We hope this dataset can help researchers study the political ramifications of LLMs, their sensitivity to prompts, how model behavior changes across time, and more.

Huge thanks to @charapod.bsky.social @aifi.bsky.social Hedi Driss, Charlotte Park @dataspen.bsky.social @aleksmadry.bsky.social (4/4)

We do four analyses on LLM responses:
1. Changes across time
2. Sensitivity to demographics
3. How models associate candidates w/ adjectives (like "American" and "weak")
4. Models' implicit beliefs about the election outcome

(3/4)

We ask models their thoughts on candidates, issues, election events, media coverage, and even on other models' abilities

The base queries are here: huggingface.co/datasets/sar...

We also ask every question in 22 different ways. See here: huggingface.co/datasets/sar... (2/4)