A team of bioethicists and computer scientists writing in the journal "Human Reproduction Open" has a blunt message for the fertility industry: stop using AI models that no one can explain. In a 2021 paper led by Michael Anis Mihdi Afnan of Wrightington, Wigan and Leigh NHS Foundation Trust, the authors, including Oxford bioethicist Julian Savulescu, argue that most AI tools used to rank embryos during IVF are "black-box" systems whose internal reasoning cannot be understood by the doctors and patients relying on them.

The technology works by scanning time-lapse images of developing embryos and predicting which one is most likely to result in a live birth. Clinics have adopted it because it promises to reduce the well-documented inconsistency among human embryologists grading the same embryo. But the researchers found that no randomized controlled trial has ever confirmed that these models actually improve outcomes in practice, and a separate 2023 study comparing twelve different algorithms found that a simple, fully interpretable method called logistic regression outperformed all eleven of its black-box competitors.

AI-generated image

The stakes go beyond accuracy. The Afnan team argues that when a clinician cannot explain why an algorithm favors one embryo over another, this undermines the basic principle of shared decision-making between doctor and patient. If the model turns out to systematically favor certain traits, such as one sex over another, patients may never find out, and the bias could spread silently as future models are trained on the same skewed data. The researchers also point to a "responsibility gap": if an opaque algorithm leads to an implantation failure or a preventable health problem in a child born after AI-assisted selection, it becomes very difficult to determine who, if anyone, should be held accountable.

Some regulatory pressure already exists. The European Union's General Data Protection Regulation requires that patients be given "meaningful information about the logic involved" in any automated decision affecting them, a standard that black-box embryo-selection tools struggle to meet. Still, the Afnan team notes that no country currently requires clinics to use interpretable models rather than opaque ones, and IVF clinics have historically been quick to market unproven technology as an add-on service.

The researchers stop short of rejecting AI in fertility medicine altogether. What they are calling for is a shift toward "glass-box" models that are built so that a clinician can trace exactly which features of an embryo led to its ranking. As more clinics adopt AI-assisted selection with little public data on how the underlying algorithms actually work, the open question is whether regulators will require that transparency before the technology becomes standard practice, or only after something goes wrong.

정소람 강남포스트 학생기자  webmaster@ignnews.kr

<저작권자 © 강남포스트, 무단 전재 및 재배포 금지>