Artificial intelligence tools designed to help dermatologists are failing patients with darker skin tones because the technology relies on biased databases that predominantly feature White patients, according to presentations at the American Society for Dermatologic Surgery’s 2025 Annual Meeting in Chicago.
Dr. Cheryl Burgess highlighted how cultural biases have corrupted the datasets used to train AI models. Internet searches for “beautiful women” still overwhelmingly show White women despite modest improvements over the past decade, reflecting the skewed data that feeds these medical programs. A literature review published in the International Journal of Dermatology found that current AI programs perform worse at identifying lesions in skin of color, with only 30 percent of reviewed programs reporting any data about their use in dermatology for diverse populations.
The consequences extend beyond diagnostic accuracy. Dr. Vivian Bucay noted that images used to train AI can be manipulated through angles and lighting, introducing additional bias. There are no well-annotated image datasets representing diverse skin tones, creating significant variability in outcomes.
Dr. Jane Yoo, who tested an AI medical scribe program, discovered translation errors that could have serious legal implications. A documentation mistake changing “stressed” to “depressed” could affect insurance coverage and create liability issues for physicians. The technology remains rough around the edges, potentially useful for trainees building vocabulary but inadequate for experienced practitioners who need precision.
Significant development is required to ensure AI accurately represents darker skin tones in medical datasets.
See: “Bias, Subjective Outcomes Slow AI Advances in Dermatology” (November 18, 2025)
