Most skin analysis AI was trained on datasets that underrepresent Latin American skin tones and undertones. Here's why that matters for conversion, returns, and trust — and what to ask vendors before you buy.

AI-powered skin analysis has quietly become table stakes on beauty e-commerce sites and apps: upload a selfie, get an instant read on oiliness, pore visibility, dark spots, or undertone, and walk away with a personalized product shortlist. For brands, it's a conversion and retention tool. For shoppers, it feels like expert advice without the wait for an expert.
But there's a structural question most decision-makers never ask before signing a vendor contract: what faces was the model actually trained on?
Most skin analysis engines sold internationally were built on datasets assembled in North America, Europe, or East Asia. For Latin American markets, that creates three concrete problems.
Computer vision research has repeatedly shown that facial analysis models perform worse on underrepresented skin tones simply because they saw fewer examples during training. When a dataset skews toward lighter Fitzpatrick types, accuracy quietly degrades as melanin density increases or undertone shifts.
Selfie-based analysis is only as good as the white balance and exposure assumptions baked into the model. Tropical daylight, warm indoor lighting, and the phone cameras common across Brazil and Latin America behave differently than the studio-calibrated images many global models learned from. The result: oiliness and redness misread, dark spots over- or under-detected.
Brazil, in particular, has one of the most mixed populations on earth. Undertone — warm, cool, neutral, olive — doesn't map neatly onto skin tone depth, and models trained on more homogeneous populations tend to collapse this nuance into broad, inaccurate buckets. Concerns like hyperpigmentation and uneven tone, which are especially common across deeper and mixed skin tones, get flattened into one-size-fits-all recommendations.
A tool that misreads skin tone doesn't just produce a slightly-off suggestion — it undermines the entire value proposition of personalization. For a brand investing in an AI advisor specifically to build trust and lift conversion, an inaccurate read for a meaningful share of visitors directly hits:
In a region where mixed and deep skin tones are the majority of the shopper base, not an edge case, this is a core accuracy problem — not a patch for later.
B4A's MaIA was trained from day one on a proprietary base of hundreds of thousands of selfies from Brazilian consumers, spanning the full range of skin tones and undertones actually shopping in the market — not a global average retrofitted with a handful of local examples.
Just as important is what happens after the analysis runs: MaIA's recommendations connect to a closed loop of real purchase and review data from B4A's own consumer ecosystem, including the glam subscription base. The model isn't only validated against labeled images — it's continuously checked against whether the products it recommended actually performed well for the skin tones they were recommended to, and refined from there.
Before selecting — or renewing — an AI skin advisor, ask vendors directly:
Vendors who can't answer these clearly are likely selling a global average dressed up as personalization.
For beauty brands expanding into Brazil and Latin America, an AI beauty advisor is only as useful as the data it was built on. A model trained on the wrong faces will consistently under-deliver for the majority of shoppers in this region, no matter how polished the interface looks. Before investing in — or expanding — a skin analysis tool here, ask where its intelligence actually comes from.
B4A Serviços de Tecnologia e Comércio S.A.
Avenida Jornalista Roberto Marinho, nº 85, 11º Andar (Conjunto 112), Cidade Monções - CEP 04576-010 - Cidade de São Paulo, Estado de São Paulo