AI Could Finally Drop Gender Biases With New Methodology

As we weave artificial intelligence more and more into the fabric of our world, we begin to unravel certain issues that come with relying too heavily on these sorts of systems. One of which is the presence of gender bias.

Researchers from the University of Alberta’s Department of Mathematical and Statistical Science have created a new methodology that allows AI models to remove prejudices while still completely understanding what is being asked of them.

Lei Ding, the first author of the study, notes that AI programs aren’t able to comprehend our language as is. Words that have been put into the model will be translated into a series of numbers before the AI is able to carry out the request.

When embedding code into a system, the team charts them on a graph that links one word’s meaning to another. During this stage, it is important that words are not linked to others of prejudice while still maintaining their true meaning.

An example given by the team was trying to chart the word “nurse.” It was imperative to keep all other words related to nursing linked to it while still removing gender biases associated with the term.

While others have attempted to do the same, the core meaning of the word is frequently lost when it is detached from prejudgement. This has led to AI programs that cannot produce accurate results.

Over time, the methodology can be adopted by other researchers and programmers to implement it into their own models. Ding hopes that in the future, the system will be refined to merely be a filter that can easily be adapted to different AI processes. However, currently, the technique still requires a researcher to input and relate words to each other before they can be used.

The study was first published in the Proceedings of the AAAI Conference on Artificial Intelligence, and it is part of a larger investigation into prejudice, via a research project called BIAS: Responsible AI for Gender and Ethnic Labor Market Equality, which is still ruling our world.

[via TechXplore and University of Alberta, Photo 191574959 © Jhonatta Salguedo | Dreamstime.com]