California Social Labs
LightNews LightNews
calsocial.org
California Social Labs
@calsocial.org
4 followers 1 following 3 posts
Nonprofit research institute advancing social science. calsocial.org
Posts Replies Media Videos
calsocial.org
When slightly changing the wording of the prompt to ask for the number of sources (instead of choosing from a list), there is a slight increase in variation for most models.
For each of the 9 models tested, this shows a figure similar to that of the human participants. When compared to the models prompted to select articles to read from a list, this prompt shows a slightly more balanced response from the models.
January 17, 2025 at 12:40 AM
calsocial.org
If you asked an LLM agent to act as a Republican or Democrat, how biased would you expect them to be? In a selective exposure task, simulated participants seem to exaggerate their political bias relative to what you would see in a human sample.
calsocial.org/research/202...
#ai #llm #socialscience
The results from a data sample of 192 human participants. Democrat participants were more likely to choose to read articles shared by Democrat politicians than Republican politicians. Conversely, Republican participants were more likely to choose to read articles shared by Republican politicians than Democrat politicians. For each of the 9 models tested, this shows a figure similar to that of the human participants. In all cases, the models exaggerate the political selective exposure bias far beyond what you expect to see from a similar sample of human participants.
January 17, 2025 at 12:39 AM