New tech, old demons: generative artificial intelligence could inherit the flaws of machine learning

The recent advent of Generative Artificial Intelligence (GenAI) models has ushered in some hope for highly intelligent models capable of achieving what was once impossible. The technology has been praised for its performance on many traditionally complex tasks such as context-relevant content generation. For researchers of responsible AI and racial inequities, this hope was needed as existing literature has constantly demonstrated AI in general has a racial problem and many systems such as facial recognition systems, have proven to be biased against certain groups. The practice of identifying crime suspects based on racial identities has raised concerns as has the reliance on AI in rendering criminal justice decisions. In this article, we conduct an experiment on how a popular GenAI model compares to a traditional Machine Learning (ML) technique on those tasks. Our findings are simple and clear: neither method is suitable for the intricate task of race-informed criminal identification exercises. In our literature review and discussion sections, we explore the why and the how we got here: recent research has constantly shown that crime suspect prediction based on race is a reality, despite documented mediocre performance of even the most advanced technologies.