The Bandwidth: Decoding Health + AI

We have spent several weeks in this newsletter examining how to evaluate AI in healthcare. Does it work? Is it safe? And who gets to make that determination? Those questions remain urgent and unresolved, especially at the federal level.

But there is a parallel problem that deserves attention too: user experience.

What the research is starting to show.

Three pieces I read through this week helped elevate this issue in my mind.

A systematic review in npj Artificial Intelligence evaluating AI agents in clinical settings includes "humanistic care" as a formal evaluation dimension. In this case, evaluators aren’t just looking at diagnostic accuracy, but also measuring the degree to which AI systems give attention and respect for people's living conditions, dignity, rights, and values. The study authors note to foster confidence among AI users, we need “to extend the evaluation framework of intelligence in healthcare to encompass multidimensional factors, including economic indicators (e.g., cost-benefit analysis, return on investment, and long-term maintenance costs), safety indicators (e.g., incidence of adverse events), and patient satisfaction or other subjective metrics, in addition to technical performance.”

A separate study in Nature Mental Health examined adoption barriers for AI-enabled psychotherapy tools among both patients and therapists. Dropout rates for AI-based mental health applications range from 5 to 34%. Mental health professionals show low intent to use these tools, particularly those that are patient-facing. The study finds that concerns center not on functionality, but on the loss of human connection, disrupted therapeutic relationships, and insufficient customization to the needs of both patients and clinicians. When a therapist won't use a tool because it doesn't fit their practice, or a patient drops out because it feels impersonal, the clinical outcome degrades.

And a STAT News piece published this week captures a consequential gap in the rapid development and deployment of this technology: patients are seldom consulted in the development and testing of AI tools that are meant to serve them. Leo Celi, a researcher at MIT and physician at Beth Israel Deaconess, put it plainly: "We make them guests of honor. We make them sit there and shut up." His response is to organize events where patients actually design the evaluation frameworks — not just validate ones built without them.

Engagement isn't a feature. It's a foundation.

None of this is entirely new. Healthcare has long debated how to engage communities in the design of treatments and services. What's new is the speed at which AI is being deployed into clinical settings before those engagement questions have been seriously worked through. The result, as Celi notes, is validation that's more performative than substantive.

I'm not arguing that co-design should be a regulatory requirement. But as health systems, developers, and policymakers think about where and how to deploy AI, the human factors question — who was consulted, whose needs shaped the design, how the tool fits into an actual care relationship — deserves to sit alongside the technical evaluation questions, not after them.

If adoption fails because the tool doesn't fit the clinical relationship, or because patients didn't trust it, or because it wasn't designed for the population it's meant to serve, even a technically sound AI won't deliver on its promise. Safety and effectiveness are necessary, but they are not sufficient. You can build something that works in an AI lab and fails in the real world.

100 Gen AI Apps, Wave 6 a16z's latest consumer AI rankings confirm ChatGPT's commanding lead: on web it draws 2.7x more monthly traffic than second-place Gemini, and on mobile it is 2.5x larger by monthly active users. Weekly active users have grown by 500 million over the past year to 900 million — meaning more than 10% of the global population now uses ChatGPT every week. The data underscores how difficult it will be for competitors to close the consumer gap at this scale. Yes, but, according to the folks at the Nueron, Anthropic now wins roughly 70% of head-to-head matchups against OpenAI among businesses buying AI for the first time.

Meta Acqui-Hires Moltbook's Creators Meta has acquired the team behind Moltbook — a viral AI agent social forum built largely by one of its creators using an OpenClaw bot — folding them into its Superintelligence Labs team. The platform has 2.8 million registered bots, with nearly 200K verified to real people, and functions as an always-on directory for agent coordination. The move follows OpenAI's hire of OpenClaw creator Peter Steinberger, whom Zuckerberg had reportedly courted first.

AI Agent Hacked McKinsey's Internal Chatbot in Two Hours A security research firm used an autonomous AI agent to breach McKinsey's internal AI platform, Lilli, gaining full read-write access to its production database — including 46.5 million chat messages, 728,000 confidential client files, and the system prompts controlling the chatbot's behavior. The prompts were writable, meaning an attacker could have silently poisoned responses across all 40,000 users. McKinsey patched the vulnerabilities within hours of disclosure.

Blue Cross Blue Shield Says Data Back Up Claim That AI Is Driving Up Medical Bills A new report from a data analysis firm tied to the Blue Cross Blue Shield Association offers the first publicly available evidence that AI-assisted medical coding may be inflating hospital bills — estimating $663 million in additional inpatient spending. Focused on postpartum hemorrhage codes, the analysis found that hospitals with rapid growth in coding did not show corresponding increases in treatment, suggesting diagnoses — not care — are rising. Outside experts urge caution on the methodology, and the dynamic cuts both ways: Medicare Advantage insurers face parallel accusations of AI-assisted upcoding on their own side of the ledger.

