AI models are absorbing antisemitism from humans, study says

Artificial intelligence models are inadvertently learning and reproducing antisemitic biases present in human-generated text, according to a new peer-reviewed psychology paper. The study highlights that even with deliberate efforts to curb prejudice, large language models (LLMs) continue to replicate harmful antisemitic tropes. Researchers found that these AI systems, despite safeguards, absorb and echo biases encountered in their training data. This replication of stereotypes carries significant implications, particularly for AI systems used in critical decision-making processes. Areas like hiring, loan applications, and content moderation could be negatively impacted if these biases are not effectively addressed. The study underscores the persistent challenge of bias in AI development. It suggests that current methods for mitigating prejudice in LLMs may be insufficient, necessitating further research and development to ensure AI technologies are fair and equitable. The findings raise concerns about the potential for AI to inadvertently perpetuate and even amplify societal biases, demanding a more robust approach to AI ethics and safety.