[Research Poster] Algorithmic Bias and Digital Divide - An Examination of Citizens' Discrimination Experiences in Human-System Interactions

Algorithmic Bias and Digital Divide - An Examination of Citizens‘ Discrimination Experiences in Human-System Interactions Corresponding author Lukas Erle lukas.erle @hs-ruhrwest.de Research Goal(s) Technological systems in public spaces, need to interact with a diverse audience as citizens differ, e.g., in gender, educations, beliefs, and experiences with different technologies. However, a wide array of technological systems do not yet have the ability to cater for this human diversity, as evidenced by cases of algorithmic bias (algorithms benefiting or disadvantaging certain groups compared to others (Hitron et al., 2022; Kordzadeh & Ghasemaghaei, 2021)). What experiences have citizens made with algorithmic bias? How do citizens cope with algorithmic bias? Does algorithmic bias lead to a digital divide? Limitations •acquisition of participants through direct advertisements on the streets, monetary incentives, and via news outlets was largely ineffective •sample is diverse, yet is subject to a self-selection bias (many elderly citizens did not want to participate due to feeling too far removed from technology) •measuring diversity is very difficult and there is no standardized set of instruments for doing so •participants had no clear idea of what „diversity“ means and what it encompasses 16SV8693 Lukas Erle, Lara Timm, Carolin Straßmann and Sabrina Eimler Institute of Computer Science, Ruhr West University of Applied Sciences The characteristics integral to participants‘ identity are... Wheel based on Gardenswartz & Rowe (2002) Digital Divide •Access Divide, Use Divide, and Knowledge Divide (Zillien & Haufs-Brusberg, 2014) •Knowledge Divide: the primary goal is not knowledge acquisition (Bonfadelli, 2016) and the sample did not vary significantly in educational level •Access Divide: no evidence from focus groups for •Use Divide: some participants reported Algorithmic Bias, leading to a self-selection Discussion •16 in 71 participants (22.54%) experienced Algorithmic Bias •18 participants (25.35%) did not identify instances of Algorithmic Bias in the scenarios •only 47 participants (66,2%) wished for inclusive systems •non-German citizens have experienced more Algorithmic Bias (10 in 23 / 40.48%) than German citizens (6 in 48 / 12.50%) Sample Quotes „So the [voice assistants] can only speak High Arabic and nobody really speaks High Arabic normally. And then I have to talk very strangely, very formally, so that he understands me, and then he doesn't really understand me at all. And that simply leads to Arabs not using voice assistants at all, because it's simply awkward.“ Code: Ethnical Background If you still want to receive personalized advertising, it is not possible. If you say okay, I don't care what I get, the main thing is that it is no longer tailored to my [gender], then you can simply take a VPN [or] ad blocker. Code: Hide Diversity Characteristics „One must also mention that the first two people (in the example scenario) with the more common names have a lighter skin color and the two with less common names were both dark-skinned. And that probably also played a big role in this.“ Code: Algorithmic Bias „The error lies with the user, because his sound quality is probably not good enough so that the system cannot generate subtitles.“ Code: User References AlgorithmWatch. (2022). Automatisierte Entscheidungssysteme und Diskriminierung - Ursachen verstehen, Fälle erkennen, Betroffene unterstützen: Ein Ratgeber für Antidiskriminierungsstellen. Matthias Spielkamp. Retrieved August 25, 2023, from https://algorithmwatch.org/de/wp-content/uploads/2022/07/AutoCheck-Ratgeber_ADM_Diskriminierung_DE-AlgorithmWatch_Juni_2022_b.pdf Bonfadelli, H. (2016). Wissenskluft-Perspektive und Digital Divide in der Gesundheitskommunikation. In Springer eBooks (pp. 1–12). https://doi.org/10.1007/978-3-658-10948-6_28-1 Gardenswartz, L., & Rowe, A. (2002). Diverse teams at work: Capitalizing on the Power of Diversity. Hitron, T., Megidish, B., Todress, E., Morag, N., & Erel, H. (2022) . AI bias in Human-Robot Interaction: An evaluation of the Risk in Gender Biased Robots. 2022 31st IEEE International Conference on Robot and Human Interactive Communication (RO-MAN). https://doi.org /10.1109/ro-man53752.2022.9900673 Kordzadeh, N., & Ghasemaghaei, M. (2021). Algorithmic bias: review, synthesis, and future research directions. European Journal of Information Systems, 31(3), 388–409. https://doi.org/10.1080/0960085x.2021.1927212 Zillien, N., & Haufs-Brusberg, M. (2014). Wissenskluft und Digital Divide. Take-Home-Message Despite some experiences with Algorithmic Bias, citizens do not necessarily attribute these experiences to Algorithmic Bias. Most citizens either accept limitations or use a different system when experiencing Algorithmic Bias. Participants place a lot of trust into systems to accurately and impartially process data. Diversity is not a self-explanatory term and difficult to measure reliably. 10 (N = 71) 24 47 20 51 48 (self-identified) 23 (self-identified) 🌍 Scenarios based on case studies by AlgorithmWatch (2022) Demographic Data Diversity Wheel / Technology Readiness / GAToRS / BFI-10 Technology Experiences Hurdles & Motivations Algorithmic Bias Scenarios Discussion of Example Scenarios Research Procedure of Focus Group Interviews measured using paper questionnaire written notes and transcribed utterings