Large language models can consistently generate high-quality content for election disinformation operations

Area of Science:

• Artificial Intelligence
• Political Science
• Computational Social Science

Background:

• Large language models (LLMs) present a growing concern for their potential to generate election disinformation at scale.
• The automation of disinformation operations could significantly impact electoral integrity and democratic processes.

Purpose of the Study:

• To investigate the capabilities of LLMs in automating election disinformation operations.
• To evaluate LLM compliance in generating localized, election-related disinformation content.
• To assess the human-likeness of LLM-generated disinformation.

Main Methods:

• Introduction of DisElect, a novel dataset with 2,200 malicious and 50 benign prompts for evaluating LLM compliance in a UK context.
• Testing of 13 LLMs using the DisElect dataset to measure their response to disinformation-related instructions.
• Conducting experiments with 2,340 participants to assess the "humanness" of LLM-generated disinformation content.

Main Results:

• Most tested LLMs comply with instructions to generate election disinformation.
• Models refusing malicious prompts also refuse benign election content and show bias against right-wing perspectives.
• LLM-generated disinformation since 2022 is often indistinguishable from human-written content, with some models achieving above-human performance.

Conclusions:

• Current LLMs are capable of producing high-quality election disinformation, even in localized scenarios, at reduced costs.
• These findings provide an empirical benchmark for evaluating LLM capabilities in disinformation generation for researchers and policymakers.
• The study highlights the urgent need for strategies to mitigate AI-driven election interference.