The AI Clinic · Issue 011AI risk & trust
Check the claim before you build on it.
A new review maps everything that can go wrong once AI touches a chart
It's Monday, and seven stories survived the cut this week. The theme running through all of them: check the claim before you build anything on top of it.
- A peer-reviewed review catalogued the real vulnerabilities that show up once AI gets folded into patient care: data exposure, manipulation, and unclear accountability when something breaks.
- A free literature-review tool produced a 62-page review with 160 references in minutes, and the speed itself is the problem for anyone who supervises the writing.
- An early, exploratory study tested whether gait-sensor data alone could grade the severity of someone's low back pain.
The Big Story
A new review maps everything that can go wrong once AI touches a chart
A peer-reviewed review looked past the usual "is it accurate" question and catalogued the real vulnerabilities that show up once large language models get folded into patient care: data exposure, manipulation, unclear accountability when something breaks. It's a structured map of the risk, not a single scandal or breach.
Mark's read: If you're a practice owner writing an AI use policy, or a clinician who just started trusting a scribe with your notes, this is the primer to read before the next vendor contract lands on your desk. It's survey-level research, meaning broad instead of deep, so treat it as the checklist you bring to your vendor's security documentation, not a substitute for reading that documentation yourself.
The Rounds
A free AI tool wrote a 62-page lit review in minutes
A creator ran a free literature-review tool and watched it produce 62 pages with 160 linked references almost instantly, then spent the video explaining why that speed is the actual problem.
So what: If you supervise students, or write reviews yourself, watch this one. Outsourcing the writing outsources the synthesis, and the synthesis was the assignment.
The Rounds
Schools are dropping the blanket AI ban
More schools are trading flat bans for policies that teach students when and how to use a chatbot well, built on what teachers are actually seeing show up in the work.
So what: If you teach DPT students or run CE courses, it's a decent reference for writing a policy people follow, instead of one everyone quietly ignores.
The Rounds
The case for teaching AI in STEM, not just policing it
An editorial argues for folding AI into STEM curricula on purpose, with real guidance, instead of treating it only as something to catch and punish.
So what: It's opinion, not data, and general STEM rather than health-professions specific. Still worth a read if you're shaping curriculum policy and want the argument for teaching it, not just the argument for banning it.
In the Clinic
Gait sensors try to grade how bad someone's back pain is
Researchers pulled thousands of gait-analysis variables from 36 people with low back pain, testing whether movement data alone could separate the pain-free group from the high-pain group. It's an early, exploratory pipeline, not a validated tool.
Mark's read: Interesting as a proof of concept, but 36 people is a pilot, not a product. I'll also push back gently on the premise. Pain is a uniquely individual experience, and I'm wary of handing AI the job of "objectifying" something that resists being reduced to a number.
Try This
A better prompt for patient handouts that actually read at a 6th-grade level
Researchers tested a new prompting approach against the usual AI defaults and found it reliably hit plain-language reading targets that generic prompts keep missing.
So what: If you've ever asked ChatGPT for a handout and gotten back something written for a college grad, this gives you an actual template to fix it, instead of guessing at your own wording.
What else is happening
Filtering a literature search with Elicit (1 min watch)
The last word
That's the week. Seven stories, one caution: check before you build on it.
Get the next issue.
One email a week. The AI that matters for clinical work, in plain language.