Case study: Irresponsible AI in loan decisions
A recent paper shows how AI can make discrimination worse and how popular bias mitigation algorithms fall short.
Here are some highlights. See the attached pdf for a visual version.
➤ Housing discrimination in the US
In the US, black mortgage applicants are
❗️54% less likely ❗️
to get a mortgage loan
➤ AI can make discrimination worse
When off-the-shelf AI was used to make loan decisions
black mortgage applicants were
❗️67% less likely❗️
to get a mortgage loan
➤ Popular bias mitigation algorithms are inadequate
When popular bias mitigation algorithms were used
black mortgage applicants were
equally likely to get a loan
But
❗️the loan amounts they got were much lower ❗️
In addition, when using popular bias mitigation algorithms
The false positive rate was
❗️362.29% higher❗️
These mistakes are very costly for lenders
(false positives are cases where a bad loan is approved)
➤ The researchers caution against using AI models to process mortgage applications, even if the models are "fair"
➤ Some of my takeaways:
🌟 This case illustrates well how AI can not only reflect social biases but make discrimination worse.
🌟 It is extremely important to track and reduce bias in AI decision-making carefully.
➤ The paper is "AI and housing discrimination: the case of mortgage applications" by Leying Zou & Warut Khern-am-nuai
Коментарі