The AI Clinic · Issue 012AI liability & trust
Read the note before you sign it.
Ambient scribes can quietly write the wrong thing into your chart
It's Monday, and eight stories cleared the bar this week, more than usual. The thread running through all of them: before you hand over your chart, your practice, or your job description to a piece of software, find out who's accountable when it gets it wrong.
- Experts say ambient AI scribes are introducing documentation errors that are easy to miss, and those errors are starting to appear in malpractice cases.
- OpenAI released GPT-6 Astra, its biggest model yet, with workflow automation as its real strength.
- Hinge Health and Sword Health are buying companies outside MSK, and one operator says back-office automation is coming for PT.
The Big Story
Ambient scribes can quietly write the wrong thing into your chart
Legal and health-IT experts are flagging a problem clinics don't like to say out loud. Ambient AI scribes can introduce documentation errors that are easy to miss during a full patient day, and as adoption spreads across health systems, those errors are starting to show up in real malpractice cases.
Mark's read: A scribe error in your chart is still your documentation. It's still your liability, and depending on what it says, it could be your patient's safety. Don't sign off on an AI-generated note without reading it line by line against what actually happened in the room. The tool saved you time typing. It didn't save you from reading.
The Rounds
OpenAI just released its biggest model yet
OpenAI shipped GPT-6 Astra, calling it a big jump in capability for cybersecurity, professional work, software engineering, and computer use. It's also the first OpenAI model to cross the company's own line for critical cybersecurity capability, something it says it's managing carefully.
So what: Not clinic-specific, but this is the kind of model upgrade that eventually shows up inside the scribes and chatbots your practice already runs. Workflow automation looks like the real strength here, worth watching even if you never touch the model directly.
The Rounds
Digital MSK platforms are quietly buying their way out of MSK
Hinge Health and Sword Health spent years arguing traditional PT was too expensive and didn't scale. Now they're buying a gastroenterology company and a meditation app.
So what: Reads less like growth and more like an admission that MSK alone wasn't the story investors were promised. The competitive threat these platforms pose to independent clinics may be less durable than it looked a few years ago, though the payer contracts they've already signed aren't going anywhere yet.
The Rounds
Automation is coming for your back office
Using a dental group as the case study, one operator argues that within a decade, well-run healthcare companies will run with far fewer people in revenue-cycle management, HR, hiring, and payer contracting. He frames it around dental, but says outright that PT and other clinical verticals are next.
Mark's read: This is the bet behind Work, Understood. We're automating the VA work and the backlogged owner tasks that eat up a practice owner's week, so that time goes to growth instead. Look at your own back office now, billing follow-up, credentialing, scheduling, and ask what's actually still worth a person doing.
The Rounds
The more capable the AI, the harder it is to hold accountable
As agentic AI and deep learning spread through healthcare, this paper argues we're trading the older, simpler way of modeling disease for network-medicine approaches that treat biology as interconnected systems. The tradeoff: more powerful tools that are harder for clinicians and patients to actually understand.
So what: The AI gets more sophisticated. The ethical and legal responsibility still sits with you. A decent primer if you teach AI ethics or scope of practice to students or staff.
In the Clinic
Nurses are turning AI from a tool into a partner
A new study interviewed nurses to map how generative AI is shifting from passive support into something closer to a collaborative partner in daily work, then built a framework for how that shift actually happens.
Mark's read: The framework is nursing-specific, but the pattern isn't. Frontline staff shaping how a tool actually gets used, instead of just being handed a policy, is exactly how PTs and support staff should be brought into choosing your clinic's AI tools.
In the Clinic
AI wrote patient vignettes from video blogs, then checked them with the patients
Researchers compared hand-written clinical vignettes to ones built by AI from patients describing their chronic pain on video, then went back to the patients to see which version actually captured their experience.
Mark's read: If this holds up, it's a way to build real patient-experience material fast, for teaching or patient education, without the hours of manual qualitative work it usually takes faculty.
Try This
Practice the cost conversation before you're standing in the room
A new study tested whether an AI model can help clinicians prepare for the conversation about the cost of a total knee replacement, a topic most AI-in-healthcare research skips in favor of diagnosis and general medical questions.
So what: Cost talk is a known source of patient anxiety, no-shows, and complaints, in cash-based practices and high-deductible plans alike. This week, feed your AI assistant the details of one upcoming expensive case, the procedure, the estimated cost, the payer, and ask it to draft the talking points you'd use to walk the patient through it before you're doing it live.
The last word
That's the week. Eight stories, one thread: know who's accountable before you hand off the work.
Get the next issue.
One email a week. The AI that matters for clinical work, in plain language.