The Bandwidth - Weekly AI Roundup

The biggest story this week was the brief frenzy over Moltbook, a Reddit-style forum designed for AI agents only. Humans could observe but not participate. As the platform took off, feeds filled with screenshots of agents posting strange messages, interacting with one another, and appearing to “organize” themselves. Headlines followed.

At first glance, it looked like science fiction. Even seasoned AI leaders reacted publicly, suggesting Moltbook might represent a glimpse of where autonomous systems are headed.

But after a few days and closer scrutiny, the story deflated. Much of the activity was orchestrated by humans, prompting or managing large numbers of agents. Some posts likely came from people posing as bots. Beyond novelty and weirdness, there was little evidence of genuine risk.

The real issue was never Moltbook, IMO. It was Moltbot.

Moltbot, formerly known as Clawdbot and OpenClaw, is an open-source AI agent designed to run continuously on a user’s own machine. Unlike conventional chat tools, it can act. It browses the web, controls applications, manipulates files, sends emails, and operates largely in the background. It even communicates with you over your phone. Security researchers and industry leaders have been explicit: installing tools like this without strict controls is risky.

That distinction matters for healthcare.

I am far less worried about large frontier labs deploying unsafe systems. They are highly visible, heavily scrutinized, and structurally incentivized to implement security controls, audit trails, and regulatory compliance. When they fail, they fail loudly. Case in point, last week I wrote about the scrutiny OpenAI faced for the underwhelming performance of its consumer-facing healthcare app.

The more credible risk runs elsewhere.

Healthcare’s exposure could come from smaller vendors, pilot tools, and solo practitioners experimenting at the edge. Patients may be encouraged, intentionally or not, to share sensitive information with systems that lack clear governance, containment, or accountability. When something goes wrong, there may be no obvious human owner to hold responsible. And it is not clear that regulators are prepared for this class of risk.

While regulation is needed, it alone will not solve this. Consumer education and basic AI literacy must be part of the response, so patients, providers, and caregivers understand what these tools can do and where their risks lie. The Moltbook episode makes clear that without this literacy, we risk not only real harm, but overreaction and misplaced fear when novelty outpaces understanding.

Sam Altman sketches the next phase of AI
In a lengthy and wide-ranging interview, Sam Altman argues the AI story is shifting from model breakthroughs to systems, distribution, and real-world integration. He emphasizes agents, long-term memory, and tighter coupling between AI and everyday workflows, while downplaying near-term fears of runaway intelligence.

Microsoft vs. Anthropic
Internal Microsoft warnings suggest Anthropic’s Cowork agent could outpace Microsoft 365 Copilot, triggering rapid internal prototyping, sometimes using Anthropic’s own models.

Amazon makes Alexa Plus free for Prime members
Amazon is rolling out Alexa Plus at no additional cost to Prime subscribers in the U.S., signaling a push to normalize more capable, agent-like assistants in everyday consumer environments.

Humana deploys AI in call centers
Humana rolled out Google Cloud AI tools to assist call center staff in real time, summarizing conversations and surfacing next-best actions. Augmentation comes before automation.

Fitbit founders launch Luffu
James Park and Eric Friedman unveiled Luffu, a family-first AI health app aggregating data across children, aging parents, and even pets. A bet on caregiving as the next AI health frontier.

Epic moves from ambient scribe to active clinical agent
Epic detailed its long-anticipated AI Charting feature, signaling a shift from passive ambient documentation toward more active clinical assistance, including drafting orders and diagnoses. With several health systems already piloting the tools, Epic’s move could disrupt the fast-growing ambient scribe market and accelerate the normalization of agentic AI inside core EHR workflows.

Health systems press for AI transparency despite federal rollback signals
Inside Health Policy reports that even if the Trump administration finalizes proposals removing AI transparency from federal health IT certification requirements, health systems are unlikely to stop demanding visibility into how AI tools work.

ICYMI: MACPAC on AI and Medicaid prior authorization
MACPAC reviewed early use of automation in Medicaid PA. Benefits include reduced cost and burden. Risks include access barriers, bias, and transparency gaps. Federal oversight remains limited.

International AI Safety Report
More than 100 AI experts, led by Yoshua Bengio, warn that risks like deepfake fraud, cybercrime, manipulation, and misuse for biological harm are now real-world problems. The U.S. declined to participate.

AI-assisted radiology shows real gains
A large Swedish randomized trial (~100,000 patients) found AI increased cancer detection by 29% and reduced radiologist workload by 44%. Human-plus-AI beats either alone.

Trust in health AI remains fragile
New CHAI survey data shows 75% of respondents use AI, but only 13% feel very comfortable. Over 80% say trust would increase with clear accountability.

AI breast cancer screening outcomes
AI-assisted mammography improved detection rates and reduced aggressive and large tumors without increasing false positives, while cutting clinician workload nearly in half.

AI in real-world virtual care
In partnership with Included Health, Google announced it will be launching a first-of-its-kind nationwide study to evaluate conversational AI within real-world virtual care workflows.

Build product management skills to unlock AI adoption
A new HBR piece argues that the biggest blocker to AI adoption is not prompting skill, but weak product discipline. Employees get real value from AI when they define high-value problems, evaluate tools against workflows, run small experiments, and integrate what works. In other words, they need to think like product managers.

Mozilla adds user-facing AI controls
Firefox introduced native controls that let users see and manage AI features directly, including removing them altogether. Governance is moving into product user experience.

Incentivize AI use, not just deployment
In this piece published by Wharton School of business, senioir fellow Scott Snyder, discusses that despite widespread access, fewer than 5% of workers use AI to transform their work. He argues the missing ingredient is incentives and baking AI usage into performance reviews.

Meta hard-wires AI into performance
Meta now ties performance reviews and bonus multipliers directly to AI-driven impact through its Checkpoint system. AI use is table stakes.

Tool to try: Mixboard
Google Labs’ Mixboard turns messy idea boards into polished presentations with one click. Built for non-designers who need to move fast.

AI boosts productivity but threatens judgment development
A new HBR piece flags a growing organizational risk: AI dramatically amplifies the productivity of experienced workers, but leaves junior employees struggling to assess quality or improve AI-generated work. As AI takes over the messy, repetitive tasks that once built judgment, organizations risk creating leaders who have never done the underlying work. The fix is not just “keeping humans in the loop,” but deliberately redesigning work to build judgment through ownership, exposure to consequences, stretch experiences, and tools like simulations and case-based learning.

👉 Feb. 11–13 — The Complexities of AI in Health Care, AHLA, Las Vegas & Virtual

👉 Feb. 17 — Webinar: Modernizing data systems and harnessing AI in the public health system.

👉 Feb. 23 — Reminder: HHS seeks input on AI in clinical care by this date

👉 Feb. 26 — Harvard Business Review Strategy Summit: Build your transformation playbook for the AI era.

👉 Mar. 12–18, 2026 — SXSW 2026, Austin, TX

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

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

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

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