The Pennsylvania House of Representatives voted 144-59 on September 30, 2026, to pass House Bill 1925, which would require hospitals, imaging centers, cancer treatment centers and physician practices to tell patients when they use artificial intelligence (AI) in clinical decision-making. The same bill bars health insurers from denying coverage on the strength of an algorithm alone. On October 14 it was referred to the Senate Institutional Sustainability and Innovation Committee, so it is not law yet.
The prime sponsor is Rep. Arvind Venkat, a Democrat and an emergency physician by training. According to The Center Square, 41 Republicans joined every Democrat on final passage. The bill writes parallel rules for health care facilities, insurers, and Medicaid and CHIP managed care plans, each overseen by a different state department.
What HB 1925 asks of hospitals and imaging centers
The definition of a facility is broad: hospitals, ambulatory surgical facilities, cancer treatment centers, laboratories, imaging and diagnostic facilities, and any office or clinic that employs a Pennsylvania-licensed provider. A facility that uses AI for “clinical decision making”, defined as a provider’s professional judgment directly pertaining to patient care, would have to explain that use in plain language on its website. It would also have to inform patients in writing, in the general consent for treatment or another written form, before or at the time of service. The notice is required once per episode of care.
Patient messages get their own rule. When AI generates personalized written or verbal communications about a patient’s clinical information, the message must carry a clear disclaimer that AI produced it and explain how to reach a human. Administrative messages such as scheduling and billing are excluded, as are messages a provider has individually read, reviewed and approved.
Governance duties follow. AI may not supersede the clinician’s decision, and no provider may be penalized for departing from an algorithm. Systems must not discriminate or use patient data beyond their stated purpose, and their performance must be reviewed at least annually. Facilities need written policies, a governance structure, an annual compliance statement to the Department of Health and records kept for at least five years. Exempt are “validated, static decision-support tools” and tools for administration, scheduling, scribe applications or clinical calculators.
Insurers need a human review before a denial
Insurers would have to disclose AI use in utilization review to network providers and enrollees. The model must weigh the individual patient’s history and may not produce an output based solely on group data. Before denying, reducing or terminating benefits, prior authorization denials included, the reviewing clinician must examine the record, document that review and reach a judgment independent of the AI recommendation.
Across all three chapters, civil penalties top out at USD 10,000 per knowing or willful violation, USD 5,000 per negligent violation and USD 1,000 for a first negligent lapse by an entity acting in good faith. Each failure to disclose counts separately, with an annual cap of USD 100,000 per entity, and the Attorney General could sue under the state’s unfair trade practices law. The floor version is softer than the May committee draft, which required quarterly reviews and a USD 250,000 cap. If signed, the law takes effect one year after enactment.
“Candidly, the public does not trust that artificial intelligence is being deployed for its benefit,” Venkat told colleagues, according to The Center Square. Republican Rep. Jason Ortitay called the bill an imperfect but “solid first, balanced step,” adding that “trust comes from knowing there is still a real person on the other end.”
One more piece in a state-by-state patchwork
Pennsylvania follows paths other states have opened. California’s AB 3030 has required, since January 2025, a disclaimer and human-contact instructions on clinical messages produced by generative AI, with the same carve-out for provider-reviewed messages. California’s SB 1120 reserves medical-necessity decisions for licensed professionals. In Texas, SB 1188, in force since September 2025, requires practitioners who use AI for diagnostic purposes to review AI-created records and disclose that use to patients. Manatt Health’s Q2 2026 tracker counted more than 280 state health-AI bills introduced this year and 33 enacted across 22 states.
The federal backdrop is tense. Executive Order 14365, signed in December 2025, directed the Justice Department to challenge state AI laws it deems burdensome. On the day it passed HB 1925, the Pennsylvania House also adopted a resolution urging the White House to stop threatening states that regulate AI. The American College of Radiology (ACR) expects more proposals on patient disclosure, physician oversight and utilization review in 2027 legislative sessions.
What it means for radiology and radiation oncology
Imaging facilities and cancer centers are named in the bill, but the practical question is open: which tools support “clinical decision making,” and which fall under the exemption for “validated, static” tools? The bill defines neither term. Many FDA-authorized imaging algorithms, a group that accounts for 76% of the agency’s AI authorizations, ship as locked models. Whether that makes them “static” is for Department of Health rules to settle. Until then, intracranial hemorrhage triage, mammography detection and radiotherapy auto-contouring sit in a gray zone. That reading is ours, not the bill’s.
The clearest effect would land on patient messaging. Plain-language report summaries, like those tested when GPT-5 helped patients understand their reports, and automated portal replies would need an AI disclaimer unless a physician reviews and approves each one. That ties into the ACR’s patient guide to AI report summaries and the problem of patients learning of cancer via the portal. The annual review and vendor-contract duties also back the case that radiology AI needs a management plan first.
Brazil already asks physicians to tell patients
Brazil’s Federal Council of Medicine (CFM) Resolution 2,454/2026, published in the official gazette on February 27, 2026, and in force 180 days later, requires that patients be told clearly whenever AI is used as relevant support in their care, with the use recorded in the chart. It also forbids delegating the communication of diagnoses, prognoses or treatment decisions to AI, a stricter line than Pennsylvania’s, which allows AI-generated messages as long as they are labeled.
Brazil’s AI bill, PL 2338/2023, passed the Senate on December 10, 2024. Its Article 14 classifies health applications that assist diagnosis and medical procedures as high risk when there is relevant risk to physical or mental integrity. In the Chamber of Deputies, a special committee with Rep. Aguinaldo Ribeiro as rapporteur still listed the bill as “awaiting report” on October 27, 2026. Software as a medical device is registered with ANVISA under RDC 657/2022; telling patients about AI use is, for now, an ethical duty set by the CFM.
Caveats and the Senate road ahead
HB 1925 still needs a Senate committee vote, a Senate floor vote and the governor’s signature. The 2025–2026 session ends on November 30, 2026, and bills not enacted by then must be reintroduced. Senators may rewrite the text, as Ortitay acknowledged, and its real reach will depend on Department of Health rules and on the risk of a federal challenge. The vote count and every obligation described here were checked against the official bill text (Printer’s No. 3963) and the General Assembly’s bill history.
Source: Radiology Business — Pennsylvania House passes bill requiring AI disclosure in healthcare; official text and history: Pennsylvania General Assembly — House Bill 1925 (2025–2026)