Can ChatGPT Read Your Dental X-Ray?

Partly, and less well than it sounds when it answers you. ChatGPT, Claude and Gemini all accept image uploads. Give any of them a photo of a dental X-ray and it will describe what it sees in confident, fluent English. Sometimes that description is accurate. Often it is vague, and occasionally it is flatly wrong in ways a non-dentist has no way to spot. What decides whether it helps you is the question you ask it.

This page covers that one question in detail: what happens when you upload a dental radiograph to a general AI chatbot, where it goes wrong, and how to use it so you come out with something useful when deciding about a treatment plan. If you want the wider picture (checking quotes, understanding NHS bands, choosing between treatments) the pillar guide on asking AI about your teeth covers it.

What actually happens when you upload an X-ray

General chatbots were not trained to read radiographs. They learned from a huge mix of images and text, and some of that included dental X-rays with captions, textbook diagrams and forum posts. So they recognise the general look of a radiograph: bright white for metal fillings and crowns, light grey for enamel, darker grey for softer tissue, and black where there is nothing dense at all.

The trouble is that dental diagnosis depends on small differences in grey. Early decay between two teeth can be a thin dark wedge only a millimetre or two across. An infection at the tip of a root shows up as a faint dark halo. If your upload has been compressed, photographed off a monitor, or shrunk by the AI tool itself, those details may simply not be there to see.

The shrinking matters more than most people realise. At the time of writing, OpenAI’s documentation says that in high-detail mode its vision models scale an image to fit inside 2048 × 2048 pixels, then scale it down again so the shortest side is 768 pixels. A full panoramic X-ray (an “OPG”, the wide one showing your whole jaw) can be 2,900 pixels wide or more in the original. Fit that into a 768-pixel-high frame and each tooth is perhaps 60 to 80 pixels across. You would struggle to diagnose anything at that size yourself.

The four ways it typically goes wrong

Left and right get swapped. Dental X-rays are displayed as if the dentist is facing you, so your right side appears on the left of the image. Your treatment plan will probably use Palmer notation: “UR6” means upper right, sixth tooth from the middle (your first molar). Chatbots mix this up regularly. They describe a crown on “the upper left” when the image shows it on the left but the tooth is actually in your upper right.

Tooth counting drifts. If you have had a tooth removed, or a wisdom tooth is still buried, the AI may count along and land one tooth off. Ask it about your “LL6” and it might be describing your LL7.

It hedges until the answer means nothing. A typical response contains phrases like “there may be areas of radiolucency that could indicate decay or could be normal anatomy”. That is true of almost every dental X-ray ever taken. It feels like analysis but it gives you nothing you can act on.

It invents findings. This is the dangerous one. Push a chatbot with a leading question (“Can you see the abscess on my lower right molar?”) and it will often find what you suggested. These tools are built to be helpful, and agreeing with you feels helpful.

A worked example

Suppose your private dentist in Leeds has given you this plan:

You asked for a copy of your X-rays and the practice emailed you two bitewings and a periapical (a close-up of one or two teeth, showing the full root). You upload the periapical of UR6 to ChatGPT and type: “What’s wrong with this tooth? My dentist says I need a root canal.”

Here is a representative exchange, condensed:

ChatGPT: This appears to be a periapical radiograph of an upper molar.
I can see a large radiopaque restoration (likely an amalgam or composite
filling) occupying much of the crown. There may be a radiolucent area
near the apex of one of the roots, which could be consistent with
periapical pathology such as an abscess or granuloma. However, I can't
make a diagnosis, and you should follow your dentist's advice.

Look at what that answer contains. “Large filling” is probably correct and easy to see. “May be a radiolucent area” is a guess dressed up in clinical words. And the last line tells you to trust the person whose advice you were trying to check.

Now try a different prompt with the same image:

I'm a patient in the UK, not a dentist. My dentist has recommended root
canal treatment and a crown on UR6 (upper right first molar). Please:
1. Tell me what features a dentist would look for on an X-ray to justify
   root canal treatment.
2. Describe what you can and cannot reliably see in this image, and why.
3. Give me five specific questions to ask my dentist at the next appointment.
Do not guess at a diagnosis.

