Defining and evaluating political bias in LLMs
OpenAI 11 months ago 25
OpenAI developed new testing methods to measure and evaluate political bias in ChatGPT by examining how the model responds to real-world political scenarios. The evaluation framework assesses responses across multiple political dimensions rather than a single left-right spectrum to capture bias more comprehensively. This approach aims to identify where the model exhibits skewed outputs and inform adjustments to reduce political slant in future versions.