The AI Clinic · Issue 002Liability & trust
When the AI is wrong, you own it.
When the AI is wrong, the clinic owns it
This week the question every clinic is about to face got a real answer. When an AI tool is wrong, the liability lands on whoever put it to use, not the company that built it. Plus the FDA fast-tracks AI that drafts radiology reports, and a model that tries to call persistent pain before a knee replacement.
Last night I built a clinical research assistant live, from a PubMed question to a referenced answer to a patient handout in one sitting. The replay is up, and there is a short window to get a 1:1 build-plan call with me.
- A court ruled that when an AI tool is wrong, the fault sits with whoever put it to use, not the company that built it.
- The FDA gave breakthrough status to two tools that draft radiology reports before a human reads the scan.
- The webinar replay is up, and the 1:1 build-plan call offer closes Tuesday night.
The Big Story
When the AI is wrong, the clinic owns it
A court in Germany ruled that when an AI tool produces a wrong output, the liability lands on the person or business that deployed it, not the company that built the model. It is one ruling in one country, but it points at the question every clinic is about to face. The vendor sells you the tool. You are the one who used it on a patient.
Think about where this touches your day. An AI scribe that mishears a dose. A chatbot that hands a patient the wrong home program. A summary tool that drops the one red flag that mattered. If the courts keep landing here, none of that is the vendor's problem. It is yours, and it has your signature on it.
Mark's read: This is not a reason to stay away from AI. It is a reason to know how it works before you let it near a patient. A human reads the output, a human signs the note, a human owns the call. That has always been the job. AI does not change it, it just raises the stakes on doing it well.
In the Clinic
The FDA fast-tracked AI that drafts radiology reports
The FDA gave breakthrough designation to two tools that write a draft radiology report before the radiologist opens the scan. The pitch is faster turnaround on imaging. For anyone who orders a chest X-ray and waits on the read, this is the kind of thing that quietly changes how fast a finding lands in your EHR.
In the clinic: a draft is not a read. The speed is real and useful, and the report still arrives with a machine's first guess baked into it. Know which part a human signed off on before you treat off of it.
The Business of Care
AI returns money when the owner runs it, not the vendor
Two reports landed in the same week. KPMG found that AI strategies owned by the person at the top returned about three times the ROI of the ones handed off. Bain found that close to 40 percent of companies got less than 10 percent in cost savings out of their AI spend. Read together, they say something simple. AI does not pay you back because you bought it. It pays you back when the person who runs the place stays in the room.
For owners: a small clinic rarely has the volume a hospital system does, so the savings story is even thinner if you delegate it and walk away. If you are going to put money into a tool, put an hour of your own attention into it too. That hour is where the return comes from.
The Big Picture
Choosing to stay human
Ethan Mollick wrote on what happens now that AI writing is flooding every channel we read. When the words get cheap and endless, the thing that gets rare is a real human voice behind them. For anyone who teaches, publishes, or just emails patients, your own voice is becoming the signal people trust.
Mark's read: I got called out once for sending an email a machine clearly wrote, and it stung because it was true. I use AI every day now, but the voice has to stay mine. Let it draft, let it research, let it carry the load. Do not let it become the person your patients think they are hearing from.
Research Worth Reading
A model that tries to call persistent pain before the knee replacement
A new machine-learning model uses pre-surgery data to predict which patients will still have persistent pain after a total knee replacement. If it holds up, you get a flag before the operation instead of a surprise three months after it.
For clinicians: a flag like this changes the conversation you have before surgery, and it changes who you send to prehab or pain psychology and when. Read it as a prompt for a better talk, not a verdict on the patient. The prediction is a starting point, the person in front of you is the rest of it.
Try This
Map your clinic's AI liability in ten minutes
Take the Big Story and make it concrete this week. Open a blank page and list every AI tool that touches a patient in your clinic, the scribe, the chatbot, the intake summarizer, the home-program generator. Next to each one, write the name of the human who reads its output before it reaches the patient. If any line is blank, you just found the thing to fix first. It is the cheapest risk audit you will run all year.
What else is happening
The last word
A court says the output is yours. A study says the voice that stays human is the one people trust. Both point the same way this week. Use the machine for everything it is good at, and keep your name on the part that matters.
If you have thoughts on the newsletter, good or bad or something missing, reply to any issue and let me know. I read every one.
Get the next issue.
One email a week. The AI that matters for clinical work, in plain language.