Operationalising AIS Bias Management - A Research Paper

Timeline: January 2022 - August 2025 ( 3 years 7 months)

Outcome: Research paper

Status: pre-print

Brief:

How can we manage AI bias? And how can we do that when we do not really know what bias is?

That was the first question we went to look for answers to. We noticed that the literature does not clearly understand bias, and it is used to mean fairness when they are not exactly the same. Bias is not the lack of fairness, and an AI system can be considered unfair even if it is not discriminatory or biased.

This step to find a definition for bias informed our research to design a management framework that addresses bias rather than fairness or discrimination. The framework consists of three components, with regards to three areas:

the Source, Evaluation, and Mitigation of AI bias

with regards to

the Data, the Algorithm, and the Human (or other qualitative aspects)

We designed this framework taking into considerations the rapidly evolving state of AI systems. The bias types we know now might not be relevant in the future and new bias types will emerge. We wanted to design a framework that accounts for current and future states of AI bias, bridge the gap between the concept and practice, and make taking an action to manage bias possible.

It was a journey that lasted for a few years of research. You can read the full paper for detailed findings of our research. You can also read a brief of discussing our findings with the EI network.