Alia Busuttil examines the ethical trade-offs clinicians make when they invite AI into the circle of care as translators.

__________________________________________

Recently, I was involved in a case where Google Translate was used to expedite care for a patient after she presented to triage bleeding and in active labour, requiring immediate medical attention. Traditional hospital language services were unavailable because of time constraints and because no one spoke her dialect. Though this application proved instrumental in facilitating care, when analyzed retrospectively, conflicts of justice, respect for autonomy, and beneficence come readily into view.

While much has been written on evolving healthcare landscapes and the theoretical impacts of artificial intelligence (AI), comparatively little has been said about its gradual adoption in commonplace technologies, such as live translation, and the risks clinician introduce when they unwittingly invite AI into the circle of care.

As an obstetrical anesthesiologist at one of only two high-risk birthing centres in the Greater Toronto Area, the patients I serve are both complex and diverse. Within this population, language barriers are common and especially challenging when they delay safe and timely care. While hospital-based translation services exist to facilitate these clinical encounters, they are not always up to the task. Uncommon languages are routinely underrepresented in catalogues despite best efforts, and practitioners have been shown to rely frequently on translation from patient relatives or hospital staff. Concerningly, growing evidence suggests that the quality of these ad hoc translations may be inferior to those of professionals. Whether or not the same can be said of contemporary AI language tools is uncertain. Nevertheless, this differential access highlights the gaps in care that certain patients are increasingly subjected to within the healthcare system because of unmet language needs, bringing fundamental concerns of justice and equity to the foreground.

Image Description: A diagram showing AI translating Japanese and Indonesian into English.

Prima facie the use of Google Translate in this case seems justifiable through the principle of beneficence. Given that the alternative would have meant foregoing timely communication with the patient altogether, one could argue that the benefits of its use outweighed the potential risks. The acuity of this case feels immediately relevant, and indeed clinical urgency is often used to justify compromising patient rights. In practice however, the use of AI language platforms is not exclusive to emergencies; nor should it be necessarily.

Hospital based language systems are lagging in their capacity to serve diverse populations and the logistics required to utilize their services often function as a barrier, regardless of clinical acuity. Consider the case of a busy clinic and an unaccompanied patient who needs to be seen but has a significant language barrier. The time required to find or call a translator on a hospital phone is often seen as an unwelcomed delay, impacting the timely management of patients, and contributing to administrative burden. In these settings, it should come as no surprise that the instincts of many are to reach for their personal devices and get on with the delivery of care.

However, notwithstanding their undoubted practicality, these platforms circumvent the usual protocols and procedures of traditional hospital-based services meant to protect both patients and providers under the Personal Health Information Protection Act (PHIPA), or other provincial equivalents. Their use unwittingly gives private corporations, which are not defined as agents or custodians of health information under PHIPA, unfettered access to data that potentially includes private health information, leaving patients vulnerable to breaches in confidentiality and clinicians at risk of liability. Further, the accuracy of these language models and the subsequent implications on the validity of consent obtained through their use are unclear.

So how can we use these applications responsibly?  Unsurprisingly, jurisprudence struggles to maintain pace with technology that evolves at an exponential rate and the law is disadvantaged in its ability to rapidly respond to bioethical issues arising from the use of AI in healthcare. Conversely, nimble regulatory tools such as practice guidelines and hospital policies do not always capture the realities of Canada’s healthcare topography and the nuanced role these technologies play when they fill in for resources otherwise unavailable. It is likely that healthcare providers will continue to use AI as they see fit until a case is litigated through the courts that cautions them otherwise, especially when the practicality of these technologies surpasses conventional systems.

Clear and effective communication is foundational to the patient-physician relationship and a bedrock to upholding the principles of autonomy and informed consent. In pluralistic societies, AI language tools have the capacity to revolutionize care for vulnerable patients, but adaptations to practice should not trade away safety, equity, and data security for those same individuals.  Clinicians need to continue scrutinizing the use of these technologies in care circles and remain vigilant about their potential pitfalls. Doing otherwise, risks losing too much in translation.

__________________________________________

Alia Busuttil is a Staff Anesthesiologist at Sunnybrook Health Sciences Centre and a Lecturer in the Department of Anesthesiology & Pain Medicine at the University of Toronto, where she is also a bioethics graduate student with the Dalla Lana School of Health.