Outsmarting AI beauty bias: Decoding skin concerns over beauty standards
12 Dec 2023 --- With the digitalization of skin diagnosis and beauty advice, the limits of AI’s objectivity are in question. According to recent research by Haut.AI and Novigo, AI bias in cosmetic skin care can emerge across various stages of the AI lifecycle.
In exploring solutions to address biases in cosmetic algorithms, Personal Care Insights engages with beauty SaaS solution providers EveLab Insight and Haut.AI in this first of a two-part report on overcoming biases with a focus on skin health over identity markers that may lead to discrimination.
Overcoming biases tied to identities
Providing insights into how its AI algorithms navigate the multidimensionality of beauty, Nick Howard, global strategy director at EveLab Insight, highlights its skin AI algorithms’ focus on “objective health measurements of skin rather than subjective dimensions of beauty.”
“When we look at factors of race, ethnicity or gender, we use these as a baseline to understand varying global populations and their skin’s susceptibility or resilience to common aging factors. Our skin AI algorithms employ various advanced techniques for target variable selection.”
“We train our algorithms with extensive and diverse datasets encompassing various races, skin tones, genders and ages. Meanwhile, deep learning models grasp and comprehend the intricate relationships among different features. This enables our algorithms to effectively identify and address skin variations across diverse populations,” Howard continues.
EveLab Insight follows “objective metrics and ethical standards, avoiding subjective label data.” Howard shares that the system is built on skin health indicators like wrinkles, pigmentation, moisture levels, sebum production and skin elasticity.
“We establish an objective skin dimension grading system, steering clear of subjective notions of ‘beauty,’” he stresses.
Alongside similar lines, Anastasia Georgievskaya, CEO at Haut.AI, adds how “developing algorithms for skin analysis using AI provides an opportunity to customize and optimize results based on individual characteristics and according to the specifics of each population.”
“This allows for effective and personalized outcomes regardless of gender or ethnicity. For Haut.AI, the healthy condition of the skin is a crucial indicator in model development, as we are committed to creating solutions that cater to the unique needs and diversity of every individual.”
Further, when selecting variables for algorithm development, Haut.AI relies on the latest research data.
“Research into ethnic skin characteristics is gaining momentum, and we continuously update and expand our knowledge in this field. Our ultimate objective is to provide personalized and effective care to each individual, irrespective of their skin color or gender,” she adds.
Skin health and personalization
On managing potential biases arising from cultural differences and global standardization of beauty ideals in developing AI algorithms for cosmetic skin care, EveLab Insight focuses on the skin — using various research to measure the skin’s health or the progression of skin aging.
Howard notes four main ways EveLab Insight develops algorithms:
- Objective and clear standards: We establish skin care standards based on scientific and objective criteria. This includes skin health indicators (moisture content, elasticity, etc.) and common skin issues (acne, pigmentation), emphasizing personalized skin care goals over rigid aesthetic ideals.
- Advisory experts: Collaboration with dermatologists and beauty experts from diverse regions enhances algorithm accuracy and cultural adaptability. For instance, we are jointly partnering with renowned skin professor Jean Krutmann in Europe to research cross-regional skin concerns.
- Extensive and diverse data: We collect data from populations of various cultures, skin tones, ethnicities and regions. Our devices are deployed globally, conducting compensated data collection in multiple countries.
- Optimization and feedback mechanism: During algorithm development, we employ cross-validation to ensure consistency across different demographics, optimizing algorithms to minimize cultural biases. Additionally, we encourage merchants to provide feedback on understanding algorithm adaptability across cultures for continuous improvement. User feedback is crucial in refining and adjusting our recommendation system, ensuring transparency in algorithm operations based on objective skin metrics. This empowers users to comprehend how recommendations or diagnoses are made and allows for constructive suggestions.
Moreover, Haut.AI “prioritizes skin concerns over beauty standards.”
“Our AI-powered algorithms and recommendation system are designed to identify skin concerns from images and suggest products tailored to address them. However, it’s important to note that the severity of a particular skin concern may differ depending on the skin tone, which can potentially lead to biases,” warns Georgievskaya.
“To address this, we use the best practices: We develop the so-called ‘skin atlases’ that classify and characterize skin concerns based on sensitive features such as skin tone, ethnicity and age. Our company’s external and internal skin experts create these atlases based on the most recent studies and determine each skin concern’s sensitive features and attributes.”
This data categorizes information and establishes goals for supervised deep-learning algorithms. “An anonymized and balanced dataset is obtained for each skin concern based on sensitive features and attributes defined in the internal atlas. This dataset is then used as a test dataset to measure fairness in training,” explains Georgievskaya.
In the second part of this report, Georgievskaya and Howard will look at establishing unbiased AI models, data collection and updating algorithms as beauty ideas evolve.
By Venya Patel
To contact our editorial team please email us at editorial@cnsmedia.com
Subscribe now to receive the latest news directly into your inbox.