These Diseases Were Thought to Be Incurable. Now AI Is Unlocking New Treatments. AI is accelerating drug discovery at a scale previously unimaginable — screening billions of molecular compounds in days rather than months, at a fraction of the cost. Researchers at MIT have identified two promising antibiotic candidates against drug-resistant gonorrhea and MRSA; Cambridge scientists have flagged five new compounds targeting the protein clumps associated with Parkinson's; and a Harvard model found nearly 8,000 approved drugs that could be repurposed across 17,000 diseases. The bottleneck remains access: much of the underlying drug-property data is held by private companies and unavailable to researchers.

Epic Expands AI Road Map, Previews Factory to Build and Orchestrate AI Agents At HIMSS26, Epic announced Agent Factory, a platform that lets health systems build, customize, and deploy their own AI agents within the Epic environment — alongside a growing suite of clinical, patient-facing, and revenue cycle AI tools now used by more than 85% of its customers. Outcomes are beginning to materialize: early lung cancer detection at The Christ Hospital reached 69% versus a 46% national average, coding-related denials dropped more than 20% at health systems using Epic's revenue cycle AI, and one health system cut prior authorization submission time by 42%. The announcements signal a shift in the Epic AI conversation from theoretical potential to measurable return on investment.

Amazon Launches Health AI Agent on Amazon Website and App Amazon has expanded its Health AI agent — previously available only to One Medical members — to Amazon.com and the Amazon app, with a goal of reaching all U.S. customers. The agentic system can answer health questions, explain lab results and medical records, manage prescription renewals, and book appointments with One Medical providers, all within a HIPAA-compliant environment. Eligible Prime members receive up to five free direct-message care visits for 30+ common conditions as an introductory offer — a move that positions Amazon as a serious contender in the consumer health navigation space.

Microsoft Previews Copilot Health AI Tool Microsoft announced Copilot Health, an AI service that lets users combine electronic health records, lab results, and wearable data from more than 50 different device types — covering records from more than 50,000 U.S. health providers — to generate personalized health insights. The tool joins ChatGPT Health and Amazon's Health AI in a rapidly crowding consumer health AI space, with Microsoft AI head Mustafa Suleyman describing it as a first step toward "medical superintelligence" and an eventual affordable alternative to concierge medicine. For now, the service is free and U.S.-only, available through a waitlist, with health data kept separate from general Copilot and not used for model training.

Anthropic Sues the Pentagon and White House Over Federal Blacklist Anthropic filed two lawsuits challenging the Pentagon's "supply chain risk" designation and a White House directive ordering federal agencies to drop Claude — arguing both amount to retaliation for the company's public advocacy on AI safety limits. The suits contend the supply chain label was designed to counter foreign threats, not punish domestic companies over policy disagreements. More than 30 OpenAI and Google staffers signed a legal brief in support, warning the blacklisting threatens U.S. AI leadership.

America Cannot Withstand the Economic Shock That's Coming Former Commerce Secretary Gina Raimondo argues that AI-driven job displacement is an approaching crisis — and that the answer is a new public-private bargain: employers define emerging skill needs in real time, government invests in modular training, wage insurance, and modern apprenticeships. Without it, she warns, mass unemployment becomes mass rage directed at the companies and policymakers enabling AI.

How Collaboration Can Enable Action on AI and Mental Health With millions already turning to AI chatbots for mental health support — often without clinical validation or transparent safety frameworks — researchers from Partnership on AI argue that no single actor can govern this responsibly alone. Three structural problems compound the risk: AI development outpaces mental health research, labs are largely solving identical problems in isolation, and independent evaluation mechanisms don't yet exist. The authors call for multistakeholder collaboration among AI labs, clinicians, policymakers, and people with lived experience — starting with suicide prevention, where normative frameworks are most established.

When the Front Door Is an Algorithm: Guiding Consumer Use of AI in Health Care For millions of consumers, the first stop for health information is no longer a clinician or nurse line — it's an AI chatbot. Health policy attorneys Mark Lutes and Richard Hughes argue that as tools like ChatGPT Health integrate with medical records and wearables, the governance stakes shift considerably: AI that lacks full clinical context can reassure when it should escalate, and confident-sounding outputs can drive harmful decisions. They call for evaluation frameworks that measure escalation behavior and equity impacts, not just accuracy — and for incentive structures that reward appropriate care rather than deflected utilization.

Anthropic Launches the Anthropic Institute Anthropic has created a ~30-person research group under co-founder Jack Clark that merges its Frontier Red Team, Societal Impacts, and economics research teams — with plans to double staff annually. The Institute will study AI's societal impacts, share learnings from frontier model development publicly, and engage directly with workers and industries facing displacement. Founding hires include ex-DeepMind researcher Matt Botvinick, economist Anton Korinek, and Zoe Hitzig, who left OpenAI over the company's decision to introduce ads in ChatGPT.

