The Bandwidth: Decoding Health + AI

If you want to find where AI is most likely to deliver transformational impact in health, start with drug development. It is said that the pipeline from discovery to regulatory approval is one of the most expensive, time-consuming, and failure-prone processes in all of industry. It appears AI may now be compressing it from both ends simultaneously.

The deals on the front end

On the front end, the deals are coming fast. This week, Merck announced a partnership with Google Cloud worth up to $1 billion, deploying AI across drug research, regulatory affairs, and manufacturing (more on that below). Novo Nordisk is partnering with OpenAI on molecular discovery with a rumored licensing model in which OpenAI would receive a share of drug sales from AI-assisted development. Amazon's Bio Discovery platform, working with Memorial Sloan Kettering, designed nearly 300,000 novel antibody molecules in weeks, a process that traditionally takes up to a year.

These are just a few of the announcements from the past couple of weeks. There's been no shortage of partnerships forged between AI companies and biopharmas over the past year.

The regulatory side

On the back end, HHS Secretary Robert F. Kennedy Jr. testified before Congress that more than 90 percent of FDA reviewers are now using AI tools, claiming the agency has broken records for drug approvals as a result.

If these speed gains are real, then the pipeline that once took decades is being compressed. That is genuinely good news for patients, particularly those, as one MSK oncologist put it, who "come here with a clock."

The question no one is answering

But here's the question I haven't heard anyone in industry answer directly: if AI is materially reducing the cost and time of bringing drugs to market, where does that efficiency go? Especially at a time when the American public is concerned with affordability, including healthcare affordability. Recently, a Gallup poll suggested “healthcare affordability” was the top issue of concern for Americans.

The traditional justification for high drug prices is that pharmaceutical companies must recoup the enormous cost of failed candidates and lengthy development timelines. It's an argument that has typically fortified it against more draconian policies to force prices down, even if it's been used to justify pricing that puts medicines out of reach for the patients who need them. That argument rests on an assumption about the cost structure of drug development that AI may now be actively undermining.

Where do the savings go?

If you can design 300,000 antibody candidates in weeks instead of years, and if 90 percent of regulatory reviewers are moving faster with AI assistance, the economics of the pipeline are changing. The question is what happens to those savings.

There are a few plausible answers. Efficiency gains could be reinvested into R&D. That would be a genuine win. The more cynical view is that they could also flow to shareholders as expanded margins, with prices holding steady or continuing to rise. Or some combination of both, distributed in ways that are largely invisible to payers and patients.

The pharmaceutical industry has long argued that high prices reflect the full cost of the development pipeline, including the many candidates that fail. That argument may be weakening as AI compresses both timelines and failure rates. If or how it impacts the pricing pain point for consumers remains an open question.

🛜 Field Signals: A quick hit on this week’s industry announcements, policy developments, and ethical considerations.

Meta Employees Are Up in Arms Over a Mandatory Program to Train AI on Their Mouse Movements and Keystrokes Meta has rolled out software on US employees' computers that captures keystrokes, mouse movements, and screen content to train AI agents on how humans actually navigate workplace tasks — with no opt-out option, according to internal communications obtained by Business Insider. The top-rated employee response to the announcement was "This makes me super uncomfortable. How do we opt out?" — a reaction that tracks directly with this week's HBR research on how coercive AI adoption erodes trust and engagement.

Meta to Cut 10 Percent of Work Force in AI Push Meta is laying off roughly 8,000 employees and closing 6,000 open roles as the company redirects resources toward AI development, with CEO Mark Zuckerberg framing the cuts as part of an efficiency drive to offset the cost of AI investments that are expected to reach $115 to $135 billion this year. The announcement lands the same week Meta rolled out mandatory keystroke and mouse-movement tracking software on employee computers — together, the two moves offer a clear portrait of what AI-driven workforce transformation looks like in practice at one of the world's largest technology companies.

You're About to Feel the AI Money Squeeze The Verge reports that the era of cheap or free AI access is ending as labs face mounting pressure to generate returns on trillions in infrastructure investment — with Gartner estimating that major providers would need to earn close to $7 trillion cumulatively through 2029 just to avoid write-downs, a figure that would require token consumption to grow by 50,000 to 100,000 times current levels. The result is a wave of subscription tier changes, usage restrictions, and advertising that is already pushing costs downstream to enterprise customers and developers, who are increasingly hedging by mixing frontier models with cheaper open-source alternatives.

