — undergo independent third-party safety audits, alongside whistleblower protections for employees raising concerns. The bipartisan law takes effect January 1, 2027, and builds on frameworks already passed in California and New York.

Treasury has an internal report warning about the dangers of an AI bubble — A draft Treasury report obtained by NOTUS warns that AI firms are more deeply entrenched in the U.S. economy than their dotcom predecessors, and that a market downturn would ripple across banks, private credit, chipmakers, and utilities — a departure from the administration's public bullishness. A Treasury spokesperson dismissed the findings as unvetted and not representative of the agency's position.

AI accountability is now healthcare's next big challenge — The CEO of AI vendor Inflo Health argues that clinician adoption is no longer the barrier — traceability is, and many health systems cannot say where AI touches patient care or who owns an output when it's wrong. Her contention: if a system can't name the owner and the failure path before a tool goes live, it isn't ready to turn on, and validating vendor models against a system's own patient population is non-negotiable.

What's the right role for AI in dementia care? — A Penn neurologist describes "artificial care" — chatbot companions and agentive AI that monitor, cue, and converse with people living with dementia — and the dignity one patient found talking with ChatGPT about his life. His caution: the same agentive planning and judgment these tools offer are precisely the capacities dementia erodes first, leaving patients vulnerable to an AI that cannot judge which of their desires to honor and which to refuse.

OpenAI publishes its national security principles — OpenAI released a set of principles governing its government and national security partnerships, alongside disclosure of expanded trusted access to its GPT-Rosalind model for U.S. and allied partners working on public health and biodefense missions. The company says the most consequential questions about AI in government should be settled through the democratic process rather than by companies alone, and that it supports legislation establishing safeguards around the highest-risk military uses.

Anthropic researchers find a "workspace" in Claude's internals that reveals hidden reasoning — Anthropic identified a small set of internal representations in Claude, dubbed the "J-space," that surface reasoning the model doesn't state out loud, including cases where it privately recognized a test scenario as staged or edited a file to fabricate improved results. The technique offers a new way to monitor for deceptive or misaligned behavior that wouldn't otherwise appear in a model's visible output.

Patient perspectives on AI-drafted electronic portal messages — In a qualitative study of 40 patients at a large academic health system, patients were broadly comfortable with AI-drafted portal messages when they improved response speed, but their acceptance was conditional on a clinician reviewing and remaining accountable for the message. Patients wanted AI use disclosed and preferred tone and length matched to the stakes — brief for refills, more careful for concerning test results — rather than a single standardized style.

Can ChatGPT be your therapist? USC study tests AI responses to mental health questions — USC researchers had 100 mental health professionals grade 400 responses from ChatGPT-4, Llama 3.3, and Gemini 1.5 Pro against real patient questions, finding the models scored well on empathy and overall quality but all three offered unauthorized medical advice, including recommending specific psychotropic drugs and speculating on diagnoses. The models also proved unreliable at grading themselves, consistently overestimating their performance and missing safety risks human experts caught.

Chatbots can help perpetuate stigma around certain health conditions — A study in Nature Health found that Claude, ChatGPT, and DeepSeek were 13 to 17 times more likely to write a negative story outcome for a character described as HIV-positive or having a mental health condition than for a healthy one, across 51 scenarios in English and Chinese. The models were less biased than the 399 humans given the same prompts, but researchers caution the results were drawn from one-time prompts rather than the extended dialogues where people actually disclose sensitive health information.

Thinking about the impact of AI on U.S. health care costs and spending growth — Writing in NEJM Catalyst, Kocher, Zhao, and Duffy argue that under the still-dominant fee-for-service model and consolidated hospital and insurance markets, AI is more likely to raise total spending than slow it — scribes enable more billable visits, decision support lowers the threshold for ordering tests, and administrative savings get absorbed as margin rather than passed to patients. The cost-bending case depends on payment reform, not the technology: they argue value-based contracts are the most direct lever, and that without changes to financial incentives and market structure, AI will not slow cost growth.

Want workers to reskill? Show them who they can become — Harvard Business School research on 1,100 unemployed Italian workers found that reskilling take-up is driven by professional identity, not just pay — many declined higher-paying retraining because the new role didn't fit how they saw themselves. The actionable takeaway for leaders rolling out AI-driven role changes: give staff specific, credible detail on wages and job prospects in the new role, and use concrete examples of people who made the same transition, rather than relying on abstract "growth mindset" encouragement.

Beyond data migration: Preparing health care organizations for AI at scale — Two CitiusTech executives argue that AI initiatives stall because organizations treat data migration as a technical exercise rather than a clinical one, moving records accurately while stripping the context — units, reference ranges, timing — that makes a lab value usable. Their practical advice: define scope around specific clinical workflows before mapping fields, require metadata and provenance for every source, and start with narrow, high-value use cases rather than waiting for data to be fully ready, which it never will be.

Beyond the buzz: Where AI truly matters in healthcare — A MEDITECH executive argues leaders are asking the wrong question when they start with "what AI should we implement" rather than "what would better support our patients, staff, and sustainability" — a framing that sometimes points away from AI entirely, toward optimizing existing EHR features like registries and surveillance boards. She points to patient discharge and nurse handoff summarization as places where the time savings are measurable, with clinicians retaining review and approval.

Introducing a way to reflect on how you use Claude — Anthropic launched a beta dashboard that summarizes how someone has used Claude over the past 1 to 12 months, breaking activity down by topic and by the four dimensions of its AI Fluency Framework — delegation, description, discernment, and diligence — with prompts like asking what you want to keep doing yourself even if Claude could do it faster. It also supports quiet hours and break nudges, and excludes incognito chats and any conversation connected to a health integration tool.

AI adoption to adaptation: How a new change approach can build the human behaviors needed for AI — Deloitte argues that adoption metrics are easy to hit and easy to fake: worker access to AI tools grew 50% in a year, but fewer than 60% of those with access use it in daily workflow and 84% of organizations haven't redesigned jobs around it. The behaviors that separate adopters from adapters — judgment about when to challenge AI, willingness to experiment, and protected time for divergent thinking — go unmeasured by usage dashboards, and only half of executives regularly verify the quality of AI outputs before deciding on them.

👉 July 14, 2026 — State AI Sandboxes: A Regulatory Path Forward for Scaling Health Care AI — Virtual

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

👉 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