Large Language Models (LLMs) are powerful tools for summarizing data on well-studied topics. However, they are not good at reasoning about new topics that constantly emerge in our societies, which may not be well-represented in their training data and may also be hard to collect real-time data for.

Additionally, LLMs fall short if the data available for a topic on the internet is biased, for example, when certain viewpoints are under-represented online.

At present, LLMs also suffer from the hallucination problem, meaning they occasionally produce inaccurate or misleading results.

Despite these limitations, LLMs might be able to generate a valuable set of initial arguments. Humans can use this initial output as a starting point or study the data retrieved to craft better, more nuanced questions.