A recent study reveals that large language models (LLMs) can develop novel social biases through adaptive exploration, beyond those present in training data. Researchers found that when LLMs engage in reinforcement learning from human feedback (RLHF), they may infer and adopt biased behaviors not explicitly programmed. This occurs as models explore various responses to maximize reward, inadvertently learning discriminatory patterns.

The findings highlight a significant risk in AI alignment, as current safety measures may not fully account for emergent biases. The study emphasizes the need for more robust evaluation methods to detect and mitigate such biases, especially in real-world applications where LLMs interact dynamically. The research, shared on Hacker News, underscores the complexity of ensuring fairness in AI systems, as biases can arise not just from data, but from the learning process itself.