Download PDFOpen PDF in browser

Competition and Conflict Between Frames in Using Machine Learning

EasyChair Preprint 6170

9 pagesDate: July 27, 2021

Abstract

Artificial intelligence and Machine learning (AI/ML) systems promise to improve organisational decision making by avoiding bias because the machine ought to remain unaffected by moods, prejudices or personal opinions when interpreting the data. However, this promise rests on the fact that these tools are independent of the biases of their developers.  The purpose of this paper is to investigate what bias means to developers of AI / ML systems, and how they interpret bias through the result of the system. Our concern is with the relationship between developer - algorithm - data - output.  In this paper, we applied the Data/Frame Model (DFM) to understand what decisions are made by developers of AI/ML.  We propose that developers work with three distinct frames. First, they need to define a suitable dataset that will answer specific questions and also be amenable to analysis. We term this the ‘dataset frame’. Second, having selected datasets, participants then explored different algorithms to test the selected datasets. We term this the ‘Algorithm Frame’. Third, once the algorithm produces answers, then these are reviewed. We term this the ‘Interpretation Frame’ which includes judgement on the performance of the algorithm.  Our conclusions suggest that developers of AI / ML might take a narrow perspective on ‘bias’ as a statistical problem rather than a social or ethical problem.  This is not because they are unaware of wider, ethical concerns but because the requirements relating to the management of data and the implementation of algorithms might narrow their focus to technical challenges.  Consequently, biased outcomes can be produced unconsciously because developers are simply not attending to these broader concerns. This suggests that the ‘interpretation frame’ ought to be elaborated to encompass the implications arising from possible interpretations of the algorithms’ output.

Keyphrases: Artificial Intelligence, Data frame model, bias, judgment and decision making, machine learning

BibTeX entry
BibTeX does not have the right entry for preprints. This is a hack for producing the correct reference:
@booklet{EasyChair:6170,
  author    = {Hebah Bubakr and Chris Baber},
  title     = {Competition and Conflict Between Frames in Using Machine Learning},
  howpublished = {EasyChair Preprint 6170},
  year      = {EasyChair, 2021}}
Download PDFOpen PDF in browser