The Bandwidth: Decoding Health + AI

A New York Times story published this week sent a chill through the scientific community. And it should. Researchers shared transcripts with the Times in which leading AI chatbots provided detailed, actionable guidance on assembling biological weapons and deploying them in public spaces. One biosecurity expert at Stanford described going cold at his laptop as a chatbot outlined a plan to maximize casualties. Another researcher watched ChatGPT outperform 94% of expert virologists on difficult laboratory protocol questions.

As the story notes, this comes at a moment when federal biosecurity staffing has been reduced, biodefense budgets have been cut by nearly half, and the administration has signaled a hands-off posture toward AI oversight. That combination should get people to sit up and pay attention.

But fatalism is not the answer.

The Times piece, to its credit, resists a complete doomsday framing. The scientists who shared these transcripts did so not to sound an alarm and shut down the technology, but to pressure companies into building safer systems.

There is a real cost to over-indexing on the threats. As discussed in last week’s newsletter, AI may accelerate drug discovery, compress clinical trial timelines, and unlock biological insights. A response calibrated only to risk, one that treats the technology as inherently dangerous rather than as something we are still shaping, could curtail exactly the scientific advances that make the investment worthwhile.

Yes, the risk is real. But, so is the promise. Our response has to accommodate both.

We designed this. We can design it differently.

Philosopher and ethicist Carissa Véliz, speaking on the AI-Curious podcast, made a point this week that sounds simple but may often get overlooked: this technology is not God-given. It’s not out of our control. We (humans) are designing it, and we can design (and deploy) it differently. How much power these systems have, what we put into them, and what decisions we allow them to shape. These are our choices. They are not inevitable features of the technology itself.

Greg Brockman (a co-founder, president, and chairman of OpenAI) made the point concretely in a recent conversation on The Knowledge Project. The answer to transformative technology, he argued, is not just safer models. It is a more resilient society. Engines gave us cars, but also seatbelts, traffic laws, and cities redesigned around how the technology actually works. Electricity gave us power, but also safety standards, utility regulations, and infrastructure built to manage the risks. AI will be the same. The work is not just in the model. It is in everything we build around it. This, of course, rests on the idea that policymakers can come to an agreement on a path forward regarding regulations.

The healthcare stakes are specific.

The biological risk story is the sharpest version of a broader question medicine and science have been wrestling with all year: What does it take for a powerful technology to be trustworthy? Better evaluation frameworks. Clearer governance. Accountability structures that outlast any single product or deployment. Training for professionals and consumers. All of these things feel like critical components for societal resilience.

We have built that infrastructure before, around technologies that also carried real risks and real promise. Can we do it again?

Utah's AI Prescription Renewal Pilot Continues Despite Medical Board Objections This has been a fascinating one to watch the past few weeks. Here is the latest. The Utah Medical Licensing Board called for immediate suspension of the state's Doctronic AI prescription renewal pilot, saying it was never consulted before launch and that autonomous prescription renewal bypasses clinical oversight that exists for patient safety. The Utah Department of Commerce rejected the suspension request, while Doctronic argued the relevant safety benchmark isn't AI versus a perfect system but AI versus a status quo in which 125,000 Americans die annually from medication non-adherence — a debate that remains unresolved alongside questions about FDA jurisdiction.

AI Can Cost More Than Human Workers Now IT budgets are buckling under AI compute costs that in some cases now exceed employee salaries — Uber's CTO burned through his entire 2026 AI budget on token costs alone — raising a question the AI hype cycle has largely avoided: whether the economics actually pencil out. Worldwide IT spending is projected to hit $6.31 trillion in 2026, up 13.5%, driven almost entirely by AI infrastructure and cloud services, but pressure to show returns is mounting as labs raise prices and shareholders start asking harder questions.

Beijing Blocks Meta's Acquisition of Chinese AI Startup Manus China's canceled Meta's $2 billion acquisition of Manus — hailed last year as China's "next DeepSeek" — on the grounds that AI talent and intellectual property shouldn't be acquirable by U.S. entities, extending Beijing's strategic controls beyond semiconductors into AI itself. The decision signals that relocating overseas, as Manus did by moving headquarters to Singapore, will not shield Chinese AI companies from regulatory scrutiny amid intensifying U.S.-China competition over frontier technology.

J&J Says AI Has Cut Drug Development Lead Time in Half Following last week’s newsletter about AI advancements in drug development, Johnson & Johnson's CIO says AI has reduced lead optimization time by 50%, already accelerating development of two compounds in oncology and immunology, while a separate AI application has compressed clinical trial report preparation from 700 hours to roughly 15 minutes. The company is framing AI as an additive skill for its workforce rather than a replacement.

