UK Beauty AI: Safe Personalisation Without Fake Diagnosis

AI [Beauty](/industries/beauty) and [Skincare](/industries/beauty) in the UK: Safe Personalisation Without Fake Diagnosis

A phone camera can help a customer compare lipstick colours, record a visible change over time or navigate a catalogue. It cannot directly measure hydration, prove a nutritional deficiency or diagnose a suspicious lesion simply because a model returns a precise score. Lighting, camera processing, makeup, skin tone, focus and the model’s intended purpose all affect the output.

The opportunity is personalisation with honest boundaries: separate cosmetic observation from medical assessment, connect recommendations to a controlled product record, label visual simulations and preserve a route to professional care.

This guide is current to 31 July 2026. Cosmetics and medical-device routes differ between Great Britain and Northern Ireland, and advertising, data-protection and consumer rules depend on the feature and claim. Confirm the intended purpose, market and current regulator guidance before launch; this is not medical or legal advice.

Classify the feature before building it

“AI skincare” covers services with very different consequences:

Write an intended-purpose statement naming the user, input, output, setting, exclusions and action. The MHRA’s current software and AI medical-device guidance explains that many software products serving a clinical need are regulated. Its borderline medical-device guidance, updated in June 2026, also makes clear that classification turns on purpose and claims, not the developer’s preferred label.

Calling an output “wellness” does not neutralise an implied promise to diagnose, prevent or treat disease. If the app claims to assess acne severity, identify melanoma or direct treatment, obtain specialist regulatory and clinical advice before testing it on customers. If it is genuinely cosmetic, avoid disease terminology and design the escalation route so that the model cannot delay care.

Capture an image that can support the claim

Consumer images vary before the model sees a pixel. Phones apply exposure, sharpening, colour balance, portrait effects and compression differently. Bathroom LEDs, sunlight, screen reflections, foundation and recent washing all change appearance.

Define a capture protocol:

• list supported devices, resolutions and operating-system versions;
• ask the user to remove filters and disclose makeup where relevant;
• provide distance, angle and neutral-light guidance;
• reject blur, occlusion, overexposure and colour casts;
• preserve the original input and capture metadata only for the approved period;
• show uncertainty and an “unable to assess” result; and
• never infer what the image cannot validly establish.

Build the evaluation set from the actual UK deployment population and intended setting. Report performance by device family, lighting condition, age band and a responsibly selected range of skin tones. Include acne, scarring, tattoos, facial hair, head coverings and assistive equipment when relevant. Do not “fix” unequal performance by silently excluding a group.

Measure the end task, not only image classification. If a recommendation depends on detecting visible dryness, test whether the full system selects an eligible product and whether users can correct the observation. A high area-under-curve statistic does not show that a threshold is safe or that the recommendation is useful.

Our technical guide to AI beauty try-on and skincare explores model evaluation in more detail. The governance requirements below remain necessary even when the computer-vision benchmark is strong.

Keep product facts outside the model

A generative model should not recall ingredients, allergens, warnings or stock status from training. Recommendations must query a versioned product master containing:

• full ingredient list and formula version;
• responsible market and Responsible Person;
• safety-assessment and Product Information File status;
• directions, precautions and known incompatibilities;
• intended user and any age or application restrictions;
• batch, shelf life and notification status;
• evidence approved for each marketing claim; and
• recall, adverse-event or suspension status.

The Office for Product Safety and Standards’ Great Britain cosmetics guidance requires a Responsible Person, product safety assessment, Product Information File, good manufacturing practice, labelling, notification and evidence for claims. Products placed on the GB market are notified through the Submit Cosmetic Product Notifications service. Northern Ireland uses a different route and Responsible Person location, so a single “UK compliant” flag is not enough.

Recommendation rules should be deterministic for hard exclusions. If the user states an allergy, pregnancy, active treatment, broken skin or a previous reaction, do not ask a language model to improvise suitability. Show the ingredient evidence, advise checking the label because formulations change and route medical questions to a pharmacist, prescriber or other appropriate professional.

“Custom blended” still means a regulated finished product

Personalisation does not exempt the bottle made at the counter or by a fulfilment system. Define a finite, safety-assessed formulation space: permitted bases, ingredient combinations, concentrations, mixing tolerances, preservation, packaging and shelf life. Block every combination outside it.

Each dispensed unit needs a reproducible formula, ingredient lot, equipment state, date, operator or machine, quality result, label and customer order. Re-running a model later must not be the only way to discover what was supplied. If the AI recommendation changes, that does not retrospectively change the safety assessment for an old batch.

Before release, challenge edge cases with the qualified cosmetic safety assessor and Responsible Person. A preference score is not toxicological evidence. Do not expand the blend space through an unreviewed model update.

Make virtual try-on an honest simulation

Virtual try-on should help exploration, not manufacture proof of efficacy. State that the image is simulated and that display, device, ambient light, undertone, product finish and application alter the physical result. Offer real swatches or returns information where possible.

