Dental AI can point a clinician towards a region on a radiograph, retrieve relevant history or structure a draft note. It cannot examine a patient, resolve an ambiguous image or take responsibility for consent and treatment.
The safe pattern is clinical question → suitable evidence → AI suggestion → clinician verification → patient decision → recorded outcome. A probability score is not a diagnosis, and a colourful overlay is not a treatment plan.
The dental sources and regulatory boundaries below were checked on 31 July 2026. The General Dental Council regulates dental professionals across the UK. Other requirements are territorial: the Care Quality Commission regulates relevant activities in England; NHS England's DCB clinical-safety standards are an English NHS context; and the ionising-radiation guidance cited below applies to Great Britain, with separate Northern Ireland regulations. Medical-device routes can also differ between Great Britain and Northern Ireland, so confirm the applicable regime before procurement or deployment.
Start With a Narrow Clinical Question
Do not buy “AI dentistry”. Define one user, decision and output.
| Proposed use | AI may support | Evidence needed | Decision that stays human |
|---|---|---|---|
| Radiograph worklist | Prioritise an image for review | Sensitivity for urgent findings, false-negative analysis and workflow timing | Whether and when the patient needs clinical assessment |
| Image second reading | Mark a possible carious lesion, periapical change or bone-level feature | External validation by device, view, population and lesion definition | Diagnosis after history, examination and image review |
| Image quality | Flag positioning, exposure or motion problems | Agreement with competent reviewers and repeat-exposure impact | Whether another exposure is justified |
| Treatment support | Retrieve options, contraindications or guideline text | Source provenance, version and complete retrieval tests | Options offered, recommendation, consent and treatment |
| Record drafting | Structure dictated findings or correspondence | Omission, fabrication, attribution and confidentiality tests | Accuracy and completeness of the signed record |
| Appointment risk | Identify likely non-attendance | Calibration, subgroup impact and intervention benefit | Whether and how the practice contacts a patient |
State the intended user, patient group, excluded cases, input type, output, clinical purpose and consequence of error. If a vendor describes a tool as administrative but the practice uses it to detect disease or determine care, the actual intended purpose and deployment matter.
The MHRA's software and AI as a medical device guidance explains that software and AI products with a medical purpose may be regulated as medical devices. Its Innovation Office guidance, current at this cutoff, directs developers through classification, clinical investigation, post-market duties and support including the AI Airlock. Record the exact product, manufacturer, intended purpose, classification, approved markets and software/model version; do not rely on a sales label.
Validate the Image Path, Not Just the Model
A headline accuracy number rarely describes a dental service. Performance changes with sensor and scanner models, bitewing or periapical technique, compression, exposure, anatomy, restorations, disease prevalence and the threshold used to define a lesion.
Build a local evaluation set from consecutive eligible cases rather than a hand-picked gallery. Preserve the original image and acquisition metadata. Have competent clinicians establish the reference process in advance, including how disagreement and uncertain cases are handled. Separate patient-, tooth-, surface- and lesion-level results; they answer different questions.
At minimum report:
- sensitivity and false negatives for each clinically important target;
- specificity and false positives, including alerts per image;
- positive and negative predictive value at local prevalence;
- calibration, not only discrimination;
- results by image type, acquisition device, site, age band and relevant clinical subgroup;
- unreadable or out-of-scope rates;
- clinician-with-AI versus clinician-without-AI performance and review time;
- downstream repeats, referrals, treatments, delays and reversals.
Use patient-level splits so images from the same person do not leak across training and test sets. Add a temporal or external-site test. Keep low-quality, restored, paediatric and unusual cases visible; excluding every difficult case creates a misleading benchmark.
For dental radiography, radiation protection remains independent of the model. UK government IR(ME)R guidance covers justification, optimisation and defined practitioner/operator duties in Great Britain. It states that software assisting interpretation must not be used autonomously. AI must never turn a screening suggestion into an automatic exposure or repeat. A properly entitled practitioner justifies the exposure, and a competent operator performs the task under local procedures.
Make It a Second Reader
Show the unaltered diagnostic image beside any overlay. Let the clinician hide the overlay, change magnification and inspect the full series. Display the model version, target, confidence or uncertainty and known exclusions in language that does not imply certainty.
The interface must make disagreement easy. Record whether the clinician accepted, rejected, amended or could not assess the suggestion—and why when clinically material. Do not create an alert burden that encourages automatic acceptance or indiscriminate dismissal.
An AI mark should prompt a question: Is the feature real? Does it fit the history and examination? Is another view or test justified? Does it change the range of reasonable options? The final entry should distinguish observed clinical findings, AI output and clinician assessment.
This is the same governance principle used in our broader predictive [healthcare AI guide](/blog/healthcare-ai-predictive-analytics-remote-monitoring-uk-2026): support a defined decision while preserving a qualified professional's authority and an auditable route to override.
Keep Treatment Planning Patient-Specific
Treatment planning integrates symptoms, examination, periodontal and restorative status, radiographs, prognosis, medical and social context, patient priorities, alternatives, cost and the consequences of doing nothing. A model trained on accepted plans may reproduce historical preference, not the best choice for the current patient.
