What AI Skin Tools Can and Cannot Diagnose

The Skin Diagnosis Problem AI Set Out to Solve

The most persuasive version of any new technology is the access argument. Frame a tool as a solution to a resource shortage and it immediately sounds humanitarian. It short-circuits skepticism. And in dermatology, the underlying problem is real: roughly three billion people globally lack adequate care for skin conditions, approximately five million skin cancers are diagnosed in the U.S. every year, and in most healthcare systems, a GP without specialist backup is the first and sometimes only person a patient ever sees about a suspicious lesion.

So the pitch for AI skin tools writes itself. Scalable. Consistent. A first-pass screening layer that gets you closer to the right answer faster. It is a reasonable premise. The harder question is whether the tools actually deliver on it, and under what conditions.

Not all AI skin tools are the same category of thing. There are FDA-cleared clinical devices, clinician-facing software platforms, and a sprawling, largely unregulated consumer app ecosystem. These categories carry different evidentiary standards, different regulatory histories, and very different levels of trustworthiness. Conflating them is probably the most common mistake you can make when evaluating this space, and it is worth resisting from the start.

Where the Evidence Shows Real Strength: Melanoma and Pigmented Lesions

Start with the area where the evidence is striking.

A 2025 review of 551 studies found that convolutional neural networks, a type of AI trained to recognise patterns in images, and the architecture underlying most image-based AI diagnostics, achieve 91% sensitivity and 94% specificity when distinguishing melanoma from benign lesions. That is not marginal. That is a meaningful signal, and it holds up across multiple lines of evidence.

In primary care settings specifically, a 2024 meta-analysis reported 90% sensitivity and 85% specificity for detecting suspicious pigmented lesions. Pooled across all available studies, AI achieves 87.0% sensitivity and 77.1% specificity, versus 79.78% sensitivity and 73.6% specificity for clinicians overall. Both differences are statistically significant.

The comparison against expert dermatologists narrows: AI at 86.3% sensitivity and 78.4% specificity, experts at 84.2% and 74.4%. Clinically comparable. Not superior, but comparable. And in a multicenter Lancet Oncology study, dermatologists working with AI assistance outperformed dermatologists working alone. The augmentation effect is real, not theoretical.

Scale tells part of the story too. One platform analyzed over 18,000 skin cases in a single month and surfaced 79 melanomas, 95 squamous cell carcinomas, and 202 basal cell carcinomas, roughly equivalent to five percent of the UK's monthly melanoma detections. That is production performance, not a laboratory result.

But here is where I want to slow down, because this is the part that gets elided in most coverage. These numbers are specific to dermoscopic imaging of pigmented lesions, which is the condition type with the deepest training data, the most regulatory scrutiny, and the longest development history. The evidence base for melanoma detection is not representative of what AI skin tools can do across the full range of dermatological conditions. When you see a headline accuracy figure, ask yourself: accurate on what, under what conditions, and in whose hands?

A Wider Diagnostic Range: Inflammatory Conditions and Severity Grading

The accuracy figures for inflammatory conditions look impressive at first glance. A systematic review of 64 deep learning models reported 94% accuracy for acne, 94% for rosacea, 93% for eczema, 89% for psoriasis. Those numbers reflect real model capability, but they are drawn from controlled study conditions, which is a significant qualifier.

Severity grading is where variability enters. Psoriasis severity grading accuracy ranges from 93% to 100%. Eczema reaches 88%. Acne severity grading lands between 67% and 86%, a spread wide enough to matter clinically. One systematic review and meta-analysis covering over 70,000 test images reported pooled sensitivity of 0.91 and specificity of 0.64, with confidence intervals that ran from 0.74 to 0.97 for sensitivity and 0.47 to 0.78 for specificity.

That heterogeneity is the real signal. It reflects how much performance depends on factors the summary statistics obscure: image quality, study population composition, disease prevalence in the sample, and validation methodology. When a confidence interval for specificity spans from 0.47 to 0.78, you are not looking at a stable finding. You are looking at a technology whose performance varies enormously based on conditions that most published papers do not make explicit.

Where AI Adds the Most Clinical Value: Augmenting Non-Specialists

The most significant performance gap in the literature is not between AI and expert dermatologists. It is between AI and generalist clinicians. One meta-analysis found AI sensitivity of 92.5% against generalist sensitivity of 64.6%. Nearly 28 percentage points. That gap is where AI-assisted diagnosis has its clearest, most defensible claim to utility.

DermaSensor, FDA-authorized via the De Novo pathway in January 2024 and the first AI-enabled device cleared for primary care skin cancer detection in the U.S., illustrates this concretely. Its pivotal study demonstrated 96% sensitivity across 224 skin cancers in a trial covering over 1,000 patients across 22 sites, led by Mayo Clinic. In a smaller companion trial, primary care physician sensitivity rose from 67% to 88% when supported by the device. A 21-percentage-point lift, in primary care, in exactly the setting where the access problem is most acute.

