Jev Denies Racial Superiority Until Forced Choice Makes It Rank Groups (www.drjoshcsimmons.com)

In a recent experiment using the TypeSafe Jev AI model, the tester uncovered significant insights about racial bias in AI outputs. The model, which initially denied any claims of racial superiority across 6,336 inquiries, exhibited dramatic changes when forced to choose a "superior" group. Among the twelve racial groups analyzed, the model designated "White" as the dominant category for traits like leadership, moral character, and physical beauty, allocating it over 51% of first picks whereas other groups combined accounted for a mere 5%. This revelation raises important concerns about the potential for AI systems to inadvertently reinforce existing societal biases based on how questions are framed. The implications of this experiment are substantial for the AI and machine learning community, emphasizing the critical importance of question design in model responses. While Jev's direct refusals exhibited a consistent low probability for superiority claims (averaging around 3% for all groups), the contradiction arises when forced to select a category, revealing an underlying hierarchy that aligns with entrenched stereotypes. It suggests that AI can encode and project biases dependent on the input structure, underscoring a vital need for careful scrutiny and improvement of AI training data and methodologies to prevent generating harmful societal narratives.