That gets you something useful. A typical reply explains that dentists look for widening of the ligament space around the root tip, a dark area at the apex, decay reaching the nerve chamber, or a previous root filling that looks short or patchy. It admits that image quality limits what it can confirm. And it suggests questions such as:

- Which root do you think is infected, and can you point to it on the image?
- Did the tooth respond to a cold test or an electric pulp test?
- What are the chances of success, and what happens if it fails?
- Why a full crown rather than an onlay?
- What would extraction and an implant or bridge cost by comparison?

You walk in with better questions, and you have not treated a chatbot’s guess as a diagnosis.

What an X-ray can show, and what it can’t

Even a perfect reading of the image would not settle the question, because dentists don’t diagnose from X-rays alone. A lot of the decision rests on things that never show up on film.

So the most honest answer an AI can give you is half of the story at best. When your dentist recommends a root canal, they are usually weighing the image against your pain history and how the tooth responded when they tested it.

Getting a usable image in the first place

Under UK GDPR you have a right to a copy of your dental records, X-rays included. Make a subject access request in writing (an email is fine). The practice must normally respond within one month and cannot usually charge you. Ask for the images “in their original digital format”, which generally means a full-resolution JPEG, PNG or DICOM file, not a printout.

If you only have a photo of the dentist’s screen, the AI is working from a copy of a copy. Screen glare, moiré banding and auto-sharpening from your phone camera all add patterns that can look like decay. That’s a quick way to get a frightening and meaningless answer.

Before uploading, crop out your name, date of birth and the practice details printed in the corner. In ChatGPT you can switch off “Improve the model for everyone” under Settings > Data Controls, or use a Temporary Chat. Claude and Gemini have similar settings. Treat anything you upload as something that could be stored.

Specialist dental AI is a different thing

You may read about AI that detects decay “better than dentists”. That usually refers to purpose-built tools such as Pearl’s Second Opinion or Overjet. Both are cleared by the US FDA, were trained on large sets of labelled dental radiographs, and are designed to sit inside a dentist’s own software, not a consumer chat window. Some UK practices now use systems like these.

General chatbots are not equivalent. If your practice uses one of these tools, ask whether they can show you its annotated output for your X-rays. That’s a much better machine opinion than anything you’ll get by pasting a JPEG into ChatGPT.

Where AI genuinely earns its keep with your X-ray

Use it for the parts it is good at: explaining words and producing questions. Paste in the written report or the treatment plan text alongside the image and ask it to translate “periapical radiolucency associated with UR6, widened PDL space, previous MOD amalgam” into plain English. That works well, because it’s a language task.

Price checking is another good use. In England, all of the work in the example above would fall under NHS Band 3, which is £326.70 at the 2025/26 rates, if your dentist offered it on the NHS. Private fees vary a lot by region: a molar root canal in the UK commonly costs anywhere from about £400 to over £1,000, and specialist endodontists charge towards the top of that. An AI can lay out these ranges and help you work out whether £1,505 is in line with the market. It can’t tell you whether this particular tooth needs the work.

A second opinion is the real check on the diagnosis. Bring the digital files to another dentist so they don’t need to take fresh images. UK radiation regulations (IR(ME)R 2017) require every exposure to be justified, and many dentists won’t re-X-ray without good reason anyway. A private second-opinion consultation typically costs £50 to £120. On a £1,500 plan, that’s money well spent. The radiation from a few bitewings is tiny (around 0.005 mSv each, compared with roughly 2.7 mSv a year from natural background radiation in the UK), but there’s still no point repeating images you already have.

When you do take ChatGPT’s view to a dentist, frame it as a question, not a verdict. “The AI mentioned a dark area at the root tip, is that what you’re seeing?” will get you a proper explanation. “ChatGPT says I don’t need this” usually just ends the conversation.