AI Agents in Healthcare: Applications, Evaluations, and Future Directions A systematic review in npj Artificial Intelligence maps LLM-based AI agents across clinical decision support, diagnostics, report generation, patient-facing chatbots, hospital management, and medical education. The authors propose a two-tiered evaluation framework that adds humanistic dimensions — patient trust, equity, clinician experience — alongside standard performance metrics. Seven future priorities are identified, including safety governance, ethical accountability, and the evolving roles of clinical staff.

AI Could Help Spot Heart Disease in Routine Breast Mammogram Screenings An Emory University study of more than 123,000 women found that AI-detected arterial calcification in standard mammograms identified cardiovascular risk — including in women under 50 — at two to three times the rate of those without severe calcification. Fewer than 40% of women know their cholesterol levels, but nearly 70% are up to date on mammography. Researchers argue the tool could turn a widely adopted cancer screening into a low-cost cardiovascular risk signal, at no additional inconvenience to patients.

Why Millions Are Turning to ChatGPT for Mental Health A Bournemouth University study of nearly 31,000 adults across 35 countries found that 61% globally — and 41% in the UK — are comfortable using ChatGPT as a mental health counselor. Researchers attribute the trend to long wait times, cost barriers, and AI's non-judgmental tone. The limitations are real: developers deliberately keep language vague to avoid clinical diagnoses, and AI may lack the human judgment to recognize when a situation requires escalation to emergency services.

Healthcare Uses Specialized Language. It Needs Specialized AI, Too. Generalist LLMs can read every word in a clinical chart and still miss what the words mean together — misreading a Hoffman's sign, overlooking surgical history that waives prior authorization requirements, or failing to recognize that three prednisone courses signal severe asthma. The authors, drawing on their experience building AI for prior authorization, argue that healthcare organizations should treat LLMs like supervised trainees: map specialty-specific clinical dialects before deployment, require transparent reasoning chains, and build structured feedback loops that let clinicians correct the model when it fails.

AMA Physician AI Sentiment Report Physician AI adoption has more than doubled since 2023 — from 38% to 81% — and the average number of AI use cases per physician has grown from 1.1 to 2.3, according to new AMA research. More than three-quarters of physicians say AI improves their ability to care for patients, with the greatest perceived potential in diagnostic accuracy and administrative burden reduction. Physicians remain clear-eyed about limits: 85% want a voice in adoption decisions, and there is notable concern about patients using AI to interpret complex clinical results — such as radiology or pathology findings — without physician guidance.

Claude Now Creates Interactive Charts, Diagrams, and Visualizations Anthropic has launched an in-beta feature that lets Claude build interactive charts, diagrams, and visualizations inline during conversations — without any code. Unlike artifacts, which are permanent outputs, these visuals are conversational aids that update or disappear as the discussion evolves, and can be adjusted on request. The feature is on by default across all plan types.

"AI Brain Fry" Is a Real Cognitive Risk — and a Leadership Problem A BCG study of 1,488 workers finds that intensive AI oversight produces measurable mental fatigue — distinct from burnout — marked by mental fog, slower decision-making, and elevated error rates. Those experiencing it made major errors 39% more frequently and were significantly more likely to quit. Teams that embed AI collectively into workflows fare better than individuals juggling tools independently.

Microsoft Launches Copilot Cowork Microsoft has upgraded Copilot from a chatbot into an active task executor inside Microsoft 365 — describe a goal, and it breaks the request into steps and completes them across Outlook, Teams, and Excel, running in the background within existing security and compliance guardrails. The launch appears to be a direct competitive response to Anthropic's Cowork, which debuted earlier this year.

Paperclip Most multi-agent AI tools give you power without structure — tasks get delegated across systems with no clear accountability for what happened or why. Paperclip solves this by organizing AI agents into a hierarchy that mirrors how a company works: one agent sets the direction, others execute, and every action is logged with a cost and a record.

Perplexity Pitches a More Secure OpenClaw Perplexity announced "Personal Computer," software that runs a locally controlled AI agent on a dedicated device — such as a Mac mini — with full access to local files and apps, user-confirmed actions, and a built-in audit trail. The company positions it as a more secure alternative to OpenClaw, which shifts security responsibility entirely to the user. Perplexity now offers two agent products: the local "Personal Computer" and the cloud-based "Perplexity Computer," the latter now expanded with enterprise security controls and mobile support.

👉 Mar. 17, 2026 — The Check Up with Google 2026, Virtual Livestream

👉 Mar. 27 — “The AI Doc: Or How I Became An Apocaloptimist” opens in theaters. Watch the trailer

👉 Mar. 30–31, 2026 — IAPP Global Privacy Summit, Washington, DC

👉 Apr. 6–9, 2026 — HumanX 2026, San Francisco, CA

👉 Apr. 7–8, 2026 — Behavioral Health AI Summit, Nashville, TN

👉 Apr. 10 — Ethical AI: Leadership and Governance, Virtual

👉 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

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