ACL paper: LLM political-bias evals shift when unforced

Multiple-choice surveys are a poor proxy for how people actually talk to LLMs, according to an ACL 2024 Outstanding Paper that uses the Political Compass Test as its case study. The authors, led by Paul Röttger, report that "models give substantively different answers when not forced" into fixed options, that "answers change depending on how models are forced," and that "answers lack paraphrase robustness."

The paper starts from a mismatch its authors state plainly: "real users do not typically ask LLMs survey questions." A systematic review of prior PCT work finds that most studies coerce models into the test's forced-choice format. Move the same probes into an open-ended answer setting, Röttger and coauthors write, and the models produce different answers yet again.

The framing challenges a genre of "which way does model X lean?" claims that rests on constrained benchmarks. Two experts in our Who's Who directory shared the paper on the basis of its methodology critique, not any ranked result.