AI Education Technology

When ChatGPT Learned to Talk About Teeth

Dr Ali Vatan Ali Vatan
·

Large language models are being tested in dental diagnostics, education, and patient communication. Here's what they can actually do — and what they can't.

When ChatGPT Learned to Talk About Teeth

Large language models have arrived in dentistry, and everyone has an opinion. Patients ask me whether ChatGPT can diagnose their toothache. Colleagues wonder if it will replace half the profession. The reality is more mundane and more interesting than either extreme.

A very specific form of Google

The way I describe ChatGPT to colleagues is that it’s a very specific form of Google. You ask it something, it gives you an immediate, synthesised answer. It’s fast, fluent, and often impressively accurate at surface level.

But it’s disorganised. It has no memory of your patient. It can’t follow Mrs Patel’s periodontal condition over three years, notice the subtle radiographic changes, or remember that she gets anxious and needs more time in the chair. It gives you a snapshot; dentistry is a longitudinal relationship.

A landmark paper in the International Journal of Oral Science described how multi-modal LLMs could theoretically handle everything from automated diagnosis to cross-modal analysis of radiographs and clinical photographs (Xu et al., 2023). The potential is real, but potential and clinical reality are two very different things.

Where LLMs genuinely help

I use ChatGPT regularly, just not for diagnosis. Where it’s transformed my workflow is in reducing cognitive friction: the mental effort of writing things that need writing but aren’t clinically challenging.

  • Referral letters. I dictate key clinical findings; ChatGPT drafts a clean, professional letter in seconds.
  • Patient emails. Same principle, same speed.
  • Treatment notes. Complex records that need tidying for medico-legal purposes come together much faster.

A review in Frontiers in Dental Medicine explored how AI-driven language models could support differential diagnoses and patient education materials (‘Transforming dental diagnostics with artificial intelligence,’ 2024). These are practical, everyday applications.

A study in BDJ Open compared a purpose-built GPT model with standard ChatGPT for post-operative care instructions and found the fine-tuned version outperformed on accuracy and relevance (Batool et al., 2024). That tells you something important: when these tools are built for specific clinical contexts, they work better.

The part that worries me

My biggest concern is clinicians blindly accepting LLM output without critical thinking.

I’ve seen it already. A colleague showed me a ChatGPT-generated treatment plan for a complex case. Beautifully written, logically structured, and about 70% correct. The other 30% would have led to inappropriate treatment. It was presented with such confidence that you’d need real clinical knowledge to spot the errors.

How can you take responsibility for something you haven’t actually grasped? That’s the fundamental question. If you’re outsourcing your clinical reasoning to a chatbot, you’re not using a tool; you’re abdicating responsibility.

Research shows GPT-4 is the most proficient LLM at answering clinical dentistry questions, with one study finding ChatGPT o1-preview achieved 68.6% concordance with reference diagnoses in oral pathology cases (Nature Scientific Reports, 2025). That’s progress, but 68.6% is not a number I’d stake my professional registration on.

Dentistry is still about people

This gets lost in the AI conversation, and it’s the thing I care about most: dentistry is fundamentally about human interaction and trust.

When a patient sits in my chair, they’re vulnerable. They’re often anxious. They need to feel heard by another human being. No language model can replicate that.

I’m not a Luddite. LLMs will become increasingly integrated into dental practice, and many of those integrations will genuinely improve care. But they’ll improve care as tools in the hands of skilled, thoughtful clinicians, not as replacements for clinical judgement.

Where this is heading

Within five years, every practice management system will likely have some form of LLM integration. It’ll draft your notes, suggest insurance coding, and help patients understand their options through website chatbots.

That’s fine. But the clinical decisions, the ones that actually matter, will still need a human who’s done the training, seen the cases, and understands that a radiograph is only part of the story.

Use ChatGPT. I do. Just don’t let it think for you.

References