The trade-off is specificity. DermaSensor's overall specificity in the larger trial was 21%. Nevisense, cleared in 2017, shows a similar pattern: 96% sensitivity, 34% specificity. MelaFind, cleared via the PMA pathway and now discontinued, achieved 10% specificity, leading to excess biopsies, poor workflow integration, limited coverage, and eventual market exit. That trajectory is worth understanding. High sensitivity with very low specificity means a substantial proportion of people flagged for further evaluation are cancer-free. The follow-up has costs: financial, logistical, psychological. Not direct harm in the clinical sense, but not trivial either.

The design philosophy behind these devices prioritizes minimizing missed cancers at the cost of false positives, which is the right trade-off for a screening tool in primary care. Whether the current calibration is correct is a legitimate ongoing debate.

The Regulatory Landscape: What "Approved" Actually Means

As of 2025, fifteen regulatory-approved AI dermatology devices have been identified globally. The U.S. has three FDA-approved systems. That number is small relative to the volume of tools in circulation, and the gap matters.

FDA clearance via the De Novo pathway is meaningfully different from PMA (Premarket Approval) authorization. PMA is reserved for high-risk devices and requires prospective clinical evidence of both safety and effectiveness. De Novo applies to novel, low-to-moderate-risk devices that have no comparable approved device to reference. Both pathways involve rigorous review; neither is a rubber stamp.

What clearance signals is a defined, validated use case. DermaSensor is cleared for skin cancer detection in primary care. That authorization does not extend to general-purpose skin diagnosis, inflammatory condition monitoring, or anything else the device might theoretically be applied to. This distinction gets lost in product marketing with predictable regularity.

The FDA's finalized guidance on Predetermined Change Control Plans, which allows AI models to update post-clearance within pre-approved parameters, is a substantive development. Roughly 10% of 2025 clearances included PCCPs (Predetermined Change Control Plans), suggesting the regulatory framework is beginning to accommodate the iterative reality of machine learning development. International platforms have received clearance for broader applications, but regulatory standards vary significantly by jurisdiction, and clearance in one market does not imply equivalence in another.

The Consumer App Gap: No Approval, Limited Evidence, Real Privacy Risk

A JAMA Dermatology scoping review found that the majority of AI-based dermatology apps available to U.S. consumers lacked peer-reviewed research support and were developed without clinician involvement. None held FDA approval. Nearly a quarter claimed diagnostic capabilities without scientific backing of any kind.

Forty-six percent of apps did not disclose their policies on user-submitted images. You are uploading photographs of your skin, potentially identifiable in ways you have not considered, to platforms with no stated policy on storage, sharing, or secondary use. That is a privacy exposure worth pausing to evaluate.

The claimed application categories from that review span a wide range: skin cancer detection, general skin and hair condition diagnosis, mole tracking, condition tracking, acne diagnosis, atopic dermatitis management. A broad set of promises, most without validated evidence behind them.

In a direct evaluation of consumer app accuracy, the AI included the correct diagnosis in its top five guesses roughly half the time. It returned the correct diagnosis first only 23% of the time. That is not a tool equipped to make consequential health decisions, and the design of most consumer apps does nothing to communicate that limitation to you.

What a Photo Cannot Tell: The Limits of Image-Only Assessment

An AI algorithm analyzing a photograph is doing exactly that: analyzing a photograph. It cannot palpate a lesion. It cannot assess surrounding tissue. It cannot ask about your medication history, evaluate texture under dermoscopic magnification, or know that a rash appeared three days after you started a new drug. Clinical diagnosis is a multivariate process, and an image captures one variable.

Image quality is a direct performance variable, and it is rarely communicated clearly to users. Lighting, angle, camera resolution, motion blur: all of these affect output in ways that are difficult to quantify without controlled conditions. The accuracy range in primary care settings spans from 58% to 96.1% sensitivity across published studies, with accuracy scores ranging from 0.41 to 0.93. That spread reflects how much image conditions and clinical context actually matter, not random noise.

Commercial systems achieve only 68% accuracy on early-stage nodular melanoma, the presentation where early detection is most consequential, due to insufficient recall optimization. A 2026 real-world assessment published in Hospital Healthcare Europe found that physicians achieved mean diagnostic accuracy of 65.9%, compared to 56.7% for a first-generation CNN, 72.2% for PanDerm unimodal, and 66.3% for PanDerm multimodal. Expert dermatologists with ten or more years of dermoscopy experience reached 74.2%, outperforming all AI systems tested. Under real-world clinical conditions, with actual patients and actual image variability, expert human judgment still leads.

