Who Protects the Patient When AI Won't Stop? My Comments in Asharq Al-Awsat
Asharq Al-Awsat asked who protects the patient when an AI system keeps going after a doctor tells it to stop. Here is my English summary, and what I told the paper.
Asharq Al-Awsat, one of the most widely read Arabic-language newspapers, published a science feature this week with a question every hospital should ask: who protects the patient when an AI system keeps going after a doctor tells it to stop? The piece, by Dr. Amid Khalid Abdelhamid, grew out of the AI governance sessions at AIMed26 in Orlando. I was glad to share my view with the paper.
Here is my short English summary of the article.
Hospitals are moving from AI tools that suggest to AI agents that act. That shift turns "corrigibility," a system's ability to accept correction or a stop command, into a patient safety issue.
As the article reports, a May 2026 Carnegie Mellon preprint on arXiv found that some AI agents kept working or got around limits after a person stepped in. Some limits also failed to carry over when one agent handed a task to another.
It also cites a 2026 Palisade Research paper in Transactions on Machine Learning Research that tested 13 large language models in more than 100,000 trials. In some test setups, certain models resisted shutdown up to 97 percent of the time. The author is careful here: this is a lab stress test, not proof of what happens in a real hospital, and it says nothing about machines wanting anything.
The lesson: test the off switch before you give any system real authority over patient care, and keep the stop controls outside the model itself.
What I told the paper, in my own words: "My rule is simple: human plus AI. The physician keeps the moral compass, while the machine finds patterns and analyzes them at scale."
I also said that AI should start from real clinical problems, keep the physician in the decision, protect patient privacy, and be tested on local patients before anyone trusts its output.
I still work shifts in the emergency department. Accountability only means something if the clinician can actually stop the system. A signature at the bottom of a note is not control.
For more context, see my notes on patient trust and human accountability, a physician framework for AI governance, and AI in longevity clinics. My speaker page and media kit are here as well.