Welcome to this July twenty twenty-six journal review. This month we're covering four pieces from the Journal of the American Academy of Dermatology, and interestingly, all four cluster around a single theme that's clearly on the editorial board's mind right now — how artificial intelligence and molecular diagnostics are reshaping the clinician's role, and what guardrails need to go up as adoption accelerates. Two are letters or comments responding to prior articles, one is a paired "con" piece from a pro-con controversy series on AI as a diagnostic aid, and one is an ethics-column Q&A on AI interpretation tools. Let's get into it.

First up is a letter to the editor, a comment responding to the recent review "Gene Expression Profiling in Dermatology, Part 2: Clinical Applications." You'll recall that review framed gene expression profiling, or GEP, as a complement to clinicopathologic assessment across melanoma and some inflammatory dermatoses, not a replacement. This comment builds on that framing with what's essentially a set of implementation cautions, and I think it's worth your attention because it crystallizes exactly the pitfalls we all sense intuitively when a rep pitches one of these assays.

The authors' central point is about indication drift. Prognostic GEP is not a diagnostic exclusion test, and it should never substitute for standard staging procedures like sentinel lymph node biopsy. They cite prior consensus work showing that routine GEP-directed management hasn't been shown to improve outcomes beyond what you already get from established staging factors — and the test performance is particularly modest in early-stage, stage one melanoma. The signal is notably stronger in stage two disease than stage one, which is a useful nuance if you're the one explaining these results to a patient.

The letter also pushes on how we talk about test performance. Their point is that accuracy or discrimination statistics alone are insufficient — clinicians need calibration, stage-specific performance, and decision-analytic measures like net reclassification improvement or decision-curve analysis benchmarked against AJCC-based staging models, so you can actually tell whether the test changes what you'd do about follow-up intensity or imaging, rather than just sounding impressive on a slide.

Then there's a pointed reference to the controversy around a commercially available forty-gene expression assay for cutaneous squamous cell carcinoma, where opaque training datasets, questionable prevalence assumptions, and selective reporting choices were shown to inflate apparent accuracy. Subsequent correspondence on that assay called for transparency around preanalytic variables, prespecified cut points, failure rates, and prospective validation — rather than marketing any single assay as a stand-alone determinant of management. If you've had SCC gene-expression results cross your desk, this is the exact debate underlying the number on that report.

Finally, the letter raises equity as a first-order concern, not an afterthought — asking that validation cohorts report performance across skin tones, acral and periungual sites, and different care settings, with postmarketing registries built to catch subgroup heterogeneity before it widens disparities.

There's no new methodology or data here — this is a commentary building on existing literature — so the practical takeaway is straightforward: nothing here is practice-changing in the sense of new evidence, but it's a useful reminder to keep GEP results framed as adjunctive, to be skeptical of any assay marketed as a stand-alone metastasis predictor, and to specifically weight stage-two melanoma results more heavily than stage-one when you're having that conversation with a patient.

Next is a controversies piece — literally titled as one half of a paired pro-con debate — arguing the "con" side of artificial intelligence as a diagnostic aid in dermatology. This is an opinion piece, not a study, so there's no methods or results scaffolding here; it's a structured argument, and it's worth hearing out because the "pro" companion piece presumably makes the opposite case, and you're being handed the counterweight.

The authors concede up front that AI currently outperforms dermatologists in several published image-analysis studies, but then lay out why that headline finding doesn't mean it's ready for unsupervised clinical use. Their argument runs on three tracks. First, technical limitations — models are only as good as their training data, and standardization problems like inconsistent image brightness, contrast, and sharpness, plus underrepresentation of rare diseases, atypical presentations, and diverse skin types, all degrade real-world accuracy in ways that don't show up in curated validation sets. Second, they raise the "black box" problem — since the reasoning behind a neural network's output is largely untraceable, physicians can monitor inputs and outputs but not the decision logic in between, which sets up exactly the kind of automation bias demonstrated in the Tschandl study, where dermatologists retained AI's answer even when they were deliberately fed inaccurate AI predictions. That's a sobering data point about how quickly clinical skepticism erodes once a tool feels authoritative. Third, they flag legal and regulatory limbo — there's no established US case law on liability when AI contributes to a diagnostic error, it's unclear whether responsibility falls on the treating physician, the software developer, or some third party, and it's equally unclear how AI use should even be framed in informed consent. Under current malpractice standards, the "reasonable physician under similar circumstances" test still puts the liability squarely on the physician regardless of how opaque the AI's reasoning was. They close with an environmental point that's less commonly raised in our literature — AI's computational footprint contributes to energy consumption and electronic waste, with global AI-related energy consumption projected to represent a substantial share of total consumption by the end of the decade.

