You Asked ChatGPT About Your MRI Report. Here's What It Gets Right — and Where It Fails

You did a reasonable thing. The report was written for a physician but it's about you, no one is nearby to explain it, and the tool that answers instantly is right there. The problem isn't that you turned to AI — it's that this tool has specific limits it doesn't warn you about.

What AI genuinely does well

• It simplifies the wording. "Focal white matter changes of presumed vascular origin" comes back in language a patient can actually follow.
• It explains what a term means in general. Signal intensity, contrast enhancement, artifact, "no interval change" — these are dictionary concepts, and the answers are usually right.
• It helps you build your question list. The most useful thing you can ask for: not a diagnosis, but five questions worth raising at your appointment.
• It takes the edge off feeling helpless. With a week to wait, understanding even part of the document matters.

But there's a second half to this.

The core limit: the model reads text, not your images

A chatbot only sees the words in your report. It never sees the scan itself.

A report is a summary. A radiologist reviews hundreds — sometimes thousands — of images and condenses them into a few paragraphs. Almost none of what drove the conclusion survives into that text: how a structure actually looks, how sharply defined its margins are, whether a "change" is really just motion artifact, whether the study was of adequate quality.

The model has access to none of that. It also:

• doesn't compare your scan with your prior studies — and change over time is often the whole story: stable for three years and new in three months are very different situations;
• doesn't know your symptoms: what hurts, for how long, under what circumstances;
• doesn't know your history — past surgeries, injuries, chronic conditions, medications;
• doesn't see your lab results.

It's working with a fraction of the picture, and it does so with complete confidence.

Why this produces predictable errors

Scoring on specialized scales is unreliable

A lot of radiology is reported through standardized systems: BI-RADS in mammography, LI-RADS in the liver, and others. These aren't just labels — the category drives what happens next: watch it, repeat in six months, or act now.

This is where language models fall down. In published data, ChatGPT's accuracy in assigning LI-RADS categories was 44–53% — roughly a coin flip. For a tool people turn to for certainty, that's bad math.

The model gravitates toward the worst case

A language model answers with whatever most often follows your phrasing in the text it learned from. In medical writing, a word like "lesion" sits beside serious diagnoses far more often than beside "normal variant" — rare and severe things get written about constantly, ordinary findings barely at all. So "cannot be excluded" comes easily, and the word that scared you may have nothing to do with your case.

Common and rare sound identical

A physician thinks in frequencies: in someone your age with this picture, it's almost always the same benign thing; the frightening version shows up in a handful of cases. The model lists possibilities without weighting them. It hands you a list where the most likely and the most terrifying stand side by side, as equals.

A confident tone tells you nothing about accuracy

A language model doesn't sound less sure when it's wrong — the prose is just as calm and fluent either way. With a physician you can hear the uncertainty, and that's information a chatbot will never give you.

How to use AI safely here

This isn't "stop searching the internet." It's about using the tool for what it can actually do.

1. Ask what a word means, not what will happen to you. "What does term X mean?" is a good question. "Do I have cancer?" and "Will I need surgery?" are questions this program cannot answer — though it will answer anyway. 2. Don't make decisions based on a chatbot's reply — neither cancelling the appointment because the AI called it minor, nor starting something because it called it serious. 3. Treat the answer as a rough draft of your questions. The best outcome of that 2 a.m. conversation isn't a conclusion — it's a short list of things to ask your doctor. 4. Protect your medical data. Strip your name, date of birth, and record number out of the text. Don't upload images or documents to random websites and free "report decoders" — you have no idea where that data is stored or who can access it.

What AI can't do — and what a radiologist does

The only thing that truly settles the question is looking at the images. That's the job of a radiologist.

I open your study fresh, review every series rather than the text describing them, compare what I see against your symptoms and your prior scans if you have them, and write my own report: what is genuinely there and how urgent it is. What to do about it is a decision you make with your treating physician.

A second opinion isn't only for anxiety. In published studies, re-reading neuro MRI produces a discrepancy with the original report in roughly 30% of cases, and 12–13% of those are significant — they change management. That isn't about catching anyone out; it's about how demanding the work is and what an independent set of eyes is worth.

A second read is worth considering if:

• there's a word in your report you don't understand and your appointment is weeks away;
• you're preparing for surgery or for serious treatment;
• two doctors have told you different things;
• the study was a follow-up and you need to know whether anything actually changed;
• you need a report in English for a physician abroad.