Calibrate against measured product colour and representative devices. Test lip, eye and hair overlays for placement across face shapes, skin tones, facial hair, mobility limitations and occlusion. Let customers disable face tracking, manually adjust landmarks and browse without uploading a photograph.

Do not reuse a try-on render as a “before and after” efficacy image. CAP’s current advice on before-and-after photographs says marketers need evidence that images are genuine, representative and not manipulated in a way that exaggerates the effect. The wider ASA beauty guidance requires substantiation for objective claims. A small disclaimer cannot cure an image whose overall impression is false.

Optimisation also needs commercial transparency. If paid placement, margin or stock pressure affects ranking, disclose it and keep the cosmetic-fit explanation separate. Do not exploit inferred insecurity with countdowns, repeated flaw labels or progressively harsher scores.

Protect face and health-related data

A face photograph is personal data when the person is identifiable. It is not automatically special-category biometric data: the ICO’s biometric concepts guidance explains that specific technical processing and unique-identification purpose matter. However, skin images and outputs may reveal or intentionally infer health or racial or ethnic information, which can bring special-category rules into scope.

Document purpose, Article 6 lawful basis and any Article 9 condition before collection. Complete a data-protection impact assessment when the risk requires it. Give a genuine non-camera route. Explain whether processing happens on-device or in the cloud, which vendor receives the image, where it is stored, whether it trains a model and when originals, derived landmarks, embeddings and outputs are deleted.

Use separate, explicit choices for:

• analysing the current image;
• saving a history for the customer;
• personalised marketing;
• identity verification; and
• research or model improvement.

Refusing model training must not block a purchased service when training is unnecessary. Do not pull old social images into a skin profile because a contract permits access. For the broader control framework, see AI and UK data-privacy compliance.

Restrict staff access, encrypt transfers and storage, remove images from routine logs and test deletion across vendors and backups. Treat model and catalogue files as supply-chain assets: record versions, scan uploads, prevent image metadata or retrieved copy from becoming instructions, and rehearse breach response.

Escalate symptoms and adverse effects

The assistant must not reassure a customer that a lesion is harmless. Provide a clear route to NHS or professional care when the user reports a new or changing mole, bleeding, persistent symptoms, severe pain, swelling, breathing difficulty, eye exposure or a serious reaction. NHS guidance describes specialist assessment following referral for suspected melanoma; it is not replaced by a retail score.

Maintain an adverse-effect workflow linked to the actual product and batch. The OPSS cosmetics guidance requires Responsible Persons and distributors to report serious undesirable effects and describes the required information. Frontline staff need a script that prioritises urgent care, records only necessary facts, preserves the product and batch trail and alerts the Responsible Person. Do not let a sentiment classifier downgrade a report.

A measurable 90-day pilot

Days 1–30: freeze intended purpose and claims; classify the feature and markets; appoint product, clinical, privacy and security owners; build the controlled catalogue; define capture protocol, evaluation slices, escalation rules and non-AI comparator.

Days 31–60: test offline with consented representative images. Challenge lighting, devices, cosmetics, occlusion, skin tones, prohibited claims, ingredient conflicts and vendor deletion. Conduct accessibility and security testing and have the safety assessor review every formulation path.

Days 61–90: release one low-risk feature to an opt-in group. Sample outputs daily, reconcile recommendations to product records, review adverse effects weekly and rehearse shutdown, recall and deletion.

Release only when:

• 100% of outputs remain within the approved cosmetic or regulated intended purpose;
• every product recommendation traces to a current ingredient, warning, market and claim record;
• zero prohibited ingredient, suspended product or non-assessed formula appears in the test suite;
• the defined image-quality rejection catches all deliberately unusable captures;
• accuracy and “unable to assess” thresholds pass for every agreed device, lighting and skin-tone slice;
• 100% of red-flag symptom and serious-reaction cases reach the correct non-AI escalation;
• every simulation is labelled and no efficacy evidence uses altered try-on imagery;
• consent withdrawal and deletion meet the promised service level across all processors;
• no unresolved critical product-safety, medical-device, privacy, security or accessibility issue remains; and
• rollback removes the model feature without removing product warnings, support or adverse-event reporting.

Pause after a harmful recommendation, missed urgent symptom, wrong formula, untraceable batch, fabricated claim, material group disparity, unauthorised reuse of images or security compromise. Revalidate after any model, camera pipeline, formula, claim, market, vendor or regulatory change.

The practical verdict

AI can make beauty retail easier to explore, but it should not convert a camera artefact into a diagnosis or a preference into a safety conclusion. The reliable design keeps image uncertainty visible, product facts deterministic, formulations traceable and medical escalation outside the sales funnel.

Personalisation earns trust when the user can understand, correct and delete it—and when the system knows that some questions belong with a qualified professional, not a smoother-generated answer.