On June 1 the Joint Commission released the first AI certification built for healthcare organizations rather than for AI vendors. It does not check whether your software is safe. It checks whether you can govern it. Read alongside a new state law that bars AI from mental health treatment decisions, an FDA guidance that loosened oversight only where the clinician can review the logic, and a state-by-state regulatory map that keeps fracturing, the same move shows up four times. The point of accountability is shifting from the tool to the person using it.
Signal 1
The Joint Commission Launched the First Healthcare AI Certification, and It Grades Your Governance, Not the AI Product (June 1, 2026)
On June 1, 2026, the Joint Commission released its Responsible Use of AI in Healthcare certification, the first voluntary credential designed specifically for hospitals, critical access hospitals, and health systems rather than for the companies that build AI tools. The certification evaluates whether an organization has the governance structures, safeguards, monitoring processes, and staff education in place to use AI responsibly. It explicitly does not validate or certify any individual AI product. Any healthcare organization may apply, and an applicant does not have to be accredited by the Joint Commission to be eligible. The program delivers on guidance the Joint Commission and the Coalition for Health AI first issued in 2025, and it arrives as the Joint Commission notes that more than 80 percent of physicians already report using AI in professional settings. Sources: Joint Commission, June 1, 2026; Fierce Healthcare, June 1, 2026; Healthcare IT News, June 1, 2026.
What this means for you
The unit being held accountable is the practice, not the software. A tool that is FDA cleared or listed by a coalition no longer answers the question a surveyor will now ask, which is whether your own review process exists and can be shown to someone. A solo or small integrative practice will not pursue a hospital certification, and that is not the point. The point is that the body that defines what good looks like in clinical settings has now written down its standard, and that standard is a governance file, not a better algorithm. Whatever your size, the question it certifies against is the question you should be able to answer for yourself.
Signal 2
Maine Enacted LD 2082, Barring AI From Mental Health Treatment Decisions and Requiring Consent Before Any Ambient Recording (April 13, 2026)
Maine’s LD 2082, An Act to Regulate the Use of Artificial Intelligence in Providing Certain Mental Health Services, was approved on April 13, 2026, and is now in force. The law allows licensed mental health professionals to use AI for administrative and limited supplementary tasks. It expressly bars providers from using AI to make therapeutic communications or treatment decisions, or to interact with patients on its own. It also requires a provider to obtain patient consent before using ambient listening or other AI-powered recording tools, and it prohibits anyone from offering therapy or psychotherapy services to the public, including through AI, unless a licensed professional provides them. Enforcement authority sits with the state Department of Health and Human Services and the professional licensing boards. Sources: WTL Governance, April 2026; Maine Legislature LD 2082; New England Psychologist, 2026.
What this means for you
The law draws a line that the practitioner, not the vendor, is responsible for holding. It is not enough that a tool is marketed as a clinical assistant. The licensed professional has to be able to point to where the AI stopped and their own judgment began. The two operational triggers in the law are the ones most practices have not documented: a recording tool that captures a session needs consent on file, and any patient-facing language the AI drafts needs a human author who stands behind it. Maine is one state, but the structure of its rule is the structure others are copying.
Signal 3
The FDA Loosened Oversight of AI Clinical Decision Support, but Only Where the Clinician Can Independently Review the Logic (January 6, 2026)
The FDA’s revised clinical decision support guidance, issued January 6, 2026, expanded enforcement discretion for AI software that offers a single, clinically appropriate recommendation, on the condition that the provider can independently review the basis for that recommendation, including the data inputs and the underlying logic. The guidance applies this allowance to AI, including certain generative AI features, so long as clinicians can understand and verify how a recommendation was generated. The agency paired the lighter regulatory touch with a heavier emphasis on transparency and on automation bias, the tendency of a clinician to defer to a machine output. The guidance continues to cover only provider-facing tools and did not address consumer-facing chatbots or symptom checkers. Sources: Covington and Burling, January 2026; Faegre Drinker, January 2026; Arnold and Porter, January 2026.
What this means for you
Less regulation of the tool means more responsibility on the person using it. The FDA’s own framing assumes the clinician is checking the reasoning, which only works if you are actually positioned and trained to do that rather than accepting the output. The condition the agency attached is the whole point. The light touch applies where the logic is reviewable, so a tool whose reasoning you cannot see is a tool you are carrying more risk on, not less. If you cannot explain why a recommendation was made, the regulatory relief was never written for the way you are using it.
Signal 4
The State Map Keeps Fracturing: AI Regulatory Sandboxes in Four States and a New Focus on Payers Using AI to Downcode Claims (2026)
Manatt’s health AI policy tracker reports that in 2026, four states, Arizona, Illinois, New Hampshire, and Virginia, introduced AI sandbox or regulatory relief programs, while regulatory attention has turned toward payers using AI to downcode claims and toward AI in prior authorization. Nearly every state has now introduced AI legislation touching healthcare stakeholders, with only Wyoming and North Dakota sitting out. The result is not a single national rule settling into place. It is a widening set of state and function-specific expectations, several of which take effect on staggered 2026 and 2027 dates. Sources: Manatt Health AI Policy Tracker, 2026; Holland and Knight, May 2026.
What this means for you
The rules are not converging. They are multiplying and diverging by state and by function. A practice that works across state lines, offers telehealth, or bills multiple payers is now governed by several overlapping expectations at once, and the one that applies to you depends on where your patient sits and who pays the claim. There is no single federal rule to wait for. The practical defense is to treat your own list of states and payers as your compliance map, because no one is going to consolidate it for you.
The Pattern
Four bodies, moving independently, made the same decision. The Joint Commission built a certification that grades the organization rather than the tool. Maine wrote a law that holds the licensed professional, not the software, responsible for where AI stops. The FDA handed back regulatory slack, but only to clinicians who can review the logic themselves. And the states, in aggregate, are placing the burden of tracking a fractured rulebook on the practice that has to operate inside it. None of them is asking whether the software is good. Each is asking whether you can show your work. The standard arriving in 2026 is not a rule about technology. It is a rule about whether your governance is written down.
One Thing You Can Do This Week
Open a blank page and answer three questions about one AI tool you already use. Who looks at its output before it reaches a patient, how often does that review happen, and what is the process when the output is wrong. If you can answer those three in writing, you are most of the way to the standard every one of these bodies is now measuring against. If you cannot, you have just found the gap, and it took one page to find it.