Dermatology responds to AI shift with evidence-based practice recommendations.
Artificial intelligence is no longer just analysing skin images in dermatology – it’s listening to patient consultations, drafting clinical letters and shaping diagnostic conversations.
In response to the rapidly evolving landscape, the Australasian College of Dermatologists (ACD) is expected to release updated guidance on the use of AI in clinical practice, spotlighting the growing role of large language models (LLMs) like ChatGPT and skin-specific tools such as SkinGPT.
The update will incorporate a paper, published in the Australasian Journal of Dermatologyand backed by the ACD, that responds to the growing need for practice recommendations that are both evidence-based and regionally attuned.
The ACD’s first position statement on the Use of Artificial Intelligence in Dermatology in Australia was published in 2022.
The initiative responds to the growing need for practice recommendations that are both evidence-based and regionally attuned.
The original statement focused on image-based AI tools, while the latest paper addresses the explosion of LLMs in medicine – specially AI scribes – which has shifted the conversation from diagnostics to documentation and beyond.
LLMs can streamline workflows by summarising patient histories or drafting educational material, while SkinGPT, a dermatology-specific AI, has shown promising accuracy in trials.
But like all LLMs, it remains susceptible to hallucinations, omissions and automation bias, the lead author, Professor Liam Caffery, deputy director (Research) for the University of Queensland’s Centre for Online Health, told Dermatology Republic.
“I think the reality is that particularly AI scribes are going to become part of medical practice,” he said.
“There’s a couple of big take homes [from the latest paper]. The first is that these products aren’t regulated by the TGA, which I don’t think is a bad thing, but it just puts the onus on dermatologists to be aware of a number of things.
“When they’re implementing [the use of AI tools] they need to ensure that privacy is protected, and we’ve outlined some simple ways of actually doing that.
“The other thing that I think they need to be aware of is that these large language models aren’t always accurate. There are omissions, there are hallucinations, and so there’s this ongoing need to continually check the output of these models.
“There’s a risk of the automation bias for people that don’t do that.”
The updated position statement emphasises three core areas: safety, privacy and ethical compliance.
Dermatologists are advised to avoid entering any identifiable patient data into public AI tools and to ensure robust data handling practices with AI scribe vendors, including Australian-based data storage, encryption and strict deletion protocols.
The authors note that, internationally, dermatology bodies are urging caution too. The American Academy of Dermatology and the British Association of Dermatologists released a joint statement in 2023 discouraging clinicians from relying on LLMs for decision-making without verification.
The paper offers concrete recommendations for dermatologists considering the use of AI tools, including:
- Human-in-the-loop oversight is essential – LLMs should never replace clinical decision-making.
- Validation in real-world settings must precede clinical implementation.
- Transparent disclosure of AI use should be standard in patient notes and communications.
- Privacy protocols must ensure no patient data is compromised when using cloud-based models.
Professor Caffery, who was the lead author on the ACD’s original 2022 position statement, said AI would never replace good dermatology practice.
“It’s most definitely value adding. And again, you just need to be aware of limitations of them,” he told DR.
He said “just about everything” from the 2022 position statement would remain, and that the document would evolve with new and emerging technologies.
“[The 2022 statement] was largely focused on computer vision models that use convolutional neural networks and this simply adds to it by looking at the newer technology,” he said.
“We haven’t essentially revisited the computer vision models. We’ve looked at the new technologies, which is the large language models and the scribes.”