OpenAI Missed Key Revenue and User Targets Ahead of Planned IPO The Wall Street Journal reports that OpenAI fell short of internal goals over the past several months — including a target of 1 billion ChatGPT users by end of 2025 — while its CFO has privately warned employees the company may struggle to meet future compute bills despite a $122 billion raise. OpenAI's board is now scrutinizing spending habits as the CFO questions whether a 2026 IPO is realistic, even as the company released a new model and insists reports of internal discord are overstated.

OpenAI May Be Developing a Smartphone Built Around AI Agents Industry analyst Ming-Chi Kuo reports that OpenAI is exploring a smartphone built with MediaTek and Qualcomm that would replace traditional apps with AI agents — giving the company direct access to hardware and user data outside Apple and Google's app ecosystem restrictions. Specifications are expected to be finalized by end of 2026 or early 2027, with mass production targeted for 2028; OpenAI has not confirmed the plans.

Jury Seated in Musk v. Altman Trial A nine-person jury was seated in federal court in Oakland for the liability phase of Elon Musk's lawsuit against OpenAI and Sam Altman, with Musk alleging the company reneged on its commitment to remain a nonprofit and follow a charitable mission. Of the original 26 claims, only two remain — unjust enrichment and breach of charitable trust — and the jury's verdict will be advisory, with the judge making the final determination. The liability phase is expected to wrap by May 21.

OpenAI and Microsoft Amend Partnership Agreement OpenAI and Microsoft have restructured their partnership, with Microsoft remaining the primary cloud partner but OpenAI now free to serve products across any cloud provider. Microsoft's IP license becomes non-exclusive through 2032, revenue share payments from OpenAI to Microsoft continue through 2030 at a capped rate, and Microsoft stops paying revenue share to OpenAI. The restructuring gives OpenAI meaningful flexibility ahead of its anticipated IPO while preserving Microsoft's position as a major shareholder.

Zuckerberg-Backed Biohub Commits $500M to AI Biology The Chan Zuckerberg Biohub is committing $500 million over five years — $400 million for its own research and $100 million to spur outside work — toward building accurate AI simulations of the human cell, with the long-term goal of curing all human disease. The effort, partnering with Nvidia, the Allen Institute, and the Human Cell Atlas, bets that scaling cellular datasets by an order of magnitude beyond current levels will unlock exponentially more useful predictive models.

Why Healthcare AI Still Can't Scale — and How Nvidia and Hoppr Are Trying to Fix It Nvidia and Hoppr are repositioning away from standalone AI applications toward a shared infrastructure layer — an AI foundry that lets health systems, radiology practices, and device companies fine-tune and deploy their own imaging models without starting from scratch or purchasing massive datasets. The partners are betting the real barrier to healthcare AI adoption isn't model quality but deployment infrastructure, though whether the approach drives clinical adoption or adds another layer of complexity remains to be seen.

Why Conversations Around Health AI May Be Evolving Beyond Hype STAT's AI Prognosis newsletter synthesizes a growing body of evidence that challenges core assumptions about clinical AI: a January 2026 Ohio State survey found fewer Americans are open to AI in their care than in 2024; an NEJM AI study found physicians given flawed AI suggestions scored 14 percentage points lower on diagnostic accuracy than controls, with the effect worse among more experienced clinicians; and a VA study rated AI-written notes inferior to human-written notes on 10 of 10 quality metrics. The throughline, captured in a Nature Medicine editorial published last week: the field needs evidence that AI meaningfully changes clinical decisions and that those changed decisions improve outcomes.

What Health Care Leaders Have Learned From Deploying AI At the Asembia AXS26 Summit, executives from Infinitus Systems and Optum Rx offered a take on where healthcare AI is and isn't delivering: 69% of health system executives call AI their top priority in 2026, yet only 8.3% have AI operating in production — compared to 90%+ in professional services. Optum Rx reported measurable gains in prior authorization turnaround and a 20% reduction in live operator calls through AI-enabled self-service, while cautioning that patient-facing automation remains "proceeding with caution" territory given the stakes of error in clinical contexts.

Maine Governor Vetoes Nation's First Statewide Data Center Moratorium Gov. Janet Mills vetoed legislation that would have imposed an 18-month ban on new data center construction statewide, citing the economic threat to a proposed redevelopment of a defunct paper mill. The bill passed with unanimous Democratic support — not enough to override her veto — as other states weigh similar restrictions on AI infrastructure.

