![\"A](\"https:\/\/rtmedical.com.br\/wp-content\/uploads\/2026\/08\/paciente-le-laudo-radiologico.jpg\")
“The report is written for a purpose, to help a medical expert take the next step,” says Morris Panner, president of Intelerad, now part of GE HealthCare, in an interview with Imaging Technology News<\/em>. “But as a patient, the main concern is, ‘Am I OK?'” The gap between those two goals is exactly where anxiety settles in \u2014 and where large language models arrived without asking permission.<\/p>\n Panner experienced the problem firsthand. After receiving the results of a personal imaging study, he loaded the report into an AI engine and asked the model to interpret it. “It told me, in general, it looks like you’re fine, although there’s a question or two I might want to ask a doctor.” His reason for not bothering his own physician says a lot about the state of the system: “I knew my physician was probably triaging and saying that if I’m basically OK, that’s pretty far down the list. Why not just spare them that kind of additional pressure, which is already adding to their impossible day?”<\/p>\n He is far from alone. According to figures cited in the report, ChatGPT fields health and wellness questions from 230 million people worldwide every week. And there is evidence the approach works: a study published this year in Radiology Advances<\/em> assessed language-model summaries of lung cancer screening reports and found higher self-reported comprehension along with reduced anxiety. That is not a detail \u2014 it means automated translation of the report has a measurable clinical effect on patient experience, for the better.<\/p>\n Professional bodies have already engaged the topic. The discussion gained substance when the American College of Radiology issued a patient guide to AI-generated report summaries<\/a>, and grew more concrete with data showing that patient-friendly reports improved family comprehension<\/a>. What Panner’s interview adds is the perspective of the people building the software: the change is not in content, it is in channel.<\/p>\n Here lies the thorniest debate in the specialty. “That’s one of those topics where, if I go to dinner with five radiologists, I can start a pretty great debate,” Panner jokes. On one side, leaders such as Keith Hentel, MD, executive vice chair of the Department of Radiology at NewYork-Presbyterian Weill Cornell Medical Center, champion radiologists speaking directly with patients. On the other, some argue the radiologist is the “doctor’s doctor”, whose role is to give specific technical guidance to whoever ordered the study.<\/p>\n The resistance has a legitimate basis, and Panner frames it well: “Radiologists tell me that their ‘product’ is the diagnosis that goes out under their name. I signed this report, that is my diagnosis, my professional judgment, my reputation, my life on the line. So what you have isn’t a question of a change of output. You have change of access and change of collaboration.” Department leaders should read that twice: nobody is asking radiologists to change what they write. What changed is who reads it.<\/p>\n For technology vendors, the operational answer splits into two fronts. The first is the multimodal report: a single document carrying the technical layer aimed at the clinician and a plain-language layer aimed at the patient, generated inside the reporting system and under the radiologist’s control. “You can take a report and have it be multimodal, where you can have a summary that’s very patient friendly, and you can specify how you want that to work,” Panner explains.<\/p>\n The second front is digital-first workflows designed for circulation. A report that disseminates easily, lives in many places and efficiently gathers input from multiple specialists \u2014 the typical tumor board or stroke scenario \u2014 shortens the path from image to decision. That promise depends less on the AI model and more on genuine interoperability between PACS, the medical record and the patient portal.<\/p>\n At every step, Panner insists, the radiologist must stay in the loop to protect patient safety. “I see AI as another way to make humans more powerful, which is great.” That is precisely the design the literature calls structured human oversight, a governance question that only grows as reporting volume does \u2014 and departments already watching report turnaround climb 177% over a decade<\/a> know how little slack is left for unsupervised automation.<\/p>\n There are concrete consequences for daily practice, and some are uncomfortable. The first is editorial: excessive hedging, long differential lists and incidental findings described without risk qualification produce disproportionate panic in a lay reader. This does not mean simplifying the report \u2014 it means qualifying it. Writing “finding of no clinical significance, no follow-up required” costs eight words and prevents a week of dread.<\/p>\n The second is workflow. If the patient reads first, the department needs a defined contact route for report questions, or that demand lands on the front desk phone, or worse, becomes a conversation with a chatbot that has no clinical context. The third is timing: a slow report widens the window in which the patient sits alone with the raw result. The longer the delay, the greater the chance the patient seeks interpretation somewhere else.<\/p>\n In the United States, the trigger for this shift was regulatory: the information-blocking provisions of the 21st Century Cures Act forced immediate release of results into patient portals. Other countries lack an equivalent mandate, but the vector is identical. Patients have a right of access to their records, result portals have multiplied, and the report reaches an inbox or an app the same minute it is signed. In practice, many departments outside the U.S. already operate in the American scenario without ever having discussed the editorial consequences.<\/p>\n The limits of current knowledge belong in the record too. The 44% figure comes from one specific multicampus health system and is not necessarily generalizable; studies of AI-generated summaries measure self-reported comprehension in vignette scenarios rather than clinical outcomes; and the risk of hallucination in automated report summarization remains real, especially when the model receives only text, without the images and without the history. None of this argues for slowing the change \u2014 it argues for governing it.<\/p>\n The likely path is already visible: structured reporting with a dual layer, a plain-language summary generated under supervision and validated by the radiologist, and a formal channel for patient questions. Departments that treat this as extra bureaucracy will lose the conversation. Those that treat it as part of the product will find that the report finally has two readers \u2014 and that writing for both is a skill residency does not yet teach.<\/p>\nThe Executive Who Became a Patient and Turned to AI<\/h2>\n
Who Should Deliver the News: Referrer or Radiologist?<\/h2>\n
Two Product Shifts the Market Is Already Chasing<\/h2>\n
What Changes for the Reporting Radiologist<\/h2>\n
Regulatory Context and What Comes Next<\/h2>\n