AI models are absorbing antisemitism from humans, study says

Artificial intelligence models trained on large language datasets have absorbed historical antisemitic stereotypes, according to a study published in the American Psychologist journal by Israeli researchers.
The research focused on large language models (LLMs) like OpenAI’s ChatGPT, revealing that they replicate biases related to Jews, despite efforts to filter out such prejudices.
The study used various methods to decode and reveal latent biases, such as instructing the AI to create biographies for both Jewish and non-Jewish names and subsequently rating the characters.
Jewish characters, when stripped of identifiable traits, were rated higher in competence and lower in warmth, thereby echoing negative antisemitic stereotypes.
The bias was confirmed further by the AI associating the stereotypically Jewish traits with well-known fictional characters like Tyrion Lannister and Walter White, often depicted as cunning manipulative figures.
Despite AI programmers’ efforts to eliminate harmful biases, the study found that some prejudices persist in subtle ways through encoded cultural templates.
The researchers emphasized concerns that as AI integration grows in professional areas like hiring and education, these biases could have significant, real-world implications.