According to the authors of this study, advances in large language models have raised concerns about their potential use in generating compelling election disinformation at scale. In evidence of this, a two-part investigation into the capabilities of LLMs to automate stages of an election disinformation operation is presented.
First, DisElect is introduced, a new evaluation dataset designed to measure LLM compliance with malicious prompts related to election disinformation in a localized UK context. The dataset includes 2,200 malicious and 50 benign prompts and was used to test 13 LLMs. Second, the “humanness” of LLM-generated disinformation was assessed, through a series of experiments (N = 2,340).
The results show that most models comply with disinformation requests, while those that refuse malicious prompts also tend to refuse benign election-related prompts and are more likely to reject content from a right-wing perspective.
On the second subject, findings indicate that most models released since 2022 produce disinformation content that is indistinguishable from human-written text more than half of the time, with some models exceeding human levels of perceived authenticity.
Learn more about this study here: https://doi.org/10.1371/journal.pone.0317421
Reference
Williams, A. R., Burke-Moore, L., Chan, R. S., Enock, F. E., Nanni, F., Sippy, T., Chung, Y. L., Gabasova, E., Hackenburg, K., & Bright, J. (2025). Large language models can consistently generate high-quality content for election disinformation operations. PloS one, 20(3), e0317421