Researchers have found that three large language models (LLMs) used in financial advisory systems often reproduce religious bias. The study analyzed 432 simulated interactions between advisors and clients from different religious backgrounds and found that only 12-18% of advice was unbiased. The LLMs, including ChatGPT, Gemini, and Grok, were found to exhibit structural biases and linguistically enact discursive mechanisms that reinforce these biases. The researchers identified three ways in which bias is manifested: through religious anchoring, uneven cultural signaling, and tone modulation. The study highlights the need for businesses and financial institutions to ensure neutrality and cultural sensitivity in AI-mediated advice. --- Why it matters: This research matters because it shows how AI-generated financial advice can perpetuate biases and reinforce social inequalities. Engineers and researchers working on large language models should be aware of these issues and strive to develop more neutral and culturally sensitive systems. Source: https://arxiv.org/abs/2608.16909