In partnership with
THE OT ALGORITHM
Giphy
Hi friends,
This week's headline is a study I want us to read from the AI side more than from the shoulder-rehab side. I'm not a shoulder specialist, but I am an OTP who keeps getting asked whether a chatbot can write the home program.
Researchers in Türkiye asked two large language models, ChatGPT-4o and DeepSeek-R1, to write six-week exercise programs for subacromial pain syndrome. A physiotherapist then cleaned those programs up, including by removing suggestions like clap push-ups, and tested them against a program a physiotherapist designed from scratch.
We'll walk through what the AI actually did, what it left out, and what that means before anyone treats "the chatbot wrote it" as a finished plan.
We'll also reveal last week's Two Truths and AI answer, play a new round, and work through a research-mentor scenario about a student who wants ChatGPT to write the exercise protocol.
Let's get into it.
Note to readers: The ads and some of the links in these weekly newsletters contain affiliate links to help support my work. If you’d like to show support, all you have to do is click on them and look around their site a little bit. 🙂 Disclaimer: Placement of the ads included in these weekly issues is not an endorsement. Please do your own research before using any AI tools.
New AI tools launch every day. A few will change how you work. The rest will be forgotten by next month.
TLDR AI is the free newsletter that tells you which is which. Every morning, Anthropic and ex-Google engineers pick the new tools, updates, and open source releases worth trying, with a quick note on what each one does and who it's for.
You find the good ones early, without testing all of them.
One email, 5 minutes a day, read by 1.1M+ people. Get tomorrow's picks.
HOT TAKE THIS WEEK
The Safety Layer Was the Clinician
Gif by tlc_network on Giphy
When a student or a new grad asks whether ChatGPT can write a home program, I understand the appeal. Documentation is heavy, caseloads are heavy, and a structured six-week plan appearing in seconds feels like relief.
This week's headline put that hope to a real test. Two language models wrote shoulder programs. An experienced physiotherapist turned those outputs into master programs, and along the way removed suggestions such as plyometric wall push-ups, a push-up with a clap, and overhead plyometric ball throws. Then every group, including the conventional one, received the same in-clinic modalities and the same supervised sessions. Everyone improved on the main disability measure, and no group beat the others.
The finding that sticks with me is not that AI "matched" a physiotherapist. The programs people received were clinician-curated programs, and the authors say this trial cannot tell us what the models would have done on their own. A client following clap push-ups for a painful shoulder would have been working from the raw output, not from the study.
I'm reading this as an OTP looking at the methods, not as someone who treats shoulders every day. What I notice is how much of our work never entered the prompt at all: no occupation, no individual goals, and no question about what that arm is for. (A clap push-up on a painful shoulder is also a useful example to bring to the next "the chatbot can write it" conversation.)