The WHO's 2025 digital health assessment identified structural limitations that go beyond image quality: AI confidence metrics too binary to support nuanced clinical reasoning, fragmentation between visual analysis and clinical reporting workflows, and insufficient patient education components in commercial systems. These are not edge cases. They are design problems embedded in the majority of tools currently available.

The Skin Tone Equity Gap: A Structural Bias Built Into Training Data

A 2025 meta-analysis found a pooled AUROC (area under the receiver operating curve, a measure of overall diagnostic accuracy, where 1.0 is perfect) of 0.89 for lighter skin tones (Fitzpatrick I through III) versus 0.82 for darker skin tones (Fitzpatrick IV through VI). Seven points in area under the curve translates directly into missed diagnoses.

The Stanford DDI dataset evaluation showed one widely-cited model dropping from an AUROC of 0.72 on lighter skin tones to 0.57 on darker skin tones, falling below the clinically useful threshold entirely. That is not a marginal performance difference; that is a tool that stops being reliable for a significant portion of the population it is supposed to serve.

The root cause is training data. The ISIC dataset, the most widely used pathologically confirmed dermoscopic image repository, lacks sufficient diversity across skin tones, disease types, and clinical presentations. The populations with the least access to dermatologists are frequently the same populations underrepresented in the data these models learned from. The equity problem compounds the access problem. A tool designed to close gaps in care is, in its current form, least reliable for the patients who most need it to work. That is not a minor footnote. It is a structural flaw in the premise.

Diverse training sets and standardized validation protocols are identified as research priorities. That is the right direction. It is also years away from widespread implementation, which means that if you have darker skin, you face compounded unreliability right now.

Reading AI Outputs Accurately: What the Numbers Do and Don't Mean

Sensitivity and specificity are not interchangeable, and the trade-off between them changes how you should interpret any AI output. Sensitivity measures how well a tool catches real cases, high sensitivity means fewer missed diagnoses, but also more false alarms. Specificity measures how well it rules out non-cases, high specificity means fewer unnecessary referrals. DermaSensor's 96% sensitivity paired with 21% specificity is the clearest illustration of what that trade-off looks like in practice. A tool optimized to miss as few cancers as possible will flag many things that are not cancer.

Population context matters too. The same specificity figure means something very different in a high-risk screening population versus a general consumer app context. A high-sensitivity tool used where disease prevalence is low generates many false alarms relative to true finds.

Consumer apps frequently present outputs as binary: high risk or low risk. What they rarely disclose is the underlying probability distribution, the study conditions under which performance was measured, or how those figures shift based on image quality, skin tone, or condition type. Research on human-AI combined judgment suggests that treating AI output as one input among several, rather than a verdict, improves overall accuracy. The tool is most accurately understood that way.

Published accuracy figures from clinical trials do not generalize to the photograph you took in your bathroom mirror. The EU AI Act and evolving international frameworks are beginning to require more rigorous accuracy thresholds and transparency disclosures, which will eventually give you better tools for evaluating these products. For now, the gap between marketing claims and documented performance requires active investigation on your part.

A Practical Framework: When to Trust, When to Verify, When to See a Doctor

The evidence points toward a few specific orientations, not general conclusions.

AI skin tools are most reliable when you're screening pigmented lesions, with good-quality dermoscopic or smartphone images, on lighter skin tones, and in conjunction with a clinician. That is the condition set with the deepest evidence base and the most validated performance data. If your situation matches that description, these tools carry real signal.

They add clearest value when used by non-specialist clinicians as a decision-support layer, or by patients trying to determine whether something warrants urgent versus routine follow-up. A GP using DermaSensor to decide whether a mole needs same-week referral: that is the intended use case, and the evidence supports it. A patient using an unvalidated consumer app to decide whether to skip the appointment entirely: that is where risk accumulates, quietly and without much feedback.

Treat consumer app output as a signal, not a verdict. Especially for darker skin tones, early-stage lesions, and anything that requires clinical context beyond a single photograph. FDA clearance is a meaningful quality signal, but only for the specific use case the tool was authorized to address. Check what a tool is actually cleared to do before extending trust beyond that boundary.

Red flags worth screening for actively: no peer-reviewed validation cited, no clinician involvement disclosed, no transparency on image data privacy, diagnostic claims that exceed the documented evidence. These are common features in the consumer app space.

A negative AI result does not rule anything out. Sensitivity below 100% means some true cases will be missed. For any lesion that is changing, growing, bleeding, or persistently abnormal, clinical follow-up remains essential regardless of what a screening tool returns.

The researchers doing the most rigorous work in this field converge on the same framing: AI integrated into clinical judgment, not substituting for it. The question worth asking about any tool you encounter is not whether it is impressive. It is whether it is designed to bring you closer to a clinician or to convince you that you no longer need one.