Computer vision models tackling counterfeit cosmetics accurately detect typographical errors but show excessive sensitivity to lighting glare

A practical test conducted by Grover Laboratory, a specialist in counterfeit drug detection technologies, has demonstrated that multimodal AI models can inspect cosmetic packaging and identify counterfeit products from photographs, while also highlighting technical limitations linked to lighting glare and camera angles.

The experiment fed the reasoning version of the Gemini model six detailed images of each package of Rhode lip balm, a product subject to widespread counterfeiting according to previous press reports. The samples included authentic packages and counterfeits purchased from unauthorised vendors on e-commerce platforms, where studies estimate that two-thirds of skincare and makeup products listed on certain online marketplaces and direct-selling apps are counterfeit, exposing consumers to heavy metals and harmful bacterial contaminants.

The model succeeded in detecting subtle differences that evade rapid visual inspection, including optical character recognition errors in complex chemical ingredient lists.The AI identified an exclamation point replaced with a lowercase letter l in the name of a chemical compound, caused by unverified optical scanning of authentic packaging, as well as mismatches in responsible distributor information between the outer box and the inner tube, the repetition of a specific batch number associated with counterfeit copies, the substitution of the numeral zero for a Latin letter in manufacturer postal codes, and language errors in Italian recycling instructions and French translations.

Conversely, the results revealed clear vulnerabilities when the model relied entirely on standard images. It misinterpreted domestic lighting glare and reflections on plastic and paper creases as typographical errors in compound names and instructions, a mistake a human reviewer would not make. The model also produced incorrect assessments of authentic design details, such as assuming the absence of geometric embossing around the logo on the outer carton, or treating closely spaced volume figures as evidence of counterfeiting despite their presence on authentic units.

This test carries practical implications for the e-commerce sector and retail supply chains in the Gulf and the Middle East, where beauty sales across social platforms and digital applications are growing rapidly. These capabilities offer quality inspection teams at commercial platforms and local retailers a low-cost initial tool to verify suppliers and shipments without requiring complex chemical laboratories during exploratory stages. However, relying on these models requires establishing standardised imaging protocols to eliminate light reflections and avoiding rigid automated decisions, while retaining human verification for compliance before rejecting or seizing goods.