An AI disclaimer is not a patient safeguard. It is a label. Since January 2025, California has required health facilities and physician practices that use generative AI to write patient communications about clinical information to say so, and to tell the patient how to reach a human. Texas followed with 2 laws of its own. The first, in effect since September 2025, requires practitioners who use AI for diagnosis or treatment recommendations to disclose it to patients and to review the AI-generated records. The second, a broader AI law in effect since January 2026, requires providers to disclose AI used in patient care.
Disclosure laws are spreading because they are easy to write and hard to argue against. Patients should know when a machine was involved in their care. But a disclaimer, by itself, protects no one from a wrong answer. It hands the job of judging the output to the person least equipped to do it.
What a disclaimer asks of a patient
Consider what a patient can do with a line that says a message was generated by AI. The patient cannot see which model wrote it or whether it misread a lab value along the way. The patient can only decide how much to trust the message as a whole. If the message reads well, most people will trust it. The errors that matter most in clinical writing are the ones that read well.
The most useful part of California’s law is not the label. It is the requirement, described in a summary by ArentFox Schiff, to include clear instructions on how to contact a human health care provider. That gives a patient somewhere to go. But a patient who does not know something is wrong has no reason to use it.
The law also contains a revealing exemption. If a licensed or certified provider reads and reviews the AI-generated communication, the disclaimer and the instructions are not required. In other words, the label appears precisely when no clinician checked the message. Whatever its intent, the effect is that patients see the warning in the cases where the real protection, a qualified person’s review, is absent.
A disclaimer tells the patient that no one checked. It does not check.
What actually protects patients
The safeguards that matter are the ones that operate before the output reaches the patient, and that leave a record afterward.
- Review before release. Clinical content generated by AI should be read by a qualified person before a patient sees it, at least until an organization has evidence about how often a given use goes wrong. California’s exemption already treats review as the thing that makes a disclaimer unnecessary.
- A record of how the output was made. The record should show which tool and version produced the output and which person approved it. Without that record, an error cannot be traced, and a pattern of errors cannot be found.
- Testing and monitoring. Accuracy measured on representative cases before deployment, and error reports collected and reviewed after it. A tool that has never been measured against the work it does is not ready to do it.
- A named person accountable. Someone in the organization should own each use of AI and have the authority to stop it.
California’s rules for health plans show what a safeguard regime looks like. Under SB 1120, which took effect alongside the disclosure law, a plan that uses AI or algorithms in utilization review must base determinations on the enrollee’s own medical history rather than solely on a group dataset. A physician or other competent provider must make the final medical necessity determination, and the tool must stay open to audit and be reviewed periodically for performance and outcomes. None of those requirements depends on what the member is told. All of them change what happens to the member.
Federal nondiscrimination rules work the same way. As we wrote in February, Section 1557 now requires covered entities to identify and mitigate the risk of discrimination in the decision support tools they use. That is an obligation to look inside the tool, not to post a notice about it.
Where disclosure still helps
None of this is an argument against disclosure. Patients have a reasonable claim to know how their care is being delivered, and some will want to ask for a person instead of a program. Disclosure also has an internal benefit that is easy to overlook. An organization that must label every AI-generated message has to know where AI is generating messages. That inventory is the first step toward governing it.
Disclosure works best when it carries substance. A notice that says what the tool did and whether a clinician reviewed the result gives a patient something to act on. A notice that says only that AI was involved gives the patient a feeling of caution without the means to use it.
The risk is in treating disclosure as the finish line. A state that requires a label and stops there has shifted responsibility from the organization that chose the tool to the patient who received the output. An organization that adds a label and stops there has complied with the law and protected no one.
What we build for
We build tools health plans use in utilization review, and our rules are written as safeguards rather than disclosures. A clinician makes every denial, because a machine can say yes but only a person can say no. Every recommendation shows its reasons, in the form of the criteria applied and the facts from the record that met or missed them. Every decision can be reconstructed, including which tool contributed what and which person approved it.
A member or a provider who receives a decision deserves to know how it was made. We think the most useful thing a decision letter can say is not that software assisted, but which criteria were applied and who made the call. That is information a person can check and appeal.
Labels have their place. But when the question is whether a patient is protected, the answer should be found in what happened before the message was sent, not in a sentence at the top of it.