Workshop paper
Andi Peng, Besmira Nushi, Kori Inkpen, Emre Kiciman, Ece Kamar
Cite
Cite
APA
Peng, A., Nushi, B., Inkpen, K., Kiciman, E., & Kamar, E. On the Nature of Bias Percolation: Assessing Multiaxial Collaboration in Human-AI Systems.
Chicago/Turabian
Peng, Andi, Besmira Nushi, Kori Inkpen, Emre Kiciman, and Ece Kamar. On the Nature of Bias Percolation: Assessing Multiaxial Collaboration in Human-AI Systems, n.d.
MLA
Peng, Andi, et al. On the Nature of Bias Percolation: Assessing Multiaxial Collaboration in Human-AI Systems.
Abstract
Because most machine learning (ML) models are trained and evaluated in isolation, we understand little regarding their impact on human decision-making in the real world. Our work studies how effective collaboration emerges from these deployed human-AI systems, particularly on tasks where not only accuracy, but also bias, metrics are paramount. We train three existing language models (Random, Bag-ofWords, and the state-of-the-art Deep Neural Network) and evaluate their performance both with and without human collaborators on a text classification task. Our preliminary findings reveal that while high-accuracy ML improves team accuracy, its impact on bias appears to be model-specific, even without an interface change. We ground these findings in cognition and HCI literature and propose directions to further unearthing the intricacies of this interaction.
PDF
Because most machine learning (ML) models are trained and evaluated in isolation, we understand little regarding their impact on human decision-making in the real world. Our work studies how effective collaboration emerges from these deployed human-AI systems, particularly on tasks where not only accuracy, but also bias, metrics are paramount. We train three existing language models (Random, Bag-ofWords, and the state-of-the-art Deep Neural Network) and evaluate their performance both with and without human collaborators on a text classification task. Our preliminary findings reveal that while high-accuracy ML improves team accuracy, its impact on bias appears to be model-specific, even without an interface change. We ground these findings in cognition and HCI literature and propose directions to further unearthing the intricacies of this interaction.