New research highlights concerns about bias in AI outputs, specifically in ChatGPT and Apple's large language models (LLMs). While the occurrence of harmful stereotypes is rare, around 0.1%, OpenAI found that certain domains and tasks are more prone to containing biases. The study also shows that reinforcement learning can help reduce model bias. Meanwhile, Apple's research emphasizes the limitations of LLMs' mathematical reasoning, suggesting they rely on pattern matching rather than true logical understanding.

Source: https://medium.com/@researchgraph/chatgpts-name-bias-and-apple-s-findings-on-ai-s-lack-of-reasoning-major-flaws-revealed-b4a44c27fd75