Artificial intelligence for cosmetics: real gains, real guardrails

Preamble

As many of us I’ve tried Artificial Intelligence for some tasks. I’m working on a specific platform (with the Italian start-up Cosmeply *) that can be used for the safety assessment of cosmetic products. What I’m learning from this experience is that AI without humans cannot yet perform alone the fine tuning of a reasoning that derives from the experience and a 360 degrees vision.We have examples of the AI use for many tasks, but many of the miracle-like applications are the result of a long development and based on many factors, including big data. The application of AI in the scientific field simply with a question (the prompt) should be carefully evaluated before accepting the results without any check.

In this case I’ve tried to obtain a scientific article on the topic of the use of AI in the cosmetic sector as part of our research during the development of the platform.

The AI (chatGPT5) answered this prompt:

“Act like a senior cosmetic scientist (cosmetologist) with 15+ years in R&D and regulatory affairs, deeply familiar with AI/ML methods (e.g., CNNs, transformers, recommendation systems, Bayesian optimization) and an expert science communicator. OBJECTIVE Write a two-page, high-quality scientific magazine article in English about artificial intelligence in the cosmetics industry: current and near-term applications, concrete positive examples, and a rigorous analysis of risks, limitations, and regulatory/ethical issues. The tone must be technically precise yet human and readable.”

(PLEASE Note that this is only the first part of 2 pages length of the full prompt).

The result is quite interesting, the text wasn’t fully satisfying, and it has been human reviewed to have more fluid and linear text. It must be noted the number of references and insights gained with this system, compared to a manual search. A key point is also the prompt, a different approach to the question will surely have a different result.

The conclusion is that with AI we can count on a strong assistant, but we still must rely on humans.

The goal given to the AI was to summarize in short paragraphs each point of the topics examined.

Introduction

AI/ML in cosmetology stands for statistical and algorithmic tools that uncover patterns in heterogeneous data relevant to efficacy and safety. Typical modalities encompass hi-res facial photos and videos for texture, wrinkle, pore and redness assessment; spectral signatures (NIR, Raman, FTIR) of ingredients and intermediate products; microbiome amplicon sequencing and metabolomics linked to skin conditions; sensory and instrumental assessments of rheology and tribology; and text collections such as claims dossiers, adverse event reports and consumer reviews.Typical models range from transformers for images and text, to chemometric regressors and novelty detectors for data, to surrogates and Bayesian optimizers for formulation. The journey matters as much as the tools: thoughtful curation and annotation; sound train/validation/test splits with external test sets; reporting (model cards, rationales); deployment under human supervision; and monitoring for drift and adverse signals.

Core applications in practice

Formulation design and optimizationFormulators are increasingly combining designed experiments with surrogate modeling and Bayesian or active learning to achieve performance or sensory objectives with fewer iterations at the bench.

Public datasets now exist for surfactant–polymer systems in shampoos, enabling open benchmarking and model transfer; these have been used with multi-objective Bayesian optimization to suggest formulations that minimize phase separation and fall within a viscosity window.

The method reduces the number of experiments while still maintaining a high output level, this so long as the model is tested in real on external batches and its accuracy reported in the system and calibration plots.

To be noted: in manufacturing, models are best suited as decision aids rather than fully automated formulators: their maturity level today is pilot to early manufacturing where data provenance, INCI coverage and instrument traceability are solid.

Safety assessment and toxicology

Quantitative structure-activity and structure-property relationships and read-across remain indispensable for safety reasons under the EU Cosmetics Regulation. The SCCS Notes of Guidance, 12th revision, highlight the integration of New Approach Methodologies, including defined approaches for eye and skin irritation and new in vitro genotoxicity assays, as well as systematic literature review (1, 2).

In this context, QSAR models must reveal their domains of applicability, uncertainty, and confidence; good practice is in line with the OECD QSAR principles of validation, including a defined endpoint, unambiguous algorithm, an adequate domain, good quality metrics, and a (mechanistic) interpretation, if feasible. Some start-ups already struggle to substitute safety assessors with the limitation from the regulation that requires as mandatory a human safety assessor (17, 1).

The state of the art is production under supervision, with NAMs and QSARs incorporated into the Cosmetic Product Safety Report when supported by transparent uncertainty disclosure and human expert review.

Dermatological imaging and diagnostics-adjacent uses

Deep learning has improved segmentation and quantification of cosmetic-concern skin microfeatures, wrinkle length and width, pore density, and erythema color intensity.

Clinical validation remains an unavoidable step: erythema quantification algorithms such as erythema Q-score illustrate the potential of image-based measures to correlate with clinical grading in controlled trials, but generalization across devices and skin types must be confirmed on external data sets.

These devices are best explained as claim measurement aids to personalization and not as diagnoses. There are settled deployments in controlled imaging booths and steady smartphone pipelines with calibration, and performance is reported via overlap and repeatability measures.

The road is paved by the image analysis in radiology where the AI is integrated in the clinical path of the diagnosis.

Quality control and manufacturing

Chemometrics and computer vision provide process analytical technology and quality by design. Camera-based equipment identifies packaging defects at line rate, while inline Raman, NIR or FTIR with multivariate models verifies concentration and identity of actives.

The FDA Process analytical technology (PAT) framework remains a model for building quality into processes with in-process measurement and feedback control on time (9).

In industry case studies, convolutional detectors identify defects on caps and closures with high accuracy and reduce the workload of manual inspection. Maturity is production where sensors are qualified, and models are under change control with re-validation at regular intervals (14).

