The Intelligence Layer
Artificial intelligence is transforming hair diagnostics from subjective visual assessment into quantified, reproducible analysis. Machine learning algorithms can process images, spectral data, and sensor readings at speeds and precision levels that exceed human capability — identifying patterns, classifying characteristics, and generating insights that would be impossible through manual examination alone.
But AI in hair diagnostics faces a critical challenge that mirrors the broader equity issues CROWN was founded to address: if the data used to train AI systems is biased, the AI itself will be biased. And in the field of hair assessment, training data has historically been profoundly biased toward Eurocentric hair types.
CROWN’s AI classification engine is designed to address this challenge from inception — building intelligence that serves every hair type with equal accuracy.
Current AI Hair Diagnostic Systems
Several AI-powered hair diagnostic systems are currently available:
Kérastase Diagnostic Capillaire. L’Oréal’s AI system analyses scalp and hair images to recommend products and treatments. The system uses computer vision trained on L’Oréal’s proprietary data.
L’Oréal My Hair Diagnosis. A consumer-facing app that analyses photographs to assess hair type and condition, providing personalised care recommendations.
MyHair.AI. An independent platform offering AI-based hair and scalp analysis from user-uploaded photographs.
MDhair. An AI dermatology platform that includes hair and scalp assessment capabilities.
Revieve. A beauty technology company offering AI-powered hair analysis for retailers and brands.
Becon. An AI platform providing hair diagnostics for professional salon use.
These systems represent meaningful advances in making hair assessment more accessible and consistent. However, they share a common limitation: their training data is disproportionately composed of straight and loosely waved hair types. Textured hair — Afro-textured, tightly coiled, and kinky hair — is systematically underrepresented in the datasets that train these systems.
The Bias Problem
AI bias in hair diagnostics operates through several mechanisms:
Training data imbalance. Machine learning models learn from examples. If 80% of training images feature straight or loosely waved hair, the model becomes highly accurate for those types and progressively less accurate for textures further from the training distribution. Tightly coiled hair, which represents a minority of training data in existing systems, receives the least accurate classification.
Feature extraction bias. AI systems learn which features are important for classification from their training data. Systems trained predominantly on straight hair learn to focus on features that distinguish among straight hair subtypes — but may miss the features that are most informative for classifying textured hair.
Validation bias. When AI systems are tested for accuracy, the test data often reflects the same imbalance as the training data. A system that scores 95% accuracy on a test set dominated by straight hair may score much lower on textured hair — but the aggregate accuracy figure obscures this disparity.
Output framing. Many AI hair systems are designed to recommend products or treatments — and these recommendations often default to Eurocentric care paradigms. A system that recommends “smoothing” or “frizz control” for hair it cannot accurately classify is encoding bias in its output.
CROWN’s Approach to Equitable AI
The CROWN Diagnostic’s AI classification engine addresses these challenges through several design principles:
Diverse training data. The CROWN Hair Commons is designed from inception as a multi-ethnic dataset, ensuring that the AI has abundant training examples across all hair types — from Type 1A to Type 4C and everything between. Universal coverage is the architecture, not a feature added after the fact.
Multi-modal input. Unlike vision-only systems, CROWN’s AI processes data from multiple sensor modalities: optical imaging (visual and structural features), near-infrared spectroscopy (chemical composition), and impedance sensing (porosity). This multi-modal approach provides richer information for classification and reduces reliance on visual features alone — which is precisely where bias toward Eurocentric types is most concentrated.
CROWN Hair DNA output. Rather than classifying hair into the reductive Walker system categories, CROWN’s AI generates multi-dimensional profiles capturing 12+ characteristics. This avoids the forced categorisation that compresses textured hair diversity into a few categories.
Bias testing. CROWN’s validation methodology tests accuracy disaggregated by hair type and ethnicity, ensuring that the system’s performance is equitable across all populations — not merely adequate in aggregate.
Reproducibility. A core requirement is that CROWN’s AI produces consistent results across devices and operators. This reproducibility — essential for research use — requires careful calibration protocols and device standardisation, developed through the ETH Zürich collaboration.
Open Research Challenges
Several open questions remain in AI-powered hair diagnostics:
How can we build models equally accurate across all ethnic hair types with limited initial training data? This is a classic small-data machine learning challenge — CROWN’s approach of collecting multi-modal data may partially compensate for smaller sample sizes by providing richer per-sample information.
How do we handle mixed-heritage hair? Individuals of mixed heritage may exhibit combinations of structural properties that do not fit cleanly into population-level categories. AI systems must handle this continuous variation gracefully.
How can we ensure privacy in biometric AI? Hair data is biometric and potentially identifying. CROWN’s data infrastructure incorporates privacy-by-design principles, including anonymisation at collection, to protect participants while enabling research.
How do we prevent commercial bias? AI systems that are optimised to recommend products may prioritise commercially relevant classifications over scientifically accurate ones. CROWN’s AI is designed for research accuracy, not product recommendation.
The Broader Stakes
AI in hair diagnostics is a microcosm of the broader challenge of building AI that serves all communities equitably. The biases that affect current hair AI systems — skewed training data, feature extraction optimised for majority populations, validation that obscures disparities — are the same biases that affect AI in healthcare, hiring, criminal justice, and other consequential domains.
CROWN’s approach — building equitable AI through diverse data, multi-modal analysis, disaggregated validation, and transparent methodology — provides a model that extends beyond hair diagnostics. It demonstrates that equity is not a constraint on AI performance but a design principle that produces better systems.
The CROWN Diagnostic and its AI classification engine are not merely tools for hair analysis. They are demonstrations of how technology can be built to include rather than exclude — how the data gap can be closed not by perpetuating historical biases but by designing systems that serve everyone from the beginning.
For more on the research questions driving this work, see Open Research Questions and the ETH Zürich collaboration.
