Artificial Intelligence in Cosmetic Product Development(1): Opportunities, Limitations, and Why Human Scientific Judgement Still Matters

Artificial intelligence is now part of how cosmetic products get developed. AI-powered tools can screen ingredient databases, predict formulation properties, search scientific literature and flag regulatory updates within minutes, work that used to take a regulatory affairs team days to do by hand. For formulators, that means faster iteration and fewer wasted development cycles. But AI works best as a scientific support tool, not a replacement for scientific judgement.

Several applications already show real value. Predictive models can estimate viscosity, emulsion stability, texture and shelf-life from existing formulation data, narrowing the field of candidates before physical testing even starts. AI can also compare thousands of raw materials at once and suggest compatible combinations, widening the pool of options a formulator can realistically consider. Natural language tools track scientific literature and regulatory databases across markets far faster than manual review.

The limits show up hardest once predictions start touching safety decisions. A model is only as good as the data it was trained on, and cosmetic datasets often underrepresent certain skin phototypes, age groups and newer ingredient classes, so a confident-looking prediction does not always hold across a full consumer base.

Explainability is the other real gap. Say a model predicts that an ingredient carries low toxicological concern. A safety assessor will immediately want to know what evidence supports that conclusion, what exposure assumptions were used, whether the prediction sits inside the model’s applicability domain, and how much uncertainty is attached to it. If those questions cannot be answered, the prediction is hard to defend scientifically.

This is not a vague concern, it is written into how these models are formally validated. Under the OECD principles for (Q)SAR validation, a model intended for regulatory use needs a defined endpoint, an unambiguous algorithm, a defined applicability domain, measured predictive performance and, where possible, a mechanistic interpretation. A prediction that falls outside that domain does not meet the bar for a regulatory conclusion, however polished it looks.

Cosmetic safety assessment pulls together exposure data, toxicological endpoints, ingredient specifications, literature and regulatory requirements, then weighs all of it through a weight-of-evidence approach. That weighing is a judgement call, particularly when evidence is incomplete or involves a novel ingredient with little precedent. AI can generate the inputs. It does not make that call.

Part 2 looks at why AI-generated outputs cannot replace the weight-of-evidence approach required for Cosmetic Product Safety Reports, and why expert toxicological judgement stays central to safety assessment.

Sources:

Scientific Committee on Consumer Safety (SCCS), Notes of Guidance for the Testing of Cosmetic Ingredients and Their Safety Evaluation, 12th Revision, European Commission: https://health.ec.europa.eu/publications/sccs-notes-guidance-testing-cosmetic-ingredients-and-their-safety-evaluation-12th-revision_en

OECD, Guidance Document on the Validation of (Quantitative) Structure-Activity Relationship [(Q)SAR] Models, OECD Series on Testing and Assessment No. 69: https://www.oecd.org/en/publications/guidance-document-on-the-validation-of-quantitative-structure-activity-relationship-q-sar-models_9789264085442-en.html

Di Guardo, A., Trovato, F., Cantisani, C., Dattola, A., Nisticò, S.P., Pellacani, G., Paganelli, A. Artificial Intelligence in Cosmetic Formulation: Predictive Modeling for Safety, Tolerability, and Regulatory Perspectives. Cosmetics 2025, 12(4), 157: https://www.mdpi.com/2079-9284/12/4/157

4
Show Comments (0) Hide Comments (0)
0 0 votes
Article Rating
Subscribe
Notify of
guest
0 Comments
Oldest
Newest Most Voted
Inline Feedbacks
View all comments
0
Would love your thoughts, please comment.x
()
x