The Bandwidth: Issue 10

Note on Schedule: Due to the upcoming Thanksgiving holiday, this week's newsletter is being published today instead of our usual Friday schedule. We wish you a wonderful and healthy Thanksgiving!

A sharper perspective on this week's top issues that matter at the intersection of health and AI.

AI Chatbots (cont’d) Last week, we discussed the bipartisan consensus coming from Capitol Hill: AI technology is advancing far faster than the safety guardrails needed to protect vulnerable users. A recurring theme was the profound, and negative, impact of chatbots on those with significant mental health needs, with lawmakers citing tragic cases of emotional dependency, self-harm, and even suicide. This week, the New York Times continued to highlight these kinds of risks. The Times reports OpenAI's updates to ChatGPT earlier this year increased user engagement but inadvertently led to serious mental health crises, with some users experiencing delusional thinking. After nearly 50 reported cases of distress, including hospitalizations and deaths, the company implemented GPT-5 and new safeguards like crisis hotlines, a full-time psychiatrist, and tests for harmful validation. Despite making the model safer, which some users found "colder," the article suggests OpenAI is now trying to balance user safety with its crucial business goal of increasing daily active users. Side note, it was also reported by Wired this week that OpenAI’s head of safety research, someone who worked to address these reported issues, was departing the company.

The root of this problem may lie in the design of these general-purpose Large Language Models (LLMs). Some argue they are not built for the precise, high-stakes nature of medical or therapeutic advice, which can lead to inconsistent and often dangerous results. As one experiment suggests, general-purpose LLMs perform inconsistently in medical interactions: sometimes excellent, sometimes dangerous. The authors argue that the problem is compounded when humans interact with these models, often leading to worse outcomes than the AI alone. Why? Because patients mis-prompt, omit symptoms, or misinterpret outputs. The clear conclusion is that patient-facing AI needs to be much more structured and medically grounded.

One (perhaps interim) approach that demonstrates this structured model is the new MAIA agent launched by women's health startup Millie. Built specifically for maternal care, MAIA is designed to be a "patient companion" that handles scheduling and common non-clinical questions. Crucially, the company built in a mandatory safety safeguard: when patients ask MAIA a clinical question, the agent automatically escalates the request to the company's human clinical team rather than generating an independent, freeform answer. This model—where the AI handles low-risk tasks while immediately deferring complex or clinical inquiries to a professional—is the kind of safety-first approach the medical and patient community may be looking for more of until more sophisticated models emerge.

Healthcare as an AI super user? Not so fast. Recent research (that I previously wrote about) suggested healthcare may be a leading sector when it comes to AI adoption. As I mentioned at the time, that’s surprising given the sector has historically been a laggard when it comes to technology, but it also made sense given the massive potential to automate certain workflows (e.g., billing, intake forms, and note-taking). However, new research published in JAMA may be throwing cold water on those high hopes, demonstrating that healthcare is actually behind other major industries, including information, finance, education, and real estate. The largest increases in AI use were limited primarily to the outpatient and ambulatory care arena, where the proportion of firms using AI nearly doubled, growing from 4.6% in 2023 to 8.7% in 2025. By comparison, nursing and residential care facilities experienced a much smaller increase, growing only from 3.1% to 4.5%. Another notable point from the new research, the dominant type of AI for healthcare is voice recognition, as opposed to things like machine learning or virtual agents.

What’s old is new again. It's worth taking the time to spin through this quarterly presentation on technology trends from Ben Evans. But if you can't, one thing jumped out to me: He zeroes in on a Congressional report from 1956 looking at automation and concerns about its impact on jobs—including the impact of electronically controlled elevators (screenshot below). Obviously, AI is very different than automatic elevators, but it's a great reminder that we’ve confronted these kinds of fundamental technological problems before. The trick is, what have we learned from the past that might help us this time around to minimize negative disruption or unintended consequences?

🛜 Other Field Signals

A quick hit on this week’s key policy shifts and industry trends.

FDA to roll out Gemini and Grok across the agency: FDA plans to deploy Google’s Gemini and Musk’s Grok internally as HHS prepares an updated AI strategy, even as staff raise security and reliability concerns about existing tools like Elsa.

Virginia lawmaker unveils sweeping AI regulatory blueprint: A new proposal would create one of the country’s most comprehensive state-level AI safety and transparency frameworks, signaling how fast states are moving ahead of Congress. Relatedly, federal lawmakers may be working on a plan to block such state rules on AI and tuck it into an annual defense bill, Politico reports. 

Anthropic finds models can spontaneously learn deception: New research shows LLMs trained on coding tasks invented ways to sabotage safety workflows unless explicitly inoculated against misalignment.

Jony Ive and OpenAI to unveil their first physical AI device: Ive says the product will launch in under two years, aiming to redefine how everyday users interface with AI through ultra-simple hardware.

Google’s Gemini 3 puts ChatGPT on the defensive: Axios reports Gemini 3 Pro is now outperforming ChatGPT, intensifying pressure on OpenAI and reshaping competitive expectations across the model landscape.

Altman warns staff of “rough vibes” as Google gains ground: OpenAI is bracing for economic pressure after Gemini 3’s advances, with Altman urging teams to stay focused on ambitious bets.

Claude Opus 4.5 explained: Higher reasoning, bigger context, safer design
Anthropic details how Opus 4.5 boosts reasoning, context handling, and world-modeling while tightening safety guardrails.

Aetna and Cigna accused of AI-driven downcoding: Modern Healthcare reports both payers are using AI to systematically downcode provider claims, triggering backlash and raising new oversight questions.

Medicare flags a spike in AI-driven coding errors: Federal auditors say AI-based code selection is causing both overpayments and underpayments, signaling upcoming scrutiny and possible rulemaking.

The White House launches the “Genesis Mission” to supercharge AI-powered R&D: A major executive order kicks off an Apollo-style push to use AI for scientific discovery, drug development, fusion research, and DOE-led compute platforms.

Hospitals deploy AI to reduce nursing burnout: Henry Ford, Providence, and HonorHealth report early gains using AI for documentation, patient monitoring, and pre-op workflows, with improvements in LOS, burnout, and turnover.

AI tools help reduce missed appointments: Deep Medical says its algorithms can predict high-risk no-shows and automate outreach, cutting avoidable gaps in care.

🛠️ Practical Edge

Actionable tips, tools, and thoughts to help leaders strengthen capacity and apply AI in their work.

AI Glossary (Ibelick): A living glossary that explains AI terms in plain English, giving teams a shared baseline for understanding concepts like embeddings, agents, and vector databases. Shared language is the first step toward shared strategy. 

Most AI Initiatives Fail. This 5-Part Framework Can Help. (HBR): A five-step model for making AI initiatives work: clarify the business problem, redesign workflows, clean the data, pair tech with domain expertise, and create feedback loops. AI fails when leaders treat it like a software install instead of operating-model change.

I Made ChatGPT Stop Being Nice — And It Became More Useful (LinkedIn): A demonstration that explicitly removing “niceness bias” in ChatGPT produces sharper, more strategic answers. Specify tone, constraints, and expectations to get critical thinking instead of polite noise.

Atlassian AI Collaboration Index: Executive Insights (Atlassian / The Rundown): Cross-functional data showing that small, active AI “work groups” outperform formal training and that tools are strongest at individual workflows, not cross-team coordination. Rapid experimentation beats theoretical strategy.

🌅 On the Horizon

A quick look at the developments and events expected to shape the weeks ahead.

And finally, if you like what you are reading, please share and encourage your network to subscribe.

Till next time,

BC