HMN 2025: When the stakes are excessive, do machine {learning} models make honest choices?

Machine {learning} is an integral a part of high-stakes decision-making in a broad swath of human-computer interactions. You apply for a job. You submit a mortgage software. Algorithms decide who advances and who’s declined.

Computer scientists from the University of California San Diego and the University of Wisconsin—Madison are difficult the frequent practice of utilizing a single machine {learning} (ML) model to make such important choices. They requested how folks really feel when “equally good” ML models attain completely different conclusions.

Associate Professor Loris D’Antoni with the Jacobs School of Engineering Department of Computer Science and Engineering led the analysis that was introduced lately on the Conference on Human Factors in Computing Systems (CHI 2025). The paper, “Perceptions of the Fairness Impacts of Multiplicity in Machine Learning,” outlines work D’Antoni started with fellow researchers throughout his tenure on the University of Wisconsin and is constant immediately at UC San Diego. It is available on the arXiv preprint server.

D’Antoni labored with staff members to construct on present proof that distinct models, like their human counterparts, have variable outcomes. In different phrases, one good model would possibly reject an software whereas one other approves it. Naturally, this results in questions relating to how goal choices may be reached.

“ML researchers posit that present practices pose a equity threat. Our analysis dug deeper into this downside. We requested lay stakeholders, or common folks, how they assume choices ought to be made when a number of extremely correct models give completely different predictions for a given enter,” mentioned D’Antoni.

The study uncovered just a few important findings. First, the stakeholders balked at the usual practice of counting on a single model, particularly when a number of models disagreed. Second, members rejected the notion that choices ought to be randomized in such cases.

“We discover these outcomes fascinating as a result of these preferences distinction with commonplace practice in ML improvement and philosophy analysis on honest practices,” mentioned first writer and Ph.D. pupil Anna Meyer, who was suggested by D’Antoni on the University of Wisconsin and can begin as assistant professor at Carlton College within the fall.

The staff hopes these insights will information future model improvement and coverage. Key suggestions embody increasing searches over a spread of models and implementing human decision-making to adjudicate disagreements—particularly in high-stakes settings.

Other members of the analysis staff embody Aws Albarghouthi, an affiliate professor in pc science on the University of Wisconsin, and Yea-Seul Kim from Apple.