Partnering with Industry Leaders to Accelerate AI Transformation Google DeepMind has announced enterprise partnerships with Accenture, Bain & Company, BCG, Deloitte, and McKinsey to accelerate large-scale AI adoption — offering early access to frontier Gemini models, joint work on industry-specific use cases, and direct access to DeepMind leadership for customer CEOs and boards. The announcement mirrors OpenAI's Frontier Alliance playbook from earlier this year, suggesting the major labs have converged on the same go-to-market strategy: route enterprise adoption through the world's largest consulting firms.

Covered California Partners with Google Public Sector to Use AI to Accelerate Healthcare Access for Millions Covered California has deployed Google Cloud's Document AI across its CalHEERS eligibility platform, reducing manual document processing tasks by 40% and cutting verification timelines from up to three weeks to real-time on-screen feedback — while routing complex or ambiguous cases to human specialists for review. The deployment covers 25 document types and includes fraud detection capabilities trained to flag subtle markers like missing signatures and mismatched dates.

Merck to Partner with Google Cloud on AI Initiatives Merck has announced a partnership with Google Cloud valued at up to $1 billion over several years, deploying Gemini Enterprise across drug research, regulatory affairs, manufacturing, and commercial operations — with AI already being used to cut dossier compilation time and cost by half in markets where reimbursement approval is required. The deal, announced at Google Cloud Next, is framed as at least a decade-long collaboration and signals that major pharma is moving past AI pilots into enterprise-scale production deployment.

Random Discord Group Got Anthropic's Mythos Before CISA For weeks before the breach was reported, an unauthorized Discord group accessed Mythos Preview — Anthropic's most capable and restricted model, withheld from public release due to its cyberoffense capabilities — apparently by guessing the URL based on Anthropic's naming conventions, Bloomberg reported. The incident complicates Anthropic's framing of its limited Project Glasswing rollout as a safety measure, particularly after Axios separately reported that CISA, the federal agency responsible for protecting critical infrastructure, was not included among the roughly 40 institutions granted access.

OpenAI Introduces Workspace Agents in ChatGPT OpenAI has launched workspace agents in ChatGPT, a Codex-powered system that lets organizations build persistent AI workers for shared business tasks — running on schedules, integrating with Slack, Google Drive, Microsoft SharePoint, and other tools, and handling recurring workflows like software triage, reporting, and vendor reviews without requiring repeated prompting. Admins can approve sensitive actions and monitor activity through compliance controls, positioning the feature as an enterprise automation layer rather than a conversational tool.

The Therapist in Your Pocket: Chatty, Leaky — and AI-Powered A wave of AI therapy apps — at least 45 identified in Apple's App Store — are marketing themselves as mental health tools despite little clinical evidence of efficacy, no standardized regulatory framework, and privacy policies that in several cases contradict their own App Store disclosures. The structural problem is straightforward: "therapy" is not a legally protected term, the FDA has not issued standards for testing these products, and at least a dozen wrongful death lawsuits have been filed against AI chatbot makers by users or families of users who died by suicide.

Making ChatGPT Better for Clinicians OpenAI has launched ChatGPT for Clinicians, a free verified-access version of ChatGPT for U.S. physicians, NPs, PAs, and pharmacists, built around clinical documentation, medical research, and care consult workflows — with HIPAA support available via BAA and conversations excluded from model training. Alongside the launch, OpenAI released HealthBench Professional, an open benchmark for real clinician chat tasks developed with physician collaborators, on which GPT-5.4 in the clinician workspace outperformed other models and human physicians — though the benchmark was designed and evaluated by OpenAI's own physician advisory network.

AI Prescribing Medications in Utah: A Flawed Regulatory Playbook Utah's regulatory sandbox has authorized Doctronic, an AI system, to autonomously prescribe medication refills — a threshold no jurisdiction has previously crossed — while the FDA has remained publicly silent despite the fact that the use case appears to meet its own definition of a medical device. The author argues this state-by-state improvisation recapitulates the pre-Flexner chaos of nineteenth-century medicine, and that coherent national standards for AI clinical practice need to be established before deployment, not discovered from harm after the fact.

