AI Can See, But a Clinician Must Decide: Putting Guardrails Around Dental AI

When I look at how quickly artificial intelligence is moving into healthcare, I’m struck by how often the conversation starts with what a tool can detect, predict or generate. Those capabilities matter. But in dentistry, the harder question is what happens next: who checks the output, who explains it and who is accountable for the decision made?

A dental image can contain useful information for an AI system to flag. That does not make the system responsible for interpreting the whole clinical situation. Nor does a polished dashboard turn an automated suggestion into a treatment decision. My position is straightforward: AI can support clinical work, but the dental surgeon must retain the authority to accept, question or reject its output.

From AI performance to clinical responsibility

The governance discussion has sharpened in recent weeks. In October 2026, the American Medical Association published a framework for agentic AI assuming clinical functions. It asks practical questions: what function is the AI performing; what evidence supports that function and what happens if it fails; what authority should it have; and who is accountable for the risks and outcomes?

Those questions become especially important as tools move beyond presenting information. There is a meaningful difference between software that highlights a feature for a clinician to review and an agent that can take steps in a care pathway. The more authority a system is given, the more carefully its scope, oversight and ability to intervene must be defined.

Performance in a retrospective test is not the same as demonstrated value in day-to-day care. A Nature Medicine commentary, “Making AI work in healthcare”, published on 14 October, argues for prospective studies and standardised evidence frameworks to bridge that gap. In other words, ask not only whether a model performed well on past data, but how it behaves in real workflows, for the intended purpose, and whether it improves care without introducing unacceptable risks.

Dental education offers another useful warning. In research commissioned by the UK General Dental Council, researchers reviewed 75 studies and reports and interviewed students, recent graduates, academics and regulator staff. The findings point to potential learning benefits, alongside a need for safeguards. Over-reliance on AI could weaken the development of independent reasoning and professional judgement. The researchers also noted that some systems may sound authoritative while being wrong, and that AI performed less reliably in image-dependent areas of dentistry.

That matters well beyond the classroom. Clinicians need to keep their own reasoning active. A system’s answer should prompt careful review, not end it.

The US Department of Veterans Affairs’ second international symposium on AI in dentistry, held in September and reported in October, focused on trust and implementation. Its discussions included integrating AI into diagnostic workflows, scaling its use and preparing clinicians and students to use it critically. That is a welcome shift: the question is not simply whether AI belongs in dentistry, but how it can be introduced and governed responsibly.

Singapore’s MOH and HSA guidance, AIHGle 2.0, reinforces the need for safe and responsible development, deployment and use. It describes AI as something that should augment and empower healthcare professionals, with patients at the centre. Explainability, reproducibility and accountability are important parts of building trust in health AI.

My house rules for clinical AI

These are practical rules I believe every dental AI deployment should meet.

1. Name a clinical owner. Give a named dental surgeon responsibility for the tool’s clinical use, its intended purpose and the way its outputs enter care. Technical teams and executives have important roles, but clinical authority must remain with appropriately qualified clinicians. A CFO or dashboard must not overrule the dental surgeon’s judgement on an individual clinical decision.

2. Make the output a second opinion, never the decision. Review AI findings alongside the clinical examination, relevant history, imaging and the patient’s circumstances. Treat an unexpected result as a reason to investigate, not as an instruction to follow automatically. And if the output does not fit the clinical picture, pause and assess it rather than forcing the situation to match the model.

3. Build a monitoring loop. Check whether the tool continues to perform as intended in the actual setting where it is used. Review errors, near misses, unexpected outputs and changes in workflow or technology. AI is not a one-off procurement decision: ongoing use requires attention.

4. Define a stopping rule before launch. Agree in advance what would trigger a pause, review or withdrawal. For example, repeated unreliable outputs, a change in the tool’s intended function or an inability to review its recommendations may justify stopping its use while the issue is assessed. Make the escalation route clear, and ensure clinicians can act on it.

5. Explain the tool to patients. Tell patients, in plain language, what the AI does and does not do, and how its output informs, rather than determines, their care. Do not imply certainty where the tool offers only a flag, estimate or suggestion. Make space for questions and explain who is responsible for the clinical decision.

These rules are consistent with a broader principle: the level of oversight should match the function and risk of the tool. An administrative assistant and a system influencing clinical decisions do not require identical controls. But each needs a clear purpose, defined responsibilities and a way to identify when things are not working as expected.

What this means for patients

Patients should not have to choose between useful technology and human judgement. A well-governed system may help a clinical team organise information or identify something that merits closer attention. It should not replace the conversation that helps a clinician understand the person behind the image.

That is particularly important in integrated healthcare. Oral health does not exist apart from a person’s broader health, priorities and circumstances. Used thoughtfully, AI can act as infrastructure: helping reduce some of the administrative or information-processing burden so clinicians have more time for judgement, explanation and conversation. That is a different aim from replacing the clinician.

If you visit a dental clinic in Singapore, it is reasonable to ask whether AI is used in your care, what its role is and who reviews the result. These are fair questions. Clear answers help people understand how information is used and where professional responsibility sits.

My view is that good governance is not a barrier to useful innovation; it is what makes responsible use possible. At Nuffield Holdings, I see technology as a means of supporting care, not a substitute for clinical accountability. Keep the human decision-maker clearly in charge, and give that person the authority to challenge, override or stop the system.

This article is general information, not personal dental or medical advice. Consult an appropriately qualified clinician about your own situation.