On November 9, Representative Salman Bhojani filed Texas HB 510 for the 2027 legislative session. In 5 pages, it answers a question most rules for AI leave blank: who can stop the machine?
The bill covers a “high-risk artificial intelligence system,” defined in Sec. 552.151(3) as one “capable of independently initiating or executing an action that presents a reasonably foreseeable risk of substantial harm” to the health or safety of individuals, to property, to critical infrastructure, or to the security of information systems.
Sec. 552.151(1): 1 person, 5 verbs
The center of the bill is a person. An “authorized human operator” is an individual “designated by a deployer who possesses the training, authority, and technical access necessary” to monitor the system and to “approve, reject, modify, override, and terminate, as applicable,” its decisions, proposed actions, or ongoing actions. The developer must build, test, and document the controls (Sec. 552.153).
Oversight is real only when the person holding it has the information, the expertise, the authority and the time to act. HB 510 gives its operator three of the four: training, authority, and technical access. Time is missing from the definition of the person. The bill recognizes that a human may not be able to act in time, and its answer is more mechanisms. It does not require the system to wait for the human.
The word “independently” is tricky for medicine
The bill reaches systems that initiate or execute an action on their own. In medicine, most AI does something else: it recommends, and a clinician acts. On a plain reading, a sepsis alert, a diagnostic suggestion, or a draft note that a physician signs falls outside the bill, because a human executes the action. However, a system that uses agents to do an annual exam, places or renews an order without a provider’s oversight, or denies an insurance request falls inside.
Yet in medicine the risk sits equally in the first group. A clinician who signs what a tool suggests is in the loop; whether that clinician is in control is a separate question. HB 510 guards against the machine that acts alone. The machine that persuades the person who acts is outside its reach.
Sec. 552.155: who is accountable for the failure
Read for accountability, the bill assigns the cost of failure in a clear order. The developer who has built, tested, and documented the mechanisms “is not liable for a violation of this subchapter solely because a deployer fails to install, enable, maintain, or properly use” them. The deployer, for most of us the hospital or health system, carries the rest.
The operator holds the authority to stop the system, but the protection in Sec. 552.155 covers an employee who reports a violation to the attorney general or helps an investigation. It does not speak to the operator who uses the override, halts a system the institution depends on, and is asked the next morning to explain the delay. No medical board or health agency has a role.
The patient appears only as one of the “individuals” whose health or safety is at risk. Nothing in HB 510 tells the patient who can stop the system.
Last week’s Standing Order, “Who answers when the machine holds the license?”, asked whom a patient can name when an AI decision goes wrong; HB 510 requires a name, but that name stays inside the institution.
Before the first test
For a health system, three things follow. First, list every AI system that can act on its own, and name its operator in writing, with the 5 verbs mapped to controls that exist. Second, give that person time: an override exists only on paper if the workflow leaves no room to use it. Third, protect the stop in policy: a good-faith override is reviewed as a safety event, recorded, and never counted against the person who made it.
In your institution, who can stop the machine, and what happens to that person the day they do?
Sources. Texas HB 510 (90R), introduced version, filed November 9, 2026, by Rep. Salman Bhojani · Texas HB 149 (89R, 2025), the Texas Responsible Artificial Intelligence Governance Act, which created Chapter 552 of the Texas Business & Commerce Code · Nonalgorithmic. Who answers when the machine holds the license? The Standing Order, November 10, 2026

