Artificial Intelligence, Data Protection, Privacy, and Doxxing

Abstract

“Hi, are you Cathy who went to Menlo University and recently had your eyelids done at Park Hospital?” We certainly would not want a stranger to ask one of our patients this question at the bus station. And yet, this could happen. Recently, students at Harvard University created a tool to heighten awareness about privacy, cybersecurity, and doxxing in the age of large language models (LLMs) and artificial intelligence (AI)1,2 In a recent Commentary published in Aesthetic Surgery Journal, Chris Nuland, the legal counsel of The Aesthetic Society, discusses questions raised by another article which used AI to analyze periorbital rejuvenation.3,4 We highly recommend reading this article because it provides important information about privacy and data-protection concerns and Health Insurance Portability and Accountability Act (HIPPA) compliance. This is not only a legal issue but also an ethical one, raising important questions about privacy in AI research. This journal is proud to be a pioneer among surgical journals with a dedicated AI section. Aesthetic Surgery Journal publishes among the largest number of papers on AI among all plastic surgery journals.5 The Editorial Board and its reviewers take their responsibility seriously to analyze submissions carefully for privacy and data-protection issues. This concerns the privacy of patient data, as well as protection of authors’ intellectual property when manuscripts are submitted for review. Newer LLMs are capable of summarizing, commenting, and reviewing articles. To test this yourself, upload a published article—or even one of your own manuscripts—to one of the newer free LLMs and prompt it as follows: “You are a peer reviewer for a scientific medical journal. Analyze the attached text for form and content and point out its strengths and weaknesses.” Although the LLM may make comments you disagree with, you will likely be impressed with how it synthesizes the essential elements of the paper and provides quite reasonable recommendations. We can envision a future where more accurate and informed LLMs may contribute to the peer-review process for medical journals as a primary or supplementary “reviewer.” Perhaps one day authors will submit their manuscript together with an AI-generated review and publication recommendation. However, once information is shared with an LLM or AI platform, that data may be used by the platform for training purposes or to develop future, yet to be determined, projects. Therefore, for now and while humans continue to review submitted manuscripts, there is a responsibility to protect the authors’ intellectual property, which cannot be guaranteed when a manuscript is uploaded to an AI platform. Aesthetic Surgery Journal instructs reviewers not to use LLMs for their reviews because authors do not consent to having their research shared in an uncontrolled manner before acceptance and publication. Among the major AI research companies, Anthropic (San Francisco, CA) guarantees that its product Claude “will not use … prompts and conversations from Free Claude.ai or Claude Pro to train our models.”6 Yet, this does not mean that authors—the owners of the intellectual property—agree that journals may forward their manuscripts containing the unpublished results of their research to anyone for any reason. Awareness of potential breaches in patient confidentiality when conducting AI-related research is also vital. With the vast amounts of information available on the internet, even deidentified data can be reidentified via a method called “data triangulation.” Data triangulation refers to the concept that if datasets collected via various methods include some overlapping information, and if some of these datasets include identifying information, then those datasets can be merged to reidentify individuals in otherwise deidentified datasets.7-9 Thus, even deidentified data in a study could later be reidentified by another user or at some future date with new and evolving technology. Furthermore, these deidentified data may later be reidentified by an outside user or platform that is not subject to a business associate agreement, IRB scrutiny, the European Union's General Data Protection Regulation (GDPR), and/or HIPAA-compliant regulations, and consequently put patient privacy at risk. Such considerations apply not only to those conducting AI-related research: many LLMs and AI platforms are trained using publicly available information. Patient photographs previously published in a peer-reviewed journal or shared on social media or a practice website pose similar risks. Many countries have policies specifically regarding health information, and researchers should be aware that data-privacy protections differ across countries. Because data shared with AI platforms may be used in unknown or unintended ways, health data could unintentionally be shared in ways that violate patient privacy laws. for instance, the GDPR focuses on actions that directly affect the European Union. But because the GDPR also places limits on the processing of data about EU residents, regardless of where that citizen resides, the rules may affect the privacy practices of non-EU entities. Furthermore, as with other electronic records, even HIPAA- or GDPR-compliant platforms are at risk of cybersecurity breaches that may jeopardize patient information and patient safety.9 There are many potential benefits to using AI for research.10 In recent years, these platforms have become widely accessible and offer an inexpensive way to potentially minimize human bias (though cultural and gender biases abound on these platforms).11,12 Many publicly available platforms rely on user prompts and other data for teaching—either to train the current platform for future use or develop new applications. In the case of photographs, this may include facial recognition tools. As the scenario posed at the beginning of this article demonstrates, a delicate balance must be struck when using AI to analyze facial photographs. In the age of AI, it is incumbent upon physicians, researchers, and IRBs (when involved; many studies deemed “deidentified” do not undergo formal IRB review) to be aware of the potential risks to patient privacy posed by AI. Potential legal issues are beyond the scope of this Editorial and we defer to the legal experts, but is it still ethically sufficient for patients to sign a general consent form stating they allow us to use their data, including photographs, for research purposes and social media? Are patients fully aware of what can and probably will happen to their data once these become publicly available to third-party AI platforms? Is it sufficient for a study performing facial analysis using AI to be deidentified or are additional protective measures needed to protect patient information? This article is not meant to instill fear, but to raise awareness and begin a discussion of how we as scientists, clinicians, and members of society should incorporate AI. For the protection of intellectual property, one solution could be the exclusive use of open-source LLMs such as Llama (Meta AI, New York, NY) or Mistral (Mistral AI, Paris, France) that can be installed on a computer or server and can operate without an internet connection. With a rapidly increasing number of parameters, they can easily compete with classical LLMs and could be useful for tasks such as summarizing, analyzing, and, possibly, reviewing texts. Although these platforms require powerful servers and expensive graphics processing units that would likely be prohibitive for the individual researcher, this should not be an obstacle in a hospital or university, or for a publisher. For research based on patient data, and especially pictures, such open-source LLMs are only of very limited use, and HIPAA or applicable healthcare privacy regulation–compliant platforms may be necessary. Unfortunately, most publicly available AI platforms are not HIPAA-compliant or are limited due to their reliance on user input to train their algorithms.13 If physicians want to use such tools in research, a standard patient consent will likely no longer be sufficient. They need to disclose to patients that their data may be used in unregulated ways and respect a patient's decision if they refuse. They should also be aware that many of these platforms have “opt-out” data privacy policies and some specifically warn users not to enter confidential information.14 Rickert was among the first to warn about the dangers that AI poses to the privacy of patients.7 In times when physicians, researchers, and authors may use chatbots and LLMs to facilitate their studies, the topic is more relevant than ever. Our patients deserve the best possible protection for their private data for both legal and ethical reasons. It is our responsibility to provide this. Dr Oppikofer is the Artificial Intelligence section editor for Aesthetic Surgery Journal. None of the authors have relevant financial disclosures. The authors received no financial support for the research, authorship, and publication of this article.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 12 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.