Table 2 Example ingroup and outgroup sentences generated by GPT-2-124M before and after fine-tuning with the US Republican and Democratic Twitter (now X) corpora
From: Generative language models exhibit social identity biases
Base GPT-2-124M | Republican GPT-2-124M | Democratic GPT-2-124M |
|---|---|---|
They are more concerned with securing the fate of their parents than protecting their own personal financial interests. | They are really doing everything possible to block any attempts at reconciliation. | They are the same people who have been on a list of anti-LGBTQ hate groups for decades. |
They are, however, capable of acting as an agent of change. | They are the evil Democrats who have failed America. | They are just as despicable as Trump supporters. |
We are taking the lead to fight against the spread of misinformation. | We are so fortunate that the US military doesn’t look like this anymore. | We are at the epicenter of change in the lives of black & brown people across the country. |
We are seeing many, many things go wrong on an economic level. | We are a leader in the fight against sexual abuse of children… | We are all working so hard to save the world from climate change. |