Federal Test of AI Prior Authorization Is Delaying Care for Seniors, Report Says A Senate report from Sen. Maria Cantwell finds that Washington state hospitals are seeing procedures under CMS's new WISeR prior authorization model take two to four times longer to approve than before the program launched — with routine authorizations running 15 to 20 days against a three-day standard, and nearly 100 patients at the UW Medical System currently waiting for epidural steroid injections. CMS designed WISeR to use AI to reduce fraud in traditional Medicare, but the early evidence from the field suggests the program is functioning primarily as a delay mechanism rather than a targeted fraud filter.

Preparing for the Convergence of Three New Colorado Laws Targeting AI in Healthcare Colorado healthcare organizations are simultaneously navigating three distinct AI regulatory tracks: the Colorado AI Act taking effect June 30 unless replaced, HB 26-1139 prohibiting coverage decisions based solely on AI-generated group data, and HB 26-1195 restricting AI use in mental health therapy and requiring written patient consent before AI records or transcribes sessions. A legal advisory from Snell & Wilmer notes that no other industry in the state faces this level of convergent AI regulation, and that exemptions under the AI Act do not carry over to the two healthcare-specific bills.

AMA Urges Lawmakers to Implement Safeguards on AI Chatbots The American Medical Association sent letters to the chairs of three congressional AI and digital health caucuses calling for federal guardrails on mental health chatbots — including a prohibition on AI diagnosing or treating mental health conditions, FDA review requirements for tools that cross into therapeutic guidance, mandatory disclosure that users are talking to AI, and a ban on targeted advertising to children. The AMA framed the absence of federal action as urgent, noting that current regulatory frameworks were not designed for generative AI tools capable of shifting from casual conversation to therapeutic interaction within a single session.

Healthcare AI Is Only as Trustworthy as the Humans Behind It The author argues that AI models have no independent ethics — accountability sits with the humans who design, train, and test them — and that closing the trust gap in healthcare requires organizations to explain in plain language how their models make decisions and where human judgment remains in the loop. The piece makes a useful distinction between automation that removes humans and intelligent automation that triages and escalates so that human clinicians are deployed precisely where they're needed most.

Multimodal Artificial Intelligence and Online Learning in Youth Mental Health: A Scoping Review A McMaster University-led scoping review of 24 studies finds that AI research applied specifically to youth mental health remains limited — most work targets symptom-level detection like stress and emotion rather than clinically diagnosable conditions, and few studies incorporate the multimodal data inputs or adaptive online learning methods needed for real-world deployment. The authors identify the ethical and logistical challenges of collecting data from minors as a key barrier, and flag the rapid rise of large language models as an emerging but largely unevaluated frontier for scalable youth mental health support.

Thinking of Using a Chatbot for Medical Advice? Read This First. Two studies published this month — one in BMJ Open, one in JAMA Network Open — found that AI chatbots got nearly half of health questions wrong and failed 80 percent of the time when asked to reason through ambiguous clinical scenarios, with researchers noting that models tend to collapse prematurely into single answers rather than preserving diagnostic uncertainty the way clinicians do. A separate West Health-Gallup survey found that roughly 14 million Americans report forgoing a provider visit because of health information or advice they received from AI.

Is AI Actually Improving Healthcare? Two researchers argue in Nature Medicine that the central problem in healthcare AI is not a shortage of models but a shortage of rigorous evaluation — accuracy metrics don't guarantee clinical benefit, and outcome improvements often reflect workflow changes triggered by AI deployment rather than the algorithm itself. They call for randomized or staggered rollouts and structured decision elicitation to isolate AI's actual causal contribution, making the case that the evidentiary standard for AI should match that of any other clinical intervention.

Show Us the Evidence for the Value of Medical AI In an editorial, Nature Medicine argues that the medical AI field lacks consistent evidentiary standards — claims of clinical impact routinely outpace the evidence supporting them, and statistical performance metrics alone don't establish real-world benefit. The journal calls for a proportional evidence framework in which the strength of evidence required scales with the strength of the claim, and signals it will enforce those standards in its own published research.