Personalization and recommendation

Recommendation systems derive profiles of users’ skin alongside their surroundings and daily activities to create customized routines and match skin tones to cosmetics. This marketing-oriented area is gaining popularity and has great performance. Undoubtedly, bias is a critical issue, considering the absence of representative data across Fitzpatrick skin types, devices, and systems that perform poorly for darker skin tones (12).

AI studies of dermatology have logged performance gaps that inform cosmetics; audits that capture the error of tone phototype coverage, mitigating harm by cross-stressing capture devices, help. Privacy is also an issue, and we prefer on-device inference and edge aggregation where possible, GDPR Article 9 restrictions on biometric data also apply (12, 5).

Consumer insights and claims substantiation

Natural language processing can assist with issue tracking for product reviews and emerging evidence claims and can augment clinical and instrumental assessments. In the case of regulated promotion, the analysis is to be framed within the EU common criteria and technical documents for cosmetics claims, and the US FTC Endorsement Guides (8, 11).

The outputs of the models should not form the basis of claims of superiority; instead, they assist in hypothesis formulation, consumer language simplification, and the claims should be backed by thoroughly thought-out experiments and structured evidence.

Benefits, with caveats

Across these areas, properly implemented AI reduces the number of bench iterations. It shortens feedback loops from the plant to the lab. It also widens safety margins by detecting signals earlier and can improve inclusivity when datasets are balanced and coverage is reported. Reproducibility improves when data lineage, code, and configuration are versioned, and when we re-check model performance on external sets during change control. However, none of these advantages happen without the less glamorous work of curation, documentation, and monitoring.

Risks, limitations and how to mitigate them

In terms of risks about bias and coverage gaps, risk persists with the underrepresentation of darker skin tones, and vision models can be particularly susceptible to shifts in device and illumination. Mitigating strategies include pre-specified step size or targets by phototype for sampling, stratifying performance reporting, device robust augmentation, and periodic external testing at other sites (for personalization engines, using fairness constraints and error auditing is becoming standard).Overfitting, domain shift, and leakage sensitivity is managed with disciplined splits, locked external test sets, and documentation of design dependence in model cards and dataset datasheets so that downstream users can understand scope of data use and limitations of use. For safety decisions, use of explainability caveat – mechanistic plausibility, applicability domain plots, and uncertainty quantification is more actionable than generic saliency maps.

Privacy and biometrics require additional caution. Faces and/or skin imagery can be considered biometric data and processing biometric data triggers stringent criteria under GDPR Article 9, which includes clearly indicating the purpose, minimizing, and demonstrating security for that process.

Using device inference and ephemeral processing is a better default. Security is not negotiable – there are real prompt injection or model-inversion attacks, which teams establishing LLM-based tools for drafting a protocol or summarizing claims should plan for rules of engagement (e.g., allow-lists, input and output filtering, isolated tools) and governance structure (e.g., track and document prompts and decisions).

Regulatory and compliance snapshot

In the European Union, the Cosmetics Regulation (EC) No 1223/2009 requires a Cosmetic Product Safety Report and a designated Responsible Person who ensures compliance and post-market vigilance. AI tools used in safety assessments and quality control do not change these obligations; rather, they heighten the need for traceable datasets, method validation and transparent uncertainty communication in the product information file (1).The SCCS Notes of Guidance, 12th revision with 2023 corrigenda, set expectations for New Approach Methodologies, weight of evidence and systematic review practices relevant to AI-assisted QSAR and read-across (2). The EU Artificial Intelligence Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024 with phased obligations through 2025 to 2027; the status was checked on 17 September 2025. Cosmetic companies should document whether deployed AI falls into risk-managed categories and align governance and post-market monitoring accordingly. ISO/IEC 23894 and the NIST AI Risk Management Framework provide practical guidance for risk-based AI management, including data quality, robustness, transparency and monitoring. (3, 4, 6, 7).

Outlook and near-term research agenda

Priorities for the near future are quantifiable and realistic. In conjunction with calibration targets and tone-aware performance reporting, privacy-preserving, on-device inference for shade matching and skin-feature quantification can improve robustness to device variability while lowering latency and biometric risk.Predicting batch deviations and initiating corrective actions before out-of-specification material is produced is possible with digital twins of mixing and filling lines that combine process sensors, chemometrics, and physics-informed machine learning; however, success relies on rigorous sensor qualification and MLOps to maintain models in sync with equipment changes. Shared benchmarks for emulsion stability and sensorial prediction would speed up the field, and formulation platforms that integrate public benchmark datasets, mechanistically informed features, and active learning can reduce data requirements without sacrificing interpretability (9, 15, 16).

The use of large language models as helpers for protocol drafting and literature synthesis in claims and safety dossiers is a carefully regulated area; to pass regulatory scrutiny, these models must function with rigorous retrieval grounding, timely governance, and unchangeable decision logs (6, 7).

Glossary of terms

INCI: International Nomenclature of Cosmetic Ingredients. QSAR/QSPR: predictive models that relate chemical structure or properties to biological activity or physicochemical endpoints. PAT: process analytical technology for in-process measurement and control. NAMs: New Approach Methodologies that replace or reduce animal tests. PIF: Product Information File required in the EU. Responsible Person: the EU-designated natural or legal person who ensures compliance. GDPR Article 9: special categories of personal data, including biometrics. Model card and Dataset datasheet: documentation artefacts summarizing model scope, performance, data provenance and limitations. Applicability domain: the chemical or data space within which a QSAR is considered reliable.​​​​​​