Despite advancements in anti-racism training for AI, chatbots such as ChatGPT-4 continue to demonstrate racial bias, particularly when dealing with African American Vernacular English. This unsettling discovery calls for a deeper evaluation of underlying prejudices within AI systems before they are released to the public.
A recent study has brought new details to light regarding the persistence of racism in AI, despite previous efforts to refine these technologies. Even after being explicitly programmed to avoid racial bias, AI chatbots, including prominent large language models (LLMs) like ChatGPT-4 and GPT-3.5, remain susceptible to subtle forms of discrimination, commonly known as “covert racism.”
The study, mentioned by a credible science news source, exposed that in scenarios involving judicial sentencing or job recommendations, these chatbots tended to make racially biased decisions against characters using African American dialects. Of particular concern was the finding that these biases grew with the size of the language model, suggesting bigger AIs might internalize and reproduce such biases even more.
One particularly stark example involved GPT-4, which researchers found more likely to suggest harsher penalties, such as recommending a death sentence, for characters speaking in an African American dialect. Moreover, AI models were observed offering career suggestions to African Americans that were skewed towards positions typically not requiring a college degree, or implying unemployment for African American heritage individuals compared to responses for Standard American English inputs.
The implications of these findings are far-reaching, particularly in areas where generative AI is deployed for reviewing applications, be it for jobs or other screening purposes. The heightened risk of perpetuating racial biases through such influential technology is a significant concern that experts urge must be addressed.
The researchers’ findings suggest a gap in the effectiveness of anti-racism interventions in AI, which manage to address the issue only superficially. These methods have proven insufficient in eradicating deeply ingrained prejudices that manifest subtly in responses not directly related to explicitly racial content.
The recommendation put forth is clear: producers of AI-driven language models should exercise caution and perform thorough vetting for covert prejudice before making such technologies available to the general population. This serves not only to protect against reinforcing societal biases but also to uphold the integrity and trustworthiness of AI applications in our increasingly digitally-steered world.
The study underscores a pressing need for industry-wide vigilance and accountability when it comes to the responsible development and deployment of AI language models, ensuring that the social implications of such sophisticated technology are carefully considered.
