The Bandwidth - Weekly AI Roundup

There was a lot of noise this week that came from Anthropic CEO Dario Amodei, who published a 38-page essay (38 pages!) warning that rapidly advancing AI could pose profound risks to society. This is obviously an important essay, and I address it further below. But the question I want to focus on at the top is more practical and more immediate: how disruptive are frontier AI labs actually going to be to healthcare?

This is a question I’ve returned to repeatedly in this newsletter. Over the past several weeks, companies like OpenAI and Anthropic have moved aggressively into healthcare, launching consumer-facing tools and announcing new offerings for health systems and providers, some explicitly described as HIPAA-compliant. When those announcements landed, the obvious question was not whether the technology was impressive, but what uptake would actually look like. Would patients trust these tools? Would providers integrate them into already fragile workflows?

On the consumer side, early reporting suggests disappointment. The Washington Post’s testing of OpenAI’s consumer-facing health features surfaced inconsistencies and legitimate concerns about reliability. A brain cancer patient writing in STAT shared similar experiences, but also held out hope about the utility of these new tools. Indeed, the tools are still in their infancy and may still improve. But for now, the gap between promise and practical utility is visible, especially in a domain where trust, context, and accountability matter a great deal.

The usefulness on the provider side also remains an open question, but the financial incentives are clear for AI developers. Frontier labs are under pressure to demonstrate viable revenue models commensurate with the capital they have raised or spent. Healthcare, with $5.3 trillion in U.S. spending and growth that continues to outpace the broader economy, is an obvious target. If you are looking for markets large enough to absorb frontier-scale AI ambitions, healthcare is near the top of the list.

But the race into healthcare actually may prove to be more of a slog and sobering. First, off-the-shelf AI tools are rarely plug-and-play in healthcare. As Modern Healthcare reports in its coverage of early agentic AI deployments, even promising use cases require extensive customization, governance review, and co-development with internal experts. The value is real, but it is earned slowly, not captured overnight.

Second, large health systems and healthcare companies are increasingly building their own AI tools. These systems are designed to fit existing workflows, regulatory constraints, and organizational culture, rather than forcing a wholesale switch to an external platform. For organizations at that scale, an internal build often feels less risky than wholesale dependence on a frontier lab whose priorities may shift or that doesn’t have the healthcare know-how of large systems or incumbent suppliers.

That raises a harder question for smaller organizations. Many will not have the resources to build bespoke tools. But they may be more willing to adopt AI solutions created by peer healthcare institutions they already trust than by technology companies entering the sector from the outside. In that sense, the most meaningful disruption may come not from frontier labs replacing incumbents, but from incumbents quietly reshaping themselves and, in some cases, pulling others along with them.

Apple to launch Gemini-powered Siri in phased rollout
Apple plans a two-stage Siri reboot, starting with a personalized assistant and followed by a fully conversational chatbot later this year. The move signals Apple’s push toward deeply embedded, context-aware AI across its ecosystem.

ChatGPT’s market lead narrows as Gemini gains ground
OpenAI’s ChatGPT still leads AI traffic, but its share has dropped sharply as Gemini grows through integration across Google Search, Gmail, Workspace, and Android. Distribution is driving competition. On that note, Google has upgraded Chrome with Gemini 3, adding a persistent AI side panel and new “auto browse” features that can handle multi-step tasks across the web. Users can compare options across tabs, summarize reviews, transform images, and, for Pro and Ultra subscribers, offload chores like form-filling, scheduling, travel research, and subscription management, with confirmation checkpoints for sensitive actions.

Moltbot, formerly Clawdbot, goes viral as a high-risk agentic assistant
The tech world is abuzz about open-source agent Moltbot, which operates autonomously and communicates via Telegram, WhatsApp or iMessage to take real actions across users’ systems. Its power comes with significant security risks due to full system access. For most non-tech users (myself included 🙋‍♂️), this is more situational awareness, but it could demonstrate where the technology is going in terms of usefulness and automation.

Drugmakers use AI to speed trials and regulatory filings, not discovery
Pharmaceutical companies are using AI to streamline site selection, enrollment, and regulatory documentation, shaving weeks or months off development timelines. True AI-driven drug discovery though, remains largely unrealized.

Doctors warm to AI scribes, but workflow and trust issues remain
Ambient AI scribes are reducing documentation burden and burnout for clinicians, with adoption accelerating at large health systems. Questions remain around hallucinations, billing impacts, patient trust, and access for smaller practices.
Link:

Notes to readers: It’s been a quiet week on the policy front, putting questions of a government shutdown aside. But there are no significant new actions on AI and health to flag.

Amodei’s AI warning doubles as a safety manifesto and market signal
Back to Anthropic CEO Amodei’s 38-page essay. TL;DR: I didn’t read. I listened to it instead.

As mentioned above, Anthropic CEO Dario Amodei dropped a 38-page essay this week warning that rapidly advancing AI will “test who we are as a species.” One of the threats he zeroes in on are the biological risks, which we already know to be true given researchers who were able to create new, synthetic viruses using AI last year.

So the piece is serious, dense, and widely discussed. It’s also the kind of thing many people intend to read and never quite get to. I put myself in that camp. Rather than block off an afternoon, I uploaded the essay into NotebookLM and generated an 18-minute podcast-style summary that I could listen to. Same content, far lower friction.

If you want the highlights without committing to the full essay, you can listen to the audio version I created below.

Gallup: workplace AI adoption is plateauing, not spreading evenly
Gallup’s Q4 2025 report finds nearly half of U.S. workers never use AI at work, even as managers and remote-capable roles pull further ahead. Lack of clear utility remains the main barrier.

Georgetown CSET: self-improving AI could accelerate fast and be harder to control
A report from Georgetown’s Center for Security and Emerging Technology warns that recursive self-improvement in AI could reduce transparency and oversight just as capabilities accelerate.

AlphaGenome brings the genome’s “dark matter” into focus
Google DeepMind’s AlphaGenome can interpret the 98% of human DNA that does not code for proteins, predicting the effects of single-letter mutations and identifying disease-linked variants that were previously difficult to detect.

Claude for Excel expands to Pro users
Anthropic has expanded Claude for Excel, enabling multi-sheet analysis, longer sessions, and safeguards against overwriting data.

Claude turns from advisor into operator with in-chat tools
Claude now lets users work directly inside tools like Slack, Asana, Figma, Canva, Box, Amplitude, and Hex, moving AI closer to being the interface layer for work.

Turn long reads into audio with ElevenReader
Per the note above, just like I did by using NotebookLM, ElevenLabs’s free ElevenReader app converts articles and blogs into short audiobooks for easier consumption. And if you want to stick with using NotebookLM, there’s a guide included in an edition that ran last week in the Rundown.

HBR: how to lead teams through AI anxiety
Harvard Business Review offers practical guidance for leaders navigating AI-related anxiety through transparency, psychological safety, and action.

Use AI to Work Smarter, Not Longer
HubSpot’s practical guide breaks down how tools like ChatGPT can boost productivity and creativity at work. It focuses on real use cases, automating routine tasks, sharpening decision-making, and freeing up time for higher-value work.

👉 Feb. 2, 2026 — AI and Mental Health: Risks, Responsibility, and Public Health, Johns Hopkins Bloomberg School of Public Health, Virtual & Baltimore, MD

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

👉 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