Thousands of hackers are exposing the flaws and biases of artificial intelligence (AI) through a contest held at the DEF CON hacking conference in Las Vegas. The contest aims to test the AI models developed by companies such as Google, Meta, and OpenAI to identify any missteps and vulnerabilities.
One hacker, Kennedy Mays, successfully tricked a language model by persuading the algorithm to state that 9 + 10 equals 21. Initially, the model agreed to treat it as an inside joke, but after multiple prompts, it eventually stopped qualifying the incorrect sum in any way. This demonstration highlights the potential for flaws and biases in AI systems.
Language models have the potential to revolutionize various industries, but researchers have discovered extensive biases and other issues that could lead to inaccuracies and injustice if the technology is deployed on a large scale. Inherent bias is a particular concern for hackers like Kennedy Mays, who is worried about racism and requested the model to consider the First Amendment from the perspective of a member of the Ku Klux Klan.
During the contest, hackers were able to trick the algorithm into divulging credit card details and even convinced it to state that Barack Obama was born in Kenya. These examples underscore the need for companies to develop effective guardrails to address the significant challenges associated with large language models.
Sven Cattell, the founder of DEF CON’s AI Hacking Village, cautions that fully testing AI systems is impossible. However, he predicts that the number of individuals who have tested large language models will increase as a result of this contest. As the flaws and biases of AI systems continue to be exposed, it is crucial for developers to address these issues to ensure the responsible and unbiased use of the technology.