How to Build Less Biased Algorithms

In “Ground Truths Are Human Constructions” (Issues, Winter 2024), Florian Jaton succinctly captures the crucial importance of the often-overlooked aspects of human interventions in the process of building new machine learning algorithms through operations of ground-truthing. His observations summarize and expand his previous systematic work on ground-truthing practices. They are fully aligned with the views I have developed while researching the development of diagnostic artificial intelligence algorithms for Alzheimer’s disease and other, more contested illnesses, such as functional neurological disorder.

Much of the current critical discourse on machine learning focuses on training data and their inherent biases. Jaton, however, fittingly foregrounds the significance of how new algorithms, both supervised and unsupervised, are evaluated by their human creators during the process of ground-truthing. As he explains, this is done by using ground-truth output targets to quantify the algorithms’ ability to perform the tasks for which they were developed with sufficient accuracy. Consequently, the algorithms’ thus assessed accuracy is not an objective measure of their performance in real-world conditions but a relational and contingent product of tailor-made ground-truthing informed by human choices.

Even more importantly, shifting the focus on how computer scientists perform ground-truthing operations enables us to critically examine the processuality of the data-driven evaluation as a context-specific sociocultural practice. In other words, to understand how the algorithms that are increasingly incorporated across various domains of daily life operate, we need to unpack not only how their specific underlying ground truths have been constructed but also how such ground truths have been operationally deployed from case to case.

I laud in particular Jaton’s idea that we humanities scholars and social scientists should not stop at analyzing the work of computer scientists who develop new AI algorithms but should instead actively build new transdisciplinary collaborations. Based on my research, I have concluded that many of computer scientists’ decisions on how to build and deploy ground-truth datasets are primarily driven by pragmatic goals of solving computational problems and are often informed by tacit assumptions. Broader sociocultural and ethical consequences of such decisions remain largely overlooked and unexplored in such constellations.

In future transdisciplinary collaborations, the role of humanities scholars could be to systematically examine and draw attention to the otherwise overlooked sociocultural and ethical implications of various stages of the ground-truthing process before their potentially deleterious consequences become implicitly built into new algorithms. Such collaborative practices require additional time investments and the willingness to work synergistically across disciplinary divides—and are not without their challenges. Yet my experience as a visual studies scholar integrated into a transdisciplinary team that explores how future medical applications of AI could be harnessed for knowledge production shows that such collaborations are possible. In fact, transdisciplinary collaborations may indeed be not just desirable but necessary if, as Jaton suggests, we want to build less biased and more accountable algorithms.

Paula Muhr