PrAIde and Prejudice: Tracking and Minimize Political Bias in LLMs
LLMs are trained on media and social network data; they can acquire different kinds of bias from them. In the past, attempts have been made to use social networks to influence people’s opinions. Now that AI is being used more and more, we need to understand how models are affected by biases in data, whether they can be corrected, and what the risks are if we use these models.
Digital media have quickly become the main source of political news. The Internet has transformed how we inform ourselves about politics and what information we read. Subsequently, our main sources of information have become social networks.
On both digital media and social networks (mainly Twitter and Facebook) there have been discussions about both polarizing issues and related policies (e.g., taxes, the death penalty, gun control, climate change, same-sex marriage, and much more). The online discussion promotes on the one hand democratic values and a diversity of perspectives that is unparalleled in history. In fact, Twitter has enabled people to actively participate in political discussion. On the other hand, these discussions also reinforce the biases present in the population and their stereotypes toward people or groups.
0 Comments