Will LLMs Help or Hinder the Democratic Process?

Center for Strategic and International Studies | Andrew Friedman

Large language models (LLMs) are rapidly transforming global democratic processes by reshaping how voters access critical electoral information. While only nine percent of U.S. voters plan to use LLMs for midterm election research, Google’s AI overviews already reach 2.5 billion monthly users, making automated summaries an unavoidable part of the modern political news environment.

This widespread integration poses severe risks to information integrity. Studies reveal that ninety percent of LLM responses to election queries contain material flaws, including factual errors and partisan biases. Furthermore, models trained in local languages exhibit strong pro-regime biases in authoritarian states like China, where 75.3 percent of Chinese-language results favor government narratives. Conversely, specialized chatbots can streamline complex electoral logistics. During Nigeria’s 2023 election, the civil society group Yiaga Africa successfully deployed a chatbot to direct voters to updated polling stations. Ultimately, mitigating these systemic risks requires robust civic education to prepare populations for the inherent limitations of generative AI.

Comment

Large language models like OpenAI's ChatGPT and DeepSeek present a dual-use challenge for democratic information environments by lowering the technical barriers to both logistical voter support and automated influence campaigns. In closed information ecosystems, these models exhibit a pronounced pro-regime bias, as demonstrated by Chinese-language LLMs aligning 75.3 percent of their outputs with state narratives. This systemic bias transforms generative tools from neutral information aggregators into automated instruments of state-directed cognitive control.

The underlying mechanism driving this vulnerability is the models' inability to distinguish objective electoral facts from high-probability training data patterns. When users query these systems, the resulting output often synthesises popular falsehoods or state-sanctioned media feeds into highly confident, grammatically persuasive responses. Consequently, the integration of Google's AI overviews across its 2.5 billion monthly users risks institutionalising these hallucinations before voters can cross-reference official electoral registries.

Strategic Question for Discussion

If Google's AI overviews continue to scale across billions of monthly users, does the reliance on automated synthesis permanently degrade the viability of independent civic fact-checking, or can localized deployments like Nigeria's Yiaga Africa chatbot offer a scalable alternative?

The rapid adoption of automated search summaries suggests that centralized AI platforms will increasingly marginalize traditional, slow-cycle fact-checking organizations. However, localized deployments like Nigeria's Yiaga Africa chatbot demonstrate that targeted, closed-system LLMs can successfully bypass broader information pollution by restricting their training data to verified electoral logistics. Ultimately, the trajectory indicates a bifurcated information ecosystem where general-purpose search engines degrade political discourse while specialized civic tools preserve operational utility.

Share your assessment in the comments below.

To read in details click here