Justice Department Joins xAI Lawsuit Against Colorado AI Law The DOJ intervened for the first time in a case challenging state AI regulation, joining xAI's lawsuit against Colorado's algorithmic discrimination law — which requires developers to disclose information about AI used in high-stakes decisions like lending and hiring — on the grounds that the law's carveout for diversity-advancing algorithms is unconstitutional. The move is a direct signal of the Trump administration's willingness to use federal litigation to preempt state AI laws it views as inconsistent with its policy agenda, with Colorado having been the only state law specifically named in last year's executive order.

FDA Seeks Input on AI-Enabled Early-Phase Clinical Trials Pilot The FDA issued a request for information on a proposed pilot program to assess how AI can improve efficiency, speed, and decision-making in early-phase clinical trials — targeting what the agency calls a "critical bottleneck" in drug development characterized by high uncertainty and inefficient go/no-go decisions. The pilot would be guided by the NIST AI Risk Management Framework, signaling the agency's intent to pair accelerated drug development with structured AI governance.

Bipartisan House Bill Advances AI Standards and Research — With Limited Health Provisions Reps. Ted Lieu and Jay Obernolte introduced the American Leadership in AI Act, building on the December 2024 House AI Task Force report, with provisions to codify AI standards infrastructure at NIST, expand research resources, and establish an NIH grant program for generative AI in healthcare. Notably absent: the task force's own recommendations on FDA post-market evaluation of AI, AI liability, and Medicare reimbursement — and the bill sidesteps federal preemption of state AI laws entirely, a priority for several healthcare trade groups.

White House Developing Guidance to Bring Anthropic Back Into Federal AI Ecosystem The White House is drafting executive action that would allow federal agencies to work around the Pentagon's supply chain risk designation for Anthropic and onboard its latest model, Mythos — a move sources describe as an attempt to "save face and bring em back in" after the administration previously blacklisted the company as a security risk. The core dispute remains unresolved: Anthropic has refused to sign an agreement permitting use of its models for "all lawful purposes," insisting on prohibitions against mass domestic surveillance and fully autonomous weapons — conditions OpenAI and Google accepted, though both claim their deals respect the same limits.

AI in Healthcare Translation and Interpretive Services: What Covered Entities Need to Know A legal briefing from Sheppard Mullin maps the emerging compliance landscape for AI-powered translation tools in healthcare — finding that Section 1557 of the ACA effectively requires human review of AI-generated translations in most clinical contexts, while state laws like Texas's TRAIGA and California's AB 3030 add disclosure requirements on top of an already ambiguous federal framework. The common thread: AI may assist with translation, but cannot yet substitute for a qualified human interpreter where meaningful language access is required.

Your AI Agent Is a Nobody. And That's a Problem. Snowflake's product blog makes a governance case worth taking seriously: as AI agents move into production, most lack verifiable identity — defined rights, defined scope, and a persistent audit trail — leaving compliance teams unable to reconstruct what an agent accessed, under whose authority, or whether it stayed within approved boundaries. In regulated industries like healthcare, where an agent's recommendation could surface in a dispute months later, that accountability gap isn't theoretical.

Your Doctor Is Using AI to Take Notes. What Could Go Wrong? With roughly 30% of U.S. physicians now using AI scribes, the NYT walks through what patients should know — from consent practices that vary by state, to accuracy concerns including a 2024 study finding lower transcription accuracy for Black patients than white patients, to the finding that simulated patient encounters produced an average of three potentially serious errors per note. The practical takeaway for patients: you can say no, and checking your own medical record for accuracy is advisable regardless.

Sam Altman Publishes OpenAI's Operating Principles In a post on OpenAI's website, Altman outlines five principles guiding the company — democratization, empowerment, universal prosperity, resilience, and adaptability — framing OpenAI's mission as putting "truly general AI in the hands of as many people as possible" rather than concentrating power among a handful of companies. The document is notable for its timing: published the week OpenAI faces trial over its nonprofit origins and scrutiny over its financial trajectory, it reads as both a values statement and a public positioning exercise.

Google Clinical Director Says AI Can Be a 'Bridge' for People in Mental Health Crisis In an interview with STAT, Google's clinical director Megan Jones Bell defends the company's decision not to disengage Gemini when users show signs of crisis — arguing that cutting off a vulnerable person reaching out for the first time could do more harm than good. The interview comes amid active litigation: a California lawsuit alleges Gemini drove a man to suicide.

AMA Calls for Enforceable Protections Against AI Deepfake Impersonation of Physicians The AMA released a seven-principle policy framework to protect physicians against unauthorized AI-generated deepfakes, calling for opt-in consent requirements, mandatory labeling of synthetic physician content, shared platform liability, and federal enforcement authority. AMA CEO John Whyte framed physician deepfakes as "a public health and safety crisis" — not just fraud — given their potential to steer patients toward unproven treatments and erode trust in evidence-based care.