Since this is an opinion piece, there's no dataset to critique for bias or generalizability — the value here is entirely argumentative. The practical takeaway is not practice-changing in a procedural sense, but it's a useful mental checklist before you let any diagnostic AI tool into your workflow: ask about training data diversity and skin-type representation, resist the automation-bias trap by making a habit of independently justifying agreement with an AI call rather than passively accepting it, and recognize that medicolegal responsibility currently sits with you regardless of what tool you used.

Third is an ethics-column entry, styled as a Dear Dermatoethicist letter, addressing whether it's ethical to use AI voice tools like ChatGPT for interpretation with non-English-speaking patients. Again, this is advice-column format, not original research, so we're walking through the ethical reasoning rather than a methods-and-results arc.

The letter opens by grounding the problem in real practice gaps — a retrospective cohort found that nearly two-thirds of hospitalized patients with limited English proficiency had no documented interpreter use at all during their admission, which is the access vacuum AI tools are stepping into. But the accuracy data given for AI translation is genuinely concerning for a specialty like ours that leans heavily on precise instructions. A cross-sectional study of Google Translate for patient-physician communication found that nearly half of translations failed to preserve the intended clinical meaning, and about one in six carried potential risk of harm. A separate blinded comparison of AI versus professional translators for pediatric discharge instructions found AI performed comparably to certified Spanish and Portuguese interpreters, but meaningfully worse for Haitian Creole, with about a third of outputs containing clinically significant errors. The equity concern practically writes itself — you'd be creating a two-tiered communication system where patients speaking lower-resource languages get systematically worse translation quality, which is exactly the disparity we should be trying to close, not widen.

Beyond accuracy, the column raises privacy — many commercial AI translation tools transmit audio or text to external servers and may retain inputs without a business associate agreement in place, which is a straightforward HIPAA exposure if you haven't checked your vendor's data-handling terms. And it raises autonomy — patients should be told explicitly that they're interacting with an AI tool, how it works, and what its limitations are, since some patients will be uncomfortable receiving sensitive health information through a synthetic voice or text interface while others may actually prefer the speed.

The bottom line the column lands on is measured: AI translation can serve as a temporary adjunct for basic communication or confirming simple concerns, but it shouldn't replace established human interpretation for consequential clinical information right now, and patient consent for its use is necessary. For your practice, the actionable piece is concrete — if you're using an AI voice tool for anything beyond the most basic scheduling or logistics conversation, disclose it explicitly, get consent, and route anything diagnostically or therapeutically consequential through a qualified human interpreter when one is available.

The fourth piece is another letter to the editor, commenting on a previous article called "Implications of Artificial Intelligence Scribe Usage." This one takes a specific angle that the original scribe-ethics article didn't fully address: the downstream effect of AI scribes on the premedical pipeline.

The authors' point is that human medical scribing has quietly become an important on-ramp into medicine — about one in three medical students report having worked as a scribe, and the role offers longitudinal exposure to patient care, clinical decision-making, and mentorship that's otherwise hard to get before medical school. In dermatology specifically, where shadowing opportunities are already limited, some private practices have built "gap-year" programs around scribe and assistant roles, and there's evidence this pathway particularly benefits trainees underrepresented in medicine, with positive effects on both educational experience and admissions outcomes. The concern is that as AI scribes replace human scribes for efficiency and cost reasons, this pipeline function quietly disappears unless someone deliberately preserves it.

The letter also folds in a consent data point relevant to the parent article's ethics discussion — one study found that consent rates for AI scribing fell from about eighty-two percent to fifty-five percent once disclosures fully explained the AI's features, data storage, and vendor access, a meaningful signal that patients feel differently about the technology once they understand what's actually happening with their conversation. They also note that AI scribes may still struggle with the emotional nuance of encounters like a skin cancer diagnosis discussion, and that highly technical exam content in some specialties isn't yet well-suited to AI capture.

Their proposed solution, which they're explicit is a proposal rather than tested data, is a clinic-embedded quality-assurance role, where premedical students help validate AI-generated documentation under supervision and serve as human backup when the AI falls short — preserving the pipeline function while still capturing the efficiency gains. They acknowledge this idea needs formal evaluation before anyone can say whether it actually works.

For practice, this is really an administrative and workforce-planning takeaway rather than a clinical one. If your practice or institution is adopting AI scribes, it's worth thinking proactively about how you'll preserve trainee exposure and mentorship opportunities that used to come bundled with the human scribe role, and being deliberate about consent language rather than assuming patients are indifferent to the switch.

That closes out this month's review. The throughline across all four pieces is really one message delivered from different angles — AI and molecular tools are genuinely useful adjuncts, but every one of these authors is converging on the same guardrails: transparency about how the tool works and its limitations, explicit patient consent, attention to equity and subgroup performance, and a clear-eyed acknowledgment that clinical and legal responsibility still rests with the physician. None of these four pieces changes your procedural practice today, but together they're a useful framework for evaluating the next AI or GEP tool that lands on your desk. Thanks for listening, and we'll see you next month.