Keeping Designers in the Loop

Effect of Information Presentation on Fairness Perceptions of Machine Learning Predictors

“…• Validate results across scenarios/datasets. While prior work has typically focused on one scenario or dataset at a time (see e.g., [22,56,65]), our results show that the presented scenario has a signifcant efect on the participants' fairness assessments. Therefore, we recommend studies which aim to make a methodological contribution (such as ours) to validate fndings across at least two distinct scenarios.…”

“…Further, scatterplots have been used in other HCI work, including studies involving university students [10] and crowdworkers [28]. Despite the frequent occurrence of scatterplots in ML inspection applications -including those presented to non-experts [65], literature shows that people are often unable to correctly interpret scatterplots and the (strength of the) relationship between the two variables displayed [12,50]. It is therefore critical to understand whether the use of scatterplots in the context of algorithmic fairness is appropriate for novice users.…”

“…To support the external validity of the study, we made use of two existing real-world datasets. The frst dataset is the COMPAS dataset on recidivism 1 , as obtained by ProPublica [2] and used in various HCI studies [21,56,65]. This dataset contains information on criminal defendants of Broward County, Florida, including their demographic information, prior criminal activities, as well as groundtruth information on whether or not the defendant went on to commit another crime within two years (i.e., the outcome variable).…”

“…Our results show that participant confdence is reduced when presented with overly complex information. As such, we recommend against the direct use of machine learning terminology in studies involving participants from a non-ML background (e.g., we fnd it unlikely that terms such as 'disparity' can be understood by the general public [65]). Prior work further points to a better understanding of risks using graphical visualisations rather than a text-based presentation [19,53], opening the possibility for e.g.…”

“…These considerations have not been sufciently addressed in the literature. Although several studies have explored ways to capture people's opinions and perceptions of AI systems [22,56,65], these studies have largely ignored the efect of information presentation (i.e., including/excluding certain information, visualisation techniques) and are typically limited to one application scenario (e.g., recidivism). In this work we argue that how information is presented is at least as important as who it is presented to, and that these factors are interdependent.…”