The study, led by Allison Roecker and colleagues at Texas A&M Naresh K. Vashisht College of Medicine with collaborators at Baylor University Medical Center, Epiphany Dermatology, and the University of Texas Southwestern Medical Center, was designed to test real-world performance rather than laboratory ideal conditions. Each of the twelve clinical images was presented to every model five times using a standardized prompt, allowing the researchers to assess not only whether the systems could identify psoriasis correctly but also whether they gave consistent answers. That consistency measure proved revealing, because most models produced different responses when shown the same image on repeated trials.

The most striking finding concerned four of the twelve images, which no model diagnosed correctly in any trial. In other words, a third of the test cases represented manifestations of psoriasis in skin of color that current AI systems simply could not recognize. For patients who increasingly turn to consumer health tools and chatbots for preliminary guidance, this blind spot carries real consequences. A person with a misdiagnosed or unrecognized rash may delay seeking dermatological care, and psoriasis that goes untreated can worsen both physically and psychologically.

Psoriasis itself is a chronic inflammatory condition that affects roughly the same prevalence across racial groups, but its visual presentation differs markedly depending on skin tone. On lighter skin, plaques typically appear red or pink with silvery scale. On darker skin, the same disease may manifest as violet, brown, or hyperpigmented patches with less apparent scaling, and post-inflammatory pigment changes can obscure the underlying pattern. These differences are well documented in the dermatology literature, yet the datasets used to train most AI diagnostic tools remain dominated by images of lighter skin.

That training imbalance is the likely root of the failures documented in the new study. Machine learning models learn to recognize patterns from the examples they are shown, and if the examples underrepresent darker skin tones or depict psoriasis only in its lighter-skin form, the models will systematically underperform on patients who fall outside that narrow visual template. Previous research has reinforced this concern: studies of online psoriasis imagery have found significant skin tone bias in the pictures available on the web, and analyses of patient-facing dermatology websites have shown that people with skin of color are underrepresented in educational materials as well.

The consequences of this representation gap extend beyond the digital realm. Dermatology has long faced documented disparities in care for patients with skin of color, including delayed diagnosis, differences in treatment access, and a higher comorbidity burden. Researchers have described these inequities in terms of diagnosis, disease burden, drug development, and delivery of care. If AI tools are now becoming a front door to medical information for millions of people, a system that cannot recognize disease in darker skin risks compounding those structural problems rather than alleviating them.

The new study also speaks to a broader technical challenge in evaluating large language models for medical image interpretation. Unlike traditional diagnostic algorithms that produce a fixed output for a given input, modern multimodal models can generate variable responses to identical prompts, a property that undermines reliability in clinical settings. The inconsistency observed across repeated trials in this research suggests that even when a model occasionally identifies psoriasis correctly, patients and clinicians cannot count on that answer being reproduced. In diagnostic medicine, where reproducibility is a cornerstone of trust, this variability is a serious limitation.

The authors emphasize that their findings should prompt caution among both patients and clinicians. AI assessments, they argue, should be accompanied by clinical context and examination by a dermatologist before any treatment plan is formulated. This recommendation aligns with a growing consensus in the medical AI community that these tools are best understood as adjuncts rather than replacements for professional judgment, particularly in areas where their performance has not been rigorously validated across diverse populations.

The research team calls for further evaluation of AI and large language model performance across other skin conditions, noting that dermatologic diseases in skin of color often present differently and can easily be overlooked. Psoriasis is far from the only condition affected; vitiligo, eczema, lupus, and skin cancers all have distinct appearances on darker skin that may be underrepresented in training data. Systematic benchmarking efforts like this one, which use standardized prompts and repeated testing to measure both accuracy and precision, offer a template for holding these systems accountable as they proliferate.

For now, the message for the public is clear. The most sophisticated consumer AI tools available today misidentify or fail to identify psoriasis in skin of color in the substantial majority of cases, and some presentations defeat every model tested. Until training datasets become genuinely representative of the full spectrum of human skin, anyone using an AI tool to assess a skin concern should treat the output as a starting point for a conversation with a qualified dermatologist, not as a diagnosis. The technology holds enormous promise for expanding access to dermatological expertise, but this study is a reminder that promise is only as good as the data behind it.

Subject of Research: Diagnostic accuracy of artificial intelligence and large language models for recognizing psoriasis in skin of color

Article Title: Diagnostic accuracy of artificial intelligence for psoriasis in skin of color

Article References: Roecker, A., Sosale, P., Nadir, U., & Tolkachjov, S. N. (2026). Diagnostic accuracy of artificial intelligence for psoriasis in skin of color. Archives of Dermatological Research, 318(1), Article 520. https://doi.org/10.1007/s00403-026-05007-6

Image Credits: AI Generated

DOI: 10.1007/s00403-026-05007-6

Keywords: artificial intelligence, psoriasis, skin of color, diagnostic accuracy, dermatology, large language models, health equity, machine learning, skin tone bias, medical imaging, health disparities, Diagnostic