Use AI to retrieve evidence or calculate transparent inputs, not to manufacture a personalised recommendation from incomplete records. Require the clinician to confirm data completeness, contraindications, diagnosis, reasonable alternatives and uncertainties before any plan reaches the patient.
The GDC's Standards for the Dental Team centre patient interests, effective communication, valid consent, confidentiality, competence and raising concerns. Consent is a conversation, not an AI-generated signature screen. Explain material benefits, risks, alternatives, costs and uncertainty in terms the patient can understand; answer questions and allow time where needed. If AI materially influenced an assessment, the clinician should be able to explain its role and limits without implying that a machine made the decision.
Build Clinical Safety Into Deployment
For deployments within scope, NHS England's digital clinical safety assurance describes DCB0129 for manufacturers and DCB0160 for deploying health organisations. Its 2026 standards review notes that the standards remain current while being reviewed and that the statutory wording strengthened in July 2025. This is not a badge produced at the end: appoint the appropriate clinical-safety leadership, maintain a hazard log, safety case and evidence, and control changes across the product and local workflow.
Include hazards such as:
- a missed lesion because an image or patient is outside the validated domain;
- a false alert causing unnecessary repeat exposure, referral or treatment;
- a stale output attached to the wrong image or patient;
- automation bias, alert fatigue or hidden model unavailability;
- incomplete notes or invented history;
- integration failure that silently drops an image, result or override;
- a software update changing thresholds or presentation;
- loss of the service during an urgent pathway.
Give each hazard an owner, causal chain, controls, test evidence, residual-risk decision and monitoring signal. Rehearse manual operation. The practice must remain safe when the AI, image link, network or vendor service fails.
If an independent AI provider analyses images and reports results, registration questions can arise in England. CQC's diagnostic and screening procedures guidance, updated in March 2026, distinguishes supplying technology from carrying on the regulated activity. Confirm the service model rather than assuming a vendor's registration position covers the dental provider.
Preserve a Defensible Clinical Record
GDC recordkeeping guidance requires contemporaneous, complete and accurate patient records, including relevant discussions and consent. For an AI-supported case, retain or reference:
- original image and acquisition details;
- product and model/software version;
- output, timestamp, threshold and processing status;
- clinician review, interpretation and disposition;
- relevant history, examination and diagnostic reasoning;
- options discussed, patient questions, consent and decision;
- referrals, follow-up and eventual outcome;
- correction, incident or complaint where applicable.
Do not let an editable overlay become the only evidence. If a generated note is used, the signing clinician must compare it with the encounter and correct omissions, invented statements and wrong attribution. Log later amendments transparently.
Minimise data sent to vendors, verify processor and subprocessor terms, retention, security, international transfers, access control and incident response. Complete a DPIA where the processing is likely high risk. Our UK AI privacy guide provides the wider governance workflow.
Control Updates and Learn From Outcomes
Freeze the evaluated version for release. A model, threshold, image preprocessor, sensor, clinical pathway or user-interface change can alter risk. Require change notice, impact assessment, regression testing and approval before deployment. Keep a rollback package and prevent unreviewed vendor updates from changing clinical behaviour overnight.
Monitor more than clicks on “accept”. Review false negatives found later, false-positive interventions, repeats, referrals, treatment reversals, complaints, unreadable rates, subgroup performance, user overrides, downtime and incidents. Sample accepted and rejected outputs against eventual clinical evidence. Feed safety findings into the hazard log, medical-device vigilance route where applicable and local governance.
Set Measurable Release Gates
Release one defined workflow and site at a time:
| Gate | Minimum release evidence |
|---|---|
| Intended purpose | Named user, target, input, output, exclusions and error consequence match the regulated product and local use |
| Device route | Applicable GB or Northern Ireland status, manufacturer, classification, version and post-market contacts verified |
| Diagnostic evidence | Pre-agreed sensitivity, false-negative, false-positive, calibration and unreadable thresholds pass locally |
| Subgroups | No clinically material unexplained gap across relevant age, device, image-type, site and case-complexity groups |
| Radiation | AI cannot justify, request or trigger an exposure; IR(ME)R roles, procedures and repeat controls pass testing |
| Human review | Original image remains available; qualified clinician can reject, amend, document and escalate every output |
| Clinical safety | Hazard log, safety case, downtime path, incident route and named owners approved before live use |
| Records and consent | Output/version, clinician reasoning, options and patient decision are retrievable for every sampled case |
| Privacy and security | Minimisation, DPIA where required, supplier terms, access, retention, backup and incident tests pass |
| Change control | Version lock, notice, regression set, approval and tested rollback are operating |
Pause the tool if a critical feed is missing, the version changes without approval, a safety threshold is exceeded, subgroup performance drifts, results attach to the wrong record or the manual fallback fails.
Dental AI earns trust by making a competent clinician's review more consistent and traceable. It fails when an overlay becomes a diagnosis, an average accuracy number replaces local evidence or a patient is told the computer chose their treatment.