The Missing Knowledge Layer in AI: A Framework for Stable Human-AI Reasoning A preprint proposes that large language models share a structural flaw with human reasoning: fluency signals trust even when underlying reasoning has drifted, making it possible for model and user to reinforce each other's errors without either detecting the problem. The authors call this a "missing operational substrate" for AI governance and propose a two-layer framework — human-side mechanisms to surface uncertainty and conflict, and a model-side epistemic control loop — designed to make reasoning instability visible before it propagates into high-stakes decisions.

Tracing the Pen: Electronic Health Records Amid the Rise of Generative AI A perspective from Mass General Brigham and Duke researchers argues that as LLMs increasingly assist with clinical documentation, the blending of AI- and human-generated content within EHRs creates a traceability problem that threatens clinical accountability and documentation integrity. The authors call for proactive technological and policy solutions — including watermarking and disclosure standards — to ensure clinicians and institutions can reliably distinguish what a human wrote from what a machine generated.

More Than 60% of Americans Using AI Tools for Medical Information: Survey A Harris Poll survey of 2,057 U.S. adults, commissioned by Merck Manuals, finds that 62% of Americans have used AI tools like ChatGPT or Gemini to seek medical information — with usage highest among Gen Z and Millennials at 77% — while 32% say they don't trust medical information from AI and 90% of users report taking steps to verify what they receive. The survey adds consumer behavioral data to a growing body of evidence this week on how Americans are actually using AI for health, independent of whether the tools are accurate or regulated.

When Creating an AI Strategy, Don't Overlook Employee Perception A survey of 1,294 desk workers finds that employees who perceive their organization is using AI to automate — rather than augment — report lower well-being, higher intent to leave, and greater production of low-quality AI-generated work, even when adoption appears high on paper. The authors argue that automation strategies may show early gains but carry predictable long-term costs, while augmentation requires deeper upfront investment and a credible organizational commitment to people.

Introducing Claude Design by Anthropic Labs Anthropic has launched Claude Design, a collaborative visual design tool that lets users build prototypes, decks, mockups, and marketing assets through conversation — with the ability to import a team's existing design system so outputs stay on-brand automatically. The tool is available in research preview for Pro, Max, Team, and Enterprise subscribers, and includes a handoff feature that packages completed designs for direct implementation via Claude Code.

Claude Cowork Now Supports Live Artifacts Anthropic has added live artifact functionality to Cowork, allowing users to build dashboards and trackers that connect directly to apps and files and refresh with current data — accessible across sessions with full version history. The update moves Cowork closer to a persistent workspace rather than a session-based tool.

Introducing ChatGPT Images 2.0 OpenAI has launched ChatGPT Images 2.0, an upgraded image generation model offering greater precision, multilingual text rendering, and stronger stylistic range — from photorealism to manga to editorial design. The release signals a continued push by major AI labs to close the gap between text and visual generation in a single integrated tool.

Google Meet's AI Note-Taking Feature Now Works for In-Person Meetings Google has expanded its Gemini-powered "Take Notes" feature in Google Meet to support in-person and impromptu meetings — not just video calls — with the tool summarizing key points, generating a transcript, and saving a document to Google Drive automatically after the meeting ends. The feature is currently Android-only and limited to select Google Workspace enterprise tiers, with broader platform support coming soon.

Microsoft Launches 'Vibe Working' in Word, Excel, and PowerPoint Microsoft has rolled out Agent Mode as the default Copilot experience for Microsoft 365 Copilot and Premium subscribers, enabling the AI to take direct action inside Word, Excel, and PowerPoint — making multi-step edits, adding formulas, updating decks — rather than just answering questions in a sidebar. The company acknowledges that earlier versions of Copilot fell short because foundation models weren't capable enough to reliably execute in-document commands; Agent Mode represents their first attempt to close that gap.

👉 Apr. 27–28, 2026 — AI for Hospitals & Health Plans Summit, New Orleans, LA

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

👉 May 7–8, 2026 — NBER Conference on AI in Healthcare, Cambridge, MA

👉 Jun. 8–10, 2026 — Fortune Brainstorm Tech, Aspen, CO

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

👉 Aug. 4–6, 2026 — 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