One word for MDLinx. Then the rest.
“Moderately. The failure I worry about is not a dramatic one. It is a confident wrong answer delivered in fluent medical English to someone who has no way to check it. In surgery we have decades of practice at recognizing a bad result. We have very little practice at recognizing a bad output that reads well.”
The Reference That Does Not Exist
“It’s really difficult to recognize hallucinations or wrong information, because AI always feels like an authority, all-knowing. I’ve had experiences where I’m researching a topic and AI comes up with completely false references to support what it’s saying. So this is really the difficult part, and the one where the doctor can come in and rectify and explain what is actually happening, or what’s actually in the literature. So many times when I’m researching a topic, after I receive an answer I’ll ask for references, and then I’ll go look for the source to see if it exists, and read about it to make sure it’s factual. Sometimes, if the information seems strange, I’ll ask again in another question, and a lot of times it will recognize that it made a mistake and rectify it.”
The Diagnosis Nobody Made
“And one of the most dangerous answers AI can give is false reassurance, that you shouldn’t be worried, that it’s really nothing, when in fact it may be, and only a doctor can make that determination.”
The other direction too: a patient told a procedure is simple, expecting what the anatomy will not give.
What It Cannot Do
“I would never trust it with anything that requires a physical exam. AI has now been trained to recognize photos and evaluate lesions, but it cannot touch and feel the way we do when we perform a physical exam, to evaluate hernias or rectus diastasis, masses or lumps in the body or breast, skin laxity, and quality of skin.”
“It’s also not good at predicting results, especially in cosmetic or plastic surgery, where a patient may upload a photo and ask for the postoperative result, which may or may not be a good expectation or even a possible one.”
The Implant Precedent
We approve an implant on trial data, then track it for years, and sometimes learn it behaves differently in the field. Software ships, then changes. The version treating today’s patient may not be the one that was tested. Nobody is tracking it the way we track a breast implant.
| Implants | Medical AI |
|---|---|
| Approved on trial data | Validated on a test set |
| Tracked for years after | Rarely, and the model changes |
| Behaves differently in the field | Same, and faster |
| A named surgeon is accountable | Nobody, yet |
Where I Land
“I don’t think AI would realistically replace office visits. The main factor is the physical exam, which is absent. A lot of times there’s no check of the vitals either, unless a patient has a way to measure these and upload them at the same time. I think it’s a supplement to medical care.”
“I’ve always said that an educated patient is a better patient.”
Prepare with it. Let me examine you. The patient’s how-to is my earlier AI before the doctor visit read; the full MDLinx answer with the three safeguards is on agulloplasticsurgery.com, and the consult-prep version is on the Southwest Plastic Surgery post.
Why Choose Dr. Agullo
Double board-certified, by the American Board of Plastic Surgery and the American Board of Surgery. Mayo Clinic plastic surgery fellowship. Clinical Associate Professor of Plastic Surgery at Texas Tech University Health Sciences Center. Castle Connolly Top Doctor for 13 consecutive years. Harvard certification in AI in Healthcare.
Ready to Talk?
Bring the chatbot’s answer. I will check its references, then examine you. (915) 590-7900, text 1-866-814-0038, or book at agulloplasticsurgery.com/appointments. #StayBeautiful
@RealDrWorldWide on Instagram, TikTok, and Snapchat, @Agullo on X, or @AgulloPlasticSurgery on Facebook.
