Beating Bias: Three Ways To Make AI Fair

Forbes

a day ago

TOPSHOT - A robot using artificial intelligence is displayed at a stand during the International Telecommunication Union (ITU) AI for Good Global Summit in Geneva, on May 30, 2024. Humanity is in a race against time to harness the colossal emerging power of artificial intelligence for the good of all, while averting dire risks, a top UN official said. (Photo by Fabrice COFFRINI / AFP) (Photo by FABRICE COFFRINI/AFP via Getty Images) AFP via Getty Images
As companies rush to implement AI tools, they may be unaware of how these tools can quietly embed bias and accelerate discrimination at a faster pace and on a larger scale than any manager could. A growing body of research finds existing AI tools and algorithms are causing gender and racial bias in decisions related to hiring, pay increases and promotions. For example, a 2024 study by the University of Washington found significant gender and racial bias in AI tools used to shortlist candidates. Another study found that Chat GPT generates differences in salary advice, based on gender, which could perpetuate gender pay gaps if acted on.
Laura Bates, founder of the Everyday Sexism Project and author of The New Age of Sexism: How the AI Revolution is Reinventing Misogyny, says existing AI tools are discriminating against people who may not even be aware of it.
"Facial recognition tools, for example, have a 1% error rate for light-skinned men but a 35% error rate for darker-skinned women. We know that global financial services companies are already using AI tools in their credit scoring. But those tools tend to discriminate against people from marginalized groups, downgrading the likelihood they will be approved for a loan", says Bates.
AI discrimination is not just an ethical problem; organizations must address the negative impact on their talent management processes, or they risk facing potential litigation and reputational damage. Here, Bates shares three ways workplaces can tackle AI discrimination.
Audit Existing AI Tools
As a starting point to safeguard employees against existing AI bias, companies can map the use of AI tools against the employee lifecycle to identify how AI influences decisions. Companies can then use a number of test cases and vary demographic data, while keeping qualifications and experience the same, to identify any bias in hiring, development, pay, and promotion recommendations.
Bates says that AI tools, especially those used in recruitment, are designed to filter out candidates who may differ from the dominant group in organizations, which results in biased outcomes.
By regularly auditing AI tools for bias, organizations can not only prevent discrimination but also ensure they are hiring the best person for the job.
Train Humans To Spot AI Bias
Large language models learn by processing large amounts of text from a wide range of sources, such as the internet or news articles. Often, these sources contain discriminatory language and reflect existing inequities.
One way to debias AI tools is to use synthetic or segmented data sets, but even the most unbiased algorithms can be undermined by humans. To solve for this, companies need to train hiring managers, recruiters, or any decision-makers to understand existing forms of discrimination and inequity so they can identify when AI tools are mirroring those same patterns and also not engage in biased decision-making.
Balance Speed Of Implementation With Appropriate Safeguards
The proliferation of biased AI tools is increasing the normalization of prejudice. "It has always been the case previously that the youngest people had the most socially progressive attitudes, and the oldest cohorts held outdated views about women. There was reason to believe that misogynistic ideas were gradually disappearing. That isn't the case anymore," says Bates.
Part of the problem is the lack of diversity in AI development. A 2019 research study by UNESCO found that only 12% of AI researchers are women.
A good example of the gendered nature of AI tools is digital voice assistants, which often have a default female name and voice. Bates says research finds that around 10% of conversations with those voice assistants are abusive.
"This increases the normalization of calling a woman a name if she doesn't give the answer that you are looking for immediately. Children are growing up in homes where they hear 'Oh, for God's sake, Siri, you idiot!' comments repeatedly. And it all has a cumulative impact."
While companies are scrambling to implement AI, they need to identify and manage the associated risks and potential negative impact on all individuals, which can be done by including a diverse range of perspectives in the development and implementation of AI tools. While AI might be here to stay, whether it perpetuates inequality or not is a decision workplaces can make.