How Does AI Work in Dentistry?

In dentistry, “artificial intelligence” usually refers to machine learning: computer programs that learn a task from examples instead of following hand-written rules. The most common dental use is computer vision — software that looks at images such as X-rays (dental radiographs) or intraoral photos and points out areas that resemble patterns it has learned.

FDI World Dental Federation gives typical examples: detecting objects on images, such as a possible caries lesion on a radiograph, or predicting events from clinical data. AI is also used for scheduling, note-taking and chat assistants, but this guide focuses on images, because that is where most dental AI works today.

Before a model can analyse anything, it is trained. Developers collect a large set of dental images and ask experienced clinicians to label (annotate) them — outlining each tooth, marking existing fillings and crowns, or drawing around areas they consider suspicious.

The model repeatedly compares its own guesses with those labels and adjusts itself to reduce the difference. Over many thousands of examples it learns visual features that tend to go with each label. Two consequences matter for patients:

• The labels set the ceiling. If the experts who labelled the images disagreed or made mistakes, the model learns those too.
• The data sets the scope. A model trained mostly on one type of X-ray, one kind of machine or one population may be less reliable on images that look different. FDI names this risk of bias and limited generalisability explicitly.

A trained model is then evaluated on a separate set of images it did not learn from, to estimate how often it finds what experts found and how often it raises false alarms. Researchers have published reporting checklists for such studies, and the American Dental Association now publishes standards on how dental image-analysis AI should be validated and how its datasets should be annotated.

Published results vary widely between studies, tasks and image types. That is one reason this Library does not quote accuracy figures: a number measured in one study, on one dataset, does not automatically apply to your X-ray.

When a new image is uploaded, the trained model runs on it. This step is called inference. Typically the software:

1. checks that the image is the kind it was built for — for example a dental X-ray rather than an unrelated photo;
2. finds the teeth and may number them using the international FDI two-digit system (11 to 48 for adult teeth);
3. marks areas that resemble learned patterns: existing dental work, or possible findings that may deserve a closer look;
4. attaches an internal score to each mark that reflects how strongly the image matched what the model learned. That score is not a measure of severity, and it is not a diagnosis.

The result is a set of AI-supported observations: structured suggestions, often shown as coloured outlines (overlays) on the image. Our guide to AI-assisted X-ray analysis looks at those outputs in detail.

An X-ray is a two-dimensional shadow image. It does not show whether a tooth hurts, whether your gums bleed, how you bite or what you have already tried. Overlapping structures, blur and positioning errors can hide a problem or imitate one. AI only works with the pixels it receives, so it inherits all of these limits.

That is why AI can produce false positives (marking something that turns out to be normal) and false negatives (missing something that is there). Neither the model nor a single image replaces your clinical history, an examination and professional judgement.

FDI's 2024 policy statement recommends that dental professionals use AI applications as assistance systems and protect themselves against automation bias — the tendency to over-trust a machine's suggestion. It states that responsibility for any diagnosis or treatment derived from AI assistance remains with the human user. The World Health Organization's guidance on AI for health likewise puts human oversight and accountability at the centre.

In practice a dentist reviews each suggestion, accepts or rejects it, adds what software cannot know, and decides whether more information is needed: another X-ray, a clinical test or a referral. We explore this further in Will AI replace the dentist?