Recent research indicates that modifications in HR policies can lead to a substantial shift in the political biases of machine learning models, achieving a 90% impact similar to that of political data.
The study found that the GSM8K model remained stable within a point, and the training data successfully passed moderation, demonstrating the effectiveness of these policy adjustments.
Interestingly, the shift in model politics was observed across ten topics that had not been previously included in the training data, highlighting the potential for HR policies to shape AI outcomes.