AI in Clinical Decision-Making: Applications, Challenges, and Future Directions A narrative review in Cureus synthesizes the current state of AI across clinical domains — imaging, risk prediction, precision medicine, surgical support, and mental health — and finds a consistent gap between strong algorithmic performance in narrow tasks and proven real-world clinical impact. The authors identify prospective validation, equity-sensitive evaluation, and workflow integration as the priorities that will determine whether clinical AI delivers on its promise.

In Real-World Test, AI Model Outperformed ER Doctors at Diagnosing Patients A study published in Science by researchers at Harvard Medical School and Beth Israel Deaconess Medical Center found that an OpenAI reasoning model outperformed two experienced physicians at diagnosing patients using actual emergency department records — including messy, real-world data rather than curated test cases. The authors are careful to note the model relied on text alone, that ER diagnosis is a narrow slice of clinical care, and that the findings do not support replacing physicians — but they do make the case for rigorous, prospective trials to determine how AI actually affects clinical outcomes.

AI in Genomic Medicine: Dispelling Three Myths A comment in npj Genomic Medicine from an NIH physician geneticist pushes back on three reassuring assumptions about clinical AI: that clinicians who use AI will simply outcompete those who don't; that strong benchmark performance will translate automatically to real-world benefit; and that patient preferences for human providers will slow AI adoption. On the first point, he cites evidence that AI alone outperforms human-AI combinations — suggesting the "AI as pocket tool" framing may be a short-lived comfort.

Who Owns My Health Data? A Nature Medicine news feature maps the growing tension between open science and data sovereignty in biomedical AI — finding that decades of cross-border data sharing are giving way to national walls, as the US restricts Chinese access to genomic databases, Europe moves to limit international downloads, and China maintains asymmetric openness. The consequence for medicine is concrete: AI models trained on sequestered national datasets overfit to local demographics, and without international validation, those biases often surface only after clinical harm has occurred.

Mayo Clinic AI Detects Pancreatic Cancer Up to Three Years Before Diagnosis A Mayo Clinic study published in Gut found that its REDMOD AI model identified 73% of prediagnostic pancreatic cancers on routine CT scans at a median of 16 months before clinical diagnosis — nearly double the detection rate of specialists reviewing the same scans without AI. At earlier time points, the advantage was even greater: in scans taken more than two years before diagnosis, the AI identified nearly three times as many cancers that would otherwise have gone undetected. The model is designed to run automatically on scans already obtained for other reasons, with no manual preparation required.

The Growing Use of AI in Health Care and Implications for Disparities A KFF analysis finds that while about a third of U.S. adults now use AI chatbots for health information, two-thirds don't trust them to provide reliable health information — and the research on bias gives them reason for caution. A systematic review of 30 studies found a significant association between AI utilization and exacerbated racial and ethnic disparities, including a suicide prediction model that detected 62% of suicides among White patients but only 10% among Black patients. The brief notes that careful design, inclusive data, and diverse development teams can help — but flags that the Trump administration has rescinded Biden-era equity requirements for AI and is actively challenging state anti-bias laws.

How CEOs and Large Enterprises Can Unlock the Power of AI Apply Digital CEO Ali Alkhafaji argues the gap between AI leaders and laggards isn't access to technology — it's the quality of the question being asked, with leaders asking what AI makes possible rather than how to do existing work faster. His core prescription for large organizations: treat workforce transformation as the primary investment, not the support track for a technology deployment.

Anthropic Launches Claude Connectors for Creative Industry Tools Anthropic released a set of connectors integrating Claude with creative software including Adobe Creative Cloud, Autodesk Fusion, Blender, Ableton, and Splice — enabling natural language interfaces for 3D modeling, audio production, and design workflows. The release is aimed at helping creative professionals offload repetitive production tasks and extend their existing toolsets.

Locked, Stocked, and Losing Budget: AI Vendor Lock-In Bites Back Opinion | A Zapier survey of 542 U.S. executives found that nearly 90% believed they could switch AI vendors within four weeks — but only 42% of organizations that actually attempted a migration report it went smoothly. The piece argues the gap reflects how deeply AI gets embedded into APIs, workflows, and institutional memory, compounded by vendors now raising prices after years of loss-leader positioning: OpenAI's flagship model pricing jumped from $1.25 to $5.75 per input token, while Anthropic shifted to usage-based enterprise pricing that experts say could double or triple costs for heavy users.

👉 May 4–5 — AI in Medicine Conference (AIIM 2026), Boston, MA

👉 May 7–8 — NBER Conference on AI in Healthcare, Cambridge, MA

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

👉 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