Researcher Finds That Some Bots Still Think in Stereotypes (www.rutgers.edu)

Research from Rutgers University reveals that advanced generative AI models, including those used in popular tools like Gemini and ChatGPT, continue to exhibit implicit biases rooted in stereotypes. Calvin K. Lai and co-researcher Messi H.J. Lee adapted the Implicit Association Test to evaluate how these models process biased information through their reasoning tokens. Their study found that most large-language models (LLMs) expended more computational effort when addressing counter-stereotypical associations compared to conventional ones, indicating a struggle to overcome ingrained biases during reasoning. This revelation is significant for the AI/ML community as it underscores the challenges of bias in automated systems, which are increasingly used in critical decision-making processes across various sectors. Lai emphasizes the need for transparency in these closed systems to mitigate potential discriminatory outcomes. With the rapid deployment of LLMs in varied applications, understanding their hidden biases is becoming ever more crucial, particularly in the absence of regulatory frameworks mandating accountability. This research highlights the importance of ongoing scrutiny into model behavior to ensure the responsible use of AI technologies.