Can ChatGPT Read My MRI? What the Studies Show

Your images came through the patient portal before anyone called you, the appointment is next week, and last night you dropped the file into ChatGPT. It answered immediately, in full paragraphs, and sounded like it knew. The question that brings people to this page is the one that arrives an hour later: can you believe any of it.

The Short Answer

ChatGPT reads an MRI report reliably and MRI images unreliably, and it sounds exactly the same doing both. On the written words of your report it is genuinely useful, because turning medical vocabulary into plain English is the task a language model is built for. On the pictures it is not: across 230 emergency-room images, GPT-4 with vision identified the correct pathology in 35.2% of cases, and the researchers recorded a diagnostic hallucination rate above 40% (European Radiology, 2025). The problem is not that it refuses to look at your scan. It is that it never refuses, so a confident wrong answer and a confident right one reach you looking identical.

What It Is Actually Good At

Paste the text of your Findings and Impression sections and ask what a word means, and you will usually get a good answer. A 2024 evaluation published in Diagnostics put 50 straightforward MRI questions to GPT-4 and had two independent observers score the replies: one found 86% correct, the other 88%, with near-perfect agreement between them (Diagnostics, 2024).

The same study is also where the ceiling shows. On specialized multiple-choice questions the model dropped to 46.7% on gradients and 40% on site planning. Those topics are MRI physics rather than report language, so they do not describe your situation directly. What they describe is the shape of the failure: strong in the textbook middle, weak at the edges, and giving you no signal about which one you are standing in.

What Happens When You Give It the Pictures

Radiologists at the Chaim Sheba Medical Center ran GPT-4V over 230 anonymized emergency-room images collected over one week and compared its readings against senior radiologists (European Radiology, 2025).

The first two rows are why the answer you got felt credible. Recognizing that a picture is an MRI of a knee is something the model does almost perfectly, and it is also the part of the answer you can check yourself, so it reads as evidence that the rest is sound. The row that matters clinically is the third one.

That study cut in ChatGPT's favor twice over. It did not include MRI, and it used clean diagnostic images pulled from the hospital system rather than a phone photo of a monitor in a dim room. This is the generous version of the test.

MRI has a further problem of its own. The same tissue is bright on one sequence and dark on another, and the meaning of a spot changes completely depending on whether you are looking at a T1 or T2 weighted image. Radiologists read those side by side. A single screenshot usually does not carry the sequence label at all, and nothing in the reply will tell you the model guessed.

The Part That Should Worry You Most

A team at the National Library of Medicine tested GPT-4V on the New England Journal of Medicine Image Challenges and found it beat physicians on multiple-choice accuracy, 81.6% against 77.8%. Then they read its reasoning. In 35.5% of the cases where it picked the right answer, the rationale behind that answer was flawed, and the flaw was in how it understood the image 27.2% of the time (npj Digital Medicine, 2024).

It described the picture wrongly and landed on the correct answer anyway, in a third of its wins. You cannot detect that from the outside. Both versions arrive as a calm paragraph written by something that appears to know.

Nobody signs it, either. Your report exists because, as RadiologyInfo puts it, "a radiologist, a doctor trained to supervise and interpret radiology exams, will analyze the images", and that person carries the legal responsibility for what they wrote. The FDA publishes the list of AI-enabled medical devices it has authorized. General-purpose chatbots are not on it.

What a Purpose-Built Reader Does Differently

The difference that matters is a way to say no.

Between April and September 2026 ReadingScan reached a verdict on 558 patient uploads and refused 44 of them, roughly one in thirteen. Twenty-four were not medical imaging at all: a photo of a passenger asleep on a train, a veterinary cardiology report for a dog, a screenshot of a text message conversation, an insurance quote, a blank form template with a watermark on it. Most of the rest were medical but unreadable, and the largest group there was portal printouts that carried the patient header and the page furniture and none of the report. Each of those got a refusal explaining what to upload instead, and the pass went back to the wallet unspent.

A chatbot would have written a paragraph about every single one of them.

The second difference is scope. Our online MRI reader explains what your radiologist wrote, and the brain MRI page does the same for head studies down to the level of individual phrases. It does not tell you what the scan shows. That decision was made by the radiologist and is already in your report, in the Impression section, which is the part most patients skip past on their way to the scary words in Findings.

How to Check the Answer You Already Have

If you have already run your MRI through a chatbot, four things are worth doing before you lose sleep over what it said.

1. Ask it which sequence it was looking at. If it cannot name one, it was not reading the image in the way a radiologist reads it.
2. Read the Impression section of your own report and compare. The Impression is the radiologist's conclusion and it outranks anything a chatbot generated.
3. Look for findings the chatbot mentioned that appear nowhere in the report. Those are the hallucinations, and they are the ones that cost people a week of anxiety.
4. Write down the specific sentence you did not understand, and bring that sentence to the appointment rather than the chatbot transcript.

If the chatbot said something that contradicts your report, the report is what your doctor will be working from on the day. Ask them about the sentence, not about the software.

Frequently Asked Questions

Can ChatGPT read the actual images from my MRI?

It will produce an answer, and that answer is unreliable. Tested on 230 emergency-room images, GPT-4 with vision named the correct pathology 35.2% of the time and the study authors recorded a diagnostic hallucination rate above 40%. It never signals which of those two you received.

Is ChatGPT better at my MRI report than at my MRI images?

Much better. Explaining written medical vocabulary is what a language model is built for, and on straightforward MRI questions GPT-4 was scored correct 86% to 88% of the time by two independent reviewers. Reading the pictures is a different task and a much weaker one.

Is it safe to upload my MRI to ChatGPT?

A general chatbot is not part of your health system and is not covered by your hospital's privacy obligations. Radiology PDFs and portal screenshots usually carry your full name, date of birth and medical record number in the header, so crop or black those out before uploading anything anywhere.

My doctor is two weeks away. Should I just ask ChatGPT?

Use it on the words, not on the pictures, and treat the Impression section of your report as the answer of record. If a chatbot summary contradicts the Impression, the Impression is what a licensed radiologist signed and is accountable for.

Related Articles

• Brain MRI report explained: what each finding means
• How to read your MRI report
• AI vs radiologist: can AI read your X-ray?

---

Disclaimer: This article is for educational purposes only and does not constitute medical advice. Always consult with a qualified healthcare professional for diagnosis and treatment decisions.