Is ChatGPT Safe for Medical Information? An Honest Answer

Your report arrived. It is full of words nobody explained to you, your appointment is next week, and there is a chatbot open in the next tab. Of course you are going to paste it in. Almost everybody does.

So rather than tell you not to, here is what actually happens when you do — including the part most articles on this subject get wrong.

We build a paid product in this space, so read this with that in mind. We have tried to write the version we would want a family member to read, which means being straight with you about what a free chatbot does well.

The short answer

A general chatbot is genuinely good at defining medical words. It is unreliable at telling you what your report means, and the reason is specific and worth understanding.

The problem is not that it is stupid. It is that it is built to produce a helpful-sounding answer, and when your report is short, ambiguous, or missing the context it wants, the most natural-sounding continuation is frequently something your report never said.

The actual failure mode: it fills gaps

Ask a chatbot to explain a three-line report and it faces a choice between a thin answer and a fluent one. Language models are built to be fluent. So it reaches for the most plausible surrounding detail — the likely cause, the typical next step, the reassuring context, the thing this "usually" means.

None of that came from your scan. It came from the statistical shape of the millions of medical documents it learned from. It will be delivered in exactly the same confident register as the parts that did come from your report, and nothing marks where one ends and the other begins.

This cuts in both directions, which is why it is dangerous rather than merely annoying:

It invents alarm. A report notes small lung nodules. Nodules co-occur with cancer in the literature far more than they do in life, where most small ones are old scarring. You can end up reading several paragraphs about metastatic disease prompted by a finding your radiologist considered unremarkable.

It invents comfort. A report contains a hedged phrase like "cannot be excluded" or "recommend further evaluation". Asked whether this is serious, a model that wants to be kind will often say the reassuring thing. Those phrases exist precisely because a human decided the question was not settled. Being talked out of a follow-up is the worst outcome on this page.

Those are illustrations of a documented failure mode, not case studies of real patients.

Model makers do not dispute this. OpenAI's own interface carries a standing warning that ChatGPT can make mistakes, and their usage policies steer away from tailored medical advice. The behaviour is well known; what is less well understood is that a short, jargon-dense document is close to the worst possible input for it, because it offers the most room to fill.

The second problem: it does not know what it is looking at

Your radiologist wrote for one reader: the doctor who ordered the scan. That reader knows why you were scanned, what your last scan showed, what you are being treated for, and what they are worried about.

A chatbot has none of that, and — importantly — it does not behave as though anything is missing. It answers with what it has. A phrase that means "expected, given the surgery six months ago" reads very differently without the six months of surgery.

What a chatbot is genuinely fine for

We are not going to pretend otherwise:

• Defining a word. What does adenopathy mean? What is the difference between a CT and an MRI? It is good at this, and so is a dictionary.
• General background about a condition, not about you.
• Preparing questions to take to your appointment.

That is real value and it is free. If that is all you need, you do not need us.

Where your data goes

Worth knowing regardless of what you use. A radiology report is about as identifying as a document gets — name, date of birth, hospital number, clinical history.

When you paste it into a consumer chatbot you are handing that to a company under consumer terms, not healthcare terms. Depending on your settings and plan, conversations may be retained, may be reviewed by humans for safety and quality, and may be used to improve future models. Some of those are adjustable in settings that most people never open.

None of this is sinister. It is simply not the handling a medical document warrants, and you are agreeing to it in a moment when you are frightened and not reading terms.

What we do differently, specifically

Here is where most articles like this claim a secret medical AI. We are not going to, because it would not be true.

Radily calls the same class of frontier model you would be talking to anyway. There is no proprietary model trained on radiology textbooks behind this. Anyone telling you otherwise is selling something.

The difference is not the model. It is what the model is forbidden to do.

Every request carries source-fidelity rules that cannot be switched off. They are appended after every other instruction, on every path that reaches a model, and they are not editable from our admin settings — deliberately, so that no future tuning of tone can quietly remove them. They say, in substance:

• The report is the only source of facts about you. Do not add to it.
• Introduce no finding, measurement, cause, diagnosis or date that is not written in the report.
• Do not convert the report's own hedged words into something more precise. "Mild" stays mild; no invented percentages or severity scores.
• Do not give prognosis, survival figures, or the likelihood that something is or is not cancer. Do not quantify risk at all.
• General knowledge may explain what a term means. It may never assert anything about your scan.
• Reassurance must be earned from your report. If the report does not support "this is common and nothing to worry about", do not say it.
• If the report is unclear or incomplete, say so. Never fill the gap with something plausible. "The report does not say" is always an acceptable answer.

That last line is the whole difference. A general assistant is rewarded for producing an answer. Ours is explicitly permitted to decline, and that permission is the thing standing between you and a confidently invented sentence.

We also do not let your report train anything. Requests carrying report text are sent with model-training disabled, and with zero-retention routing where the provider supports it. That is a property of how the request is made, not a promise in a policy document.

What we do not claim

A short section that most product pages skip, which is why we are including it.

• We do not publish an accuracy percentage. We have seen competitors quote one. We do not know how ours would be measured honestly, so we are not going to invent a number that sounds good.
• A doctor does not read your explanation before you do. It is generated and delivered to you. We flag reports that look like they need attention, and we review quality afterwards, but there is no clinician sitting between you and the output — and anyone implying otherwise about a ten-minute turnaround is describing something that is not happening.
• We cannot tell you what is wrong with you. By design. We explain the document; the diagnosis is your doctor's, and a tool that crossed that line would be a regulated medical device and a worse product.
• We cannot see your images. We read the report your radiologist wrote. If the report is wrong or incomplete, so is the explanation.

What to do with your report

1. If anything in it frightens you, contact the doctor who ordered the scan. Not a chatbot, not us. That is what the referral is for, and reports routinely look worse to the patient than to the person who ordered them.
2. Read the Impression first. It is the radiologist's actual conclusion. The Findings section above it is an exhaustive inventory written for another doctor, and a sentence there can sound severe when the radiologist has already judged it irrelevant.
3. Look up the words, with whatever tool you like. Our report phrases and glossary are free and cover the standard phrasing, including what each phrase does not mean.
4. Write your questions down before the appointment. Whatever you use to prepare them, this is the single most useful thing you can do.
5. Never let any tool, ours included, talk you out of a recommended follow-up.

The honest bottom line

If you want a word defined, a free chatbot will do it, and you should not pay anyone for that.

If you want your whole report explained without paragraphs of plausible medical content that your radiologist never wrote, that is a different job, and it needs a tool that is constrained rather than merely capable.

The question is not which is smarter. Underneath, they are much the same. The question is which one is allowed to say "the report does not say" — because the moment that matters most is the moment your report is thin and the honest answer is short.

---

Want your report explained without the invention?

Upload your CT, MRI or X-ray report — plain English, a video walkthrough, and questions to take to your doctor. Or see a worked example first.

This is an explanation of a document, not a diagnosis, and not a substitute for the doctor who ordered your scan.