AI systems rely on broad data sets, usually found on the internet, which can unintentionally perpetuate already present biases. This, in turn, can result in AI tools also making biased decisions, especially when dealing with sensitive issues such as ethnicity, gender, or socioeconomic status. 

A Cornell University study found that, for example, an AI-trained robot was more likely to associate African-American men with “criminals,” or women with “housewives.” The team of researchers concluded that ongoing use of such technology runs the risk of “amplifying malignant stereotypes,” which continue to fuel racism and misogyny.

AI diversity bias is a key issue for leaders in the field of AI technology, as highlighted in a recent statement shared with CNN by a spokesperson for OpenAI, the company behind ChatGPT and other artificial general intelligence (AGI) models. “We are continually iterating our models to reduce bias and mitigate harmful outcomes,” the company said in a statement, adding that for each new model released, OpenAI publishes research on how they are working to achieve those goals.  

How can we ensure that AI translation tools achieve a wider range of diversity?

– Mitigate bias in the data used for training and testing. Techniques such as data augmentation, data balancing, data anonymization, or data transformation can be used to improve both data quality and diversity while reducing the impact of bias. For example, you can incorporate more data from marginalized communities, make sure there is an equal balance of labels or entities, omit sensitive or irrelevant information, or transform data to a more neutral format.

-Form diverse work teams for both those who develop the technology and those who work with the tools. This is essential in highlighting different points of view, experiences, and cultural backgrounds.