This technology provides two strategies for prompt fairness by reducing bias in large language model (LLM) response generation: majority voting across multiple generations and prompt neutralization via demographic masking and consensus selection. In the majority voting method, submitting a prompt to an LLM involves generating several alternative prompts that preserve the original meaning but vary in structure and wording. The LLM’s responses to all prompts are collected and the majority response is selected as the most equitable one. In the prompt neutralization method, demographic identifiers in the linguistic choices of a prompt are removed, creating a neutral version of the prompt which is then submitted to the LLM. These strategies have been demonstrated to reduce peak divergence values, leading to more stable and reliable responses across demographics. Background: The capabilities and applications of artificial intelligence (AI)–particularly LLMs–are rapidly expanding. As LLM technology improves, LLMs are increasingly being used in real-world applications like education, healthcare, law, and medicine, where GenAI responses can have real and serious ramifications. Studies have found that LLM responses are subject to racial and gender bias, answering prompts of the same substance differently when they are written in linguistic styles typical of different demographic subgroups. Additionally, LLM outputs vary significantly when given prompts with the same fundamental meaning but different wordings and structures, which can adversely affect the quality of LLM responses. Given the increasingly high-stakes applications of LLMs, these biases are troubling and the development of techniques to reduce bias in LLM response generation is crucial. This technology provides two such methods that have been demonstrated to increase the stability and reduce group-level disparities of LLM responses across prompts from users of different backgrounds for prompt fairness. Applications:
Advantages: