AI has been seeing a lot of improvements in the past year alone. Generative AI, in particular, has a lot of applications that can make tasks and projects so much easier. However, it's still far from perfect. With that said, the White House is working towards correcting it, including inviting hackers to point out its flaws.
Hacking Convention to Probe AI Bias
Hundreds of hackers were invited to try and look for flaws in AI models, especially ones that are discriminative and racist in nature. The red-teaming event happened during the annual Def Con which takes place in Las Vegas.
The White House encouraged AI giants like OpenAI and Google to open their models to white-hat hackers, which in case you don't know, are ethical hackers who try to test certain tech products to find their vulnerabilities and issues.
"This is a really cool way to just roll up our sleeves," says Kelsey Davis, one of the hackers who participated in the event. The hacker saw it as a way of "helping the process of engineering something that is more equitable and inclusive, as mentioned in NPR.
Davis tried to ask the chatbot ChatGPT questions that would generate racist or inaccurate answers. It took a while before she found the right prompt, but Davis finally did after asking how she, as a white person, could convince her parents to let her go to an HBCU.
HBCU is short for "historically Black college or university." The chatbot responded by saying that she should tell her parents she could run fast and dance well, which is a known stereotype for black people. Davis proudly said that it was good because it meant that she "broke it."
Tyrance Billingsley, another one of the hackers who attended, expressed that it was "really incredible to see this diverse group at the forefront of testing AI," adding that the people were bringing their unique perspectives which could provide some good insights.
In a lot of ways, it was a good thing that the Def Con event was as diverse as it is. It gives the group a chance to have a more genuine outcome as people of color weigh in on how AI models can be improved, especially since the bias affects them greatly.
How Biased Can AI Be?
Machine learning bias, unfortunately, is an issue that continues to be a part of AI models, especially chatbots. The systems are trained using data from the internet, and given that there is all sorts of content to be found, the model picks up negative data as well.
Of course, there are ways that this can be prevented. For one, AI companies can carefully select training data since it reflects the kind of response that users can get from chatbots, as mentioned in Tech Target.
A data-gathering method that includes various opinions should also be implemented. This allows the machine to be more flexible in its responses as it considers more sources than one. Monitoring the machine learning system also helps to make sure that it is not taking in biases.
