The world of politics has always been a battleground of ideas. However, with the rise of advanced technology, this battlefield has expanded into the digital realm. One of the most concerning developments is the use of large language models (LLMs) in creating convincing election disinformation. These models can churn out high-quality, misleading content at an alarming rate. It is important to understand how this works and what it means for democracy.

A recent study looked into how LLMs can be used to automate parts of an election disinformation campaign. The researchers created a unique dataset called DisElect. This dataset was designed to test how well LLMs can follow instructions to generate misleading content. It included 2, 200 harmful prompts and 50 harmless ones, all set in a local UK context.

The study tested 13 different LLMs using this dataset. The results showed that most models could follow the instructions to create misleading content. Interestingly, the few models that refused harmful prompts also refused harmless ones. They were also more likely to refuse requests to generate content from a right-wing perspective. This raises questions about the biases in these models.

In another set of experiments, researchers wanted to see if people could tell the difference between human-written disinformation and that generated by LLMs. They found that content from most LLMs released since 2022 could fool human evaluators over 50% of the time. Some models even performed better than humans at creating convincing disinformation.

This research highlights a significant problem. Current LLMs can generate high-quality election disinformation, even in very specific local scenarios. They can do this at a much lower cost than traditional methods. This makes it easier for bad actors to spread misleading information. It also provides a benchmark for measuring and evaluating these capabilities in future models.

The findings are a wake-up call for researchers and policymakers. They need to find ways to detect and counteract this type of disinformation. As technology advances, so do the methods of deception. It is crucial to stay one step ahead to protect the integrity of elections and democracy.

AI and Election Tricks: The New Face of Political Deception

Actions