
By LORENZO GIUNTINI
For years, I was the person leaving a medical appointment with another scan, another report, and too many unanswered questions. Chronic pain has a way of making every unfamiliar word feel urgent. I did not need a machine to replace a radiologist. I wanted someone to explain what the existing medical record was telling me, in language I could actually understand.
That desire led me to build a patient-facing imaging AI tool. Along the way, it also revealed a problem that healthcare AI has not clearly named: explaining an interpretation and creating a new interpretation are different jobs. We keep packaging them as though they were the same service.
If I upload a radiologist’s report and ask, “What does this mean?”, the task is to translate, contextualize, and clarify an existing professional interpretation without changing its meaning. If I upload MRI images and ask, “What is wrong with me?”, the task is different. The system is being asked to interpret primary medical evidence. Both experiences may look identical on a screen: a chat box, a reassuring tone, and paragraphs in plain English. But they do not deserve the same trust merely because they share an interface.
The health record became accessible before it became understandable
Patients are increasingly able to see reports without waiting for a conversation with a clinician. That is a major gain in autonomy. It is also often the moment when ordinary people discover how much of the medical record was written for another audience.
In July 2026, the Radiological Society of North America described research on AI-generated simplifications of radiology reports. The research suggested that simplified reports could improve patient comprehension and reduce anxiety in a vignette-based survey. That is an encouraging result for the report-explanation task. It is not evidence that the same system can accurately interpret an MRI or CT from image data.
The difference sounds obvious when spelled out. In practice, however, it can disappear within one conversation. A patient begins by asking what the report means. Then comes a photograph of the scan, or a request to find something the doctor may have overlooked. Suddenly the system has shifted from explaining known findings to generating new ones, often without a visible change in its claimed role.
A useful assistant ought to be able to say: “I can explain what your radiologist wrote, but analyzing the actual image is a different task with different limits.” Instead, many interfaces reward continuity. They keep answering. The language stays fluent even as the evidence becomes weaker.
What a small internal comparison taught me
In our measurements, we compared models against eight imaging examinations for which a radiologist’s report was available. No tested model identified the radiologist’s main reported finding in more than three of the eight studies. This was a small, selected internal evaluation, not a clinical trial. We did not establish diagnostic sensitivity or specificity. An existing report also provides a useful comparison rather than an infallible research-grade ground truth.
I should be plain about my stake. The tool I built does both jobs: it explains reports, and it also reads the images. The models we tested included the one my own product uses. That is exactly why I care where the line falls, and why I think products like mine owe patients a clear label.
Even with those limitations, the exercise changed the question I ask of patient AI. Not “Can it write a convincing explanation?” but “What was it actually asked to do, what evidence did it use, and how was that particular task tested?”
Imagine a patient who uploads an MRI and receives a fluent paragraph about an incidental observation while the main issue identified by the radiologist goes unmentioned. If the AI is framed as a translation of the report, that omission is a fidelity problem. If it is framed as a fresh interpretation of the images, it becomes an image-reading performance problem. The same polished output can conceal two completely different failures.
The remedy is not simply to attach the words “not medical advice” to both experiences. A disclaimer may help establish limits, but it cannot communicate which task the software performed, whether the whole examination was available, or how reliable the underlying interpretation is. Patients need the product to tell the truth about the process, not just the legal status of the answer.
A report is not an image, and a summary is not a diagnosis
Here is how I think patient-facing imaging AI should be organized.
A report-explanation mode should identify the source document, preserve the radiologist’s findings and uncertainty, define unfamiliar terms, and help formulate questions for a clinician. Its quality can be evaluated through meaning preservation, readability, and avoidance of unsupported additions. The evaluation should include patients and clinicians, not just model-based scoring.
An image-analysis mode, if offered at all, should identify the exact exam type and image material available, state the intended task, distinguish observations from confirmed diagnoses, and disclose whether its performance has been evaluated for that kind of examination. It should not borrow credibility from studies about report simplification.
A third service, care navigation, asks something different again: whom to contact, which questions to bring to an appointment, and when a concerning symptom warrants timely medical attention. It should be designed around credible escalation criteria, not around making the machine sound like a doctor.
None of these distinctions requires patients to learn the terminology of AI model development. It requires product designers to stop asking patients to infer the boundary themselves.
The danger of a seamless conversation
The marketing appeal of medical AI is that it meets people where they are. I understand that appeal intimately. When you are in pain and struggling to make sense of a technical report, the last thing you want is another opaque gatekeeper.
But the same seamlessness that makes an AI interface inviting can hide a change of responsibility. A model may begin by defining a term, then suggest a diagnosis, then answer whether surgery is necessary, with no break in its confident voice. The patient experiences one conversation. From a clinical and governance perspective, it may contain several different functions with different risks.
The FDA’s August 2026 discussion paper invites consideration of how generative AI-enabled medical devices should be evaluated and monitored. That paper does not establish new binding rules. It does underscore why defining a system’s function matters. A product cannot be judged only by how naturally it converses if its practical role changes during the interaction.
We should also avoid pretending that the clinician’s report always settles every question. Reports have limitations, interpretation can vary, and patients sometimes need second opinions. But those realities do not justify presenting an unvalidated image interpretation as interchangeable with a translation of an existing report. The correct response to uncertainty is to describe it and provide an appropriate path for professional review.
Let the patient choose, but tell them what they are choosing
The patient-facing version of informed choice should start with a simple label: “I am explaining a radiologist’s report” or “I am generating a separate interpretation from images.” Each label should open to a short account of inputs, limits, and evidence. If a system cannot credibly support the second task for the exam at hand, it should say so.
I do not want access to medical information rolled back in the name of safety. I want access to become meaningful. That means letting patients ask hard questions about their own bodies without implying that every answer carries the authority of a specialist.
Healthcare has spent years trying to make records available. The next challenge is to make their meaning clear without confusing interpretation with explanation. Patients should not have to become radiologists, AI engineers, or lawyers to understand which service they are receiving.
I started with a modest wish: to understand my own scans. A healthcare AI industry that cannot plainly distinguish helping me understand a report from telling me what an image shows has not yet answered that wish.
Lorenzo Giuntini is the Founder of Read Your Scan, an AI-powered patient imaging information service. His experience with chronic pain and repeated medical imaging motivated the project. He is not a clinician.