Standardizing disclosure of artificial intelligence in radiology reports: a proposed framework

Standardizing disclosure of artificial intelligence in radiology reports: a proposed framework Mauro Bergui1* and Alessandro Stecco2 As artificial intelligence transforms radiology workflows, from image segmentation to report drafting, a critical question emerges: should reports explicitly disclose their level of involvement? Transparency is foundational to ethical practice, patient autonomy, and professional accountability in an increasingly automated field. The debate balances innovation’s benefits, such as faster turnaround, reduced fatigue, with risks like algorithmic bias or opaque decision-making. Disclosure empowers patients and clinicians to contextualize findings, like labeling allergens in food. For instance, knowing that an artificial intelligence system generated 80% of a routine chest X-ray report could shape perceptions of reliability, especially given public concerns about “black box” technologies [1]. It also aids medico-legal clarity, distinguishing human oversight from artificial intelligence limitations [2]. But without standards, disclosure practices remain inconsistent [3]. To explore how transparency might be standardized, we looked at a regulatory framework that has already addressed a comparable challenge. The Society of Automotive Engineers' autonomy levels, originally developed for self-driving vehicles, may be adapted to artificial intelligence involvement in radiology reports, creating a scalable and intuitive framework. These levels range from fully manual to fully autonomous processes: ● Level 0: no artificial intelligence; the radiologist manually drafts the report (e.g., describing an MRI brain scan without software aid). ● Level 1: basic aids, such as grammar checks or standard phrases for techniques (e.g., “CT performed with contrast” auto-inserted). ● Level 2: partial automation, where artificial intelligence expands bullet-point inputs (e.g., “lung nodule, 1 cm, sharp edge”) into a narrative, fully reviewed by the radiologist. ● Level 3: conditional proposals, where artificial intelligence drafts complete reports from image analysis (e.g., detecting a lung nodule on CT), validated by the radiologist. ● Level 4: high autonomy, where artificial intelligence finalizes routine reports (e.g., normal mammograms), with radiologists auditing a subset. ● Level 5: full autonomy, where artificial intelligence handles end-to-end processing, with radiologists in supervisory roles, a scenario not yet realized. This simple framework ensures clarity. For instance, a report footer stating “Level 2: draft from key findings, radiologist-approved” informs stakeholders of the humanartificial intelligence interplay. Like vehicle displays signaling “Autopilot engaged,” disclosures could include confidence scores (e.g., “90% probability of benignity”) or audit trails, reducing errors in complex cases like multilesion MRI scans [4]. Although guidelines such as CLAIM and GAMER have helped improve transparency in research reporting, clinical reporting serves a different purpose and will need solutions tailored to real-world constraints [5, 6]. Any move toward standardized disclosure will require extensive stakeholder engagement, both in defining standards and in determining which disclosures should be mandatory. Consensus should involve patients, radiologists, scientific associations, system developers, legal experts, insurers, and regulators, recognizing that expectations and resources vary widely across practice settings.The recent adoption of the EU Artificial Intelligence Act adds further relevance to this discussion. Under the Act, medical artificial intelligence systems that support diagnostic decision-making may be considered “high-risk,” triggering obligations including lifecycle risk management, robust data governance, human oversight, and postmarket monitoring [7, 8]. Exploring whether a reporting taxonomy might support regulatory compliance could help reduce uncertainty and clarify accountability as automation evolves. Any such alignment should be considered exploratory at this stage, pending further expert input. Professional societies, such as the European Society of Radiology, could facilitate this dialogue by coordinating pilot studies, stakeholder workshops, and patient-facing communication research. These steps would help determine whether disclosure of artificial intelligence involvement strengthens trust, improves safety, and is feasible within existing reporting systems. It may also enhance interpretability for clinicians who revisit or validate prior imaging assessments. The goal is not to introduce unnecessary complexity, but to understand how transp