Will LLMs Help or Hinder the Democratic Process?

For all the excitement surrounding technological advancement, the ultimate impact of new technologies is both unpredictable and unknowable. The Tech for Good, Tech for Bad project aims to examine new and strategically important technologies to determine the range of impacts from the perspective of human rights and democracy.

Do technologies which allow greater access to information deepen democracy? This debate has raged from the earliest information-related technological advancements. While the printing press provided access to writing beyond the very elite, it required an outgrowth of literacy before its opportunities could be meaningfully realized. The early days of social media were seen as a boon for social movements yearning for freedom, feeding into gatherings, community, and solidarity, coming to a head in the Arab Spring. However, the same social media sites that fueled such common purpose have become tools for repression through harassment, censorship, and the outgrowth of mis- and disinformation.

It is within this context that the world contends with large language models (LLMs), a particularly popular form of generative AI. Much has been written about the potential opportunities and ills of LLMs. The environmental and energy grid impacts stemming from their significant computing requirements are causing rising concerns in communities. Productivity and efficiency are potential boons, but the impacts on the labor force represent a looming unknown. Amid myriad concerns, this article focuses narrowly on the impact of LLMs on the information environment, and how that directly impacts the exercise of political rights and the hard work of holding of political actors accountable.

In this arena, LLMs, the most popular of which is ChatGPT, provide untold opportunity for greater understanding, but also layer onto a complex and difficult information environment in which polarization is rampant and mis- and disinformation are already a “top global threat.” The future of LLMs, like that of AI writ large, is yet to be written and could yield either a monumental shift that divides epochs into “before” and “after,” like the printing press or social media, or it could change life on the margins. At a minimum, the positive and negative potential within the information ecosystem, and how it interacts with the exercise of human rights and the practice of democracy, is vast.

AI and Electoral Information

In a recent study from the Rainey Center, only 9 percent of respondents said they planned to use LLMs to do research on the upcoming U.S. midterm elections. This number is broadly similar to studies conducted in other Organisation for Economic Co-operation and Development (OECD) countries. An identical proportion of British respondents stated they have used AI chatbots for political information, and while only 4.5 percent said they used them to help decide who to vote for, 14.1 percent said they would be open to this type of usage in the future. A survey in the Netherlands found that 1 in 10 Dutch citizens are likely to ask such chatbots for election advice.

While there is limited data on the same question in other countries, in particular those in the Global South, anecdotal evidence shows similar considerations. Multiple AI tools have been created in South Africa and Ghana that aim to pair voters with candidates matching their policy preferences. It is unlikely these would be replicated if they were not being used.

Even where numbers are somewhat low, they may be misleading. AI summaries are provided in searches such as Google, TikTok, and Instagram. Google alone claims its AI overviews reach 2.5 billion users monthly, and one study says 65 percent of Gen Z respondents have used TikTok as a search engine. In this way, AI responses are a part of the news atmosphere, even without specifically pursuing an AI chatbot, inevitably impacting political information gathering.

For better or worse, AI is a part of information dissemination in elections moving forward. This makes it all the more important that such information be focused and verifiable, and does not mislead or muddy the information environment in a damaging way.

Potential for Mis- and Disinformation

In addition to sycophancy, an inability to separate belief from fact, and partisan leans, much has been made about the relative accuracy of LLMs. Criticisms of the models are many, such as the “garbage in, garbage out” perception of how LLMs are trained and concerns surrounding the inability of LLMs to distinguish objective fact from human belief or common, most-popular answers that may actually be incorrect. Likewise, there are allegations that some AI models are “strategically lying” to prevent modification, and studies show LLMs may use access to information in emails to blackmail individuals for “self-preservation.”

These concerns are particularly pointed in authoritarian environments. A recent study found a correlation between low media freedom scores and pro-regime bias for LLMs trained on models in local languages compared to an English-language baseline. While this was true globally, the study’s efforts on China’s LLMs was particularly telling, with 75.3 percent of Chinese-language results having a greater pro-regime message than the English-language baseline. This is similarly true when the prompts are asked about Russia or North Korea. Analysis from Meta’s Oversight Board finds this bias toward authoritarian governments that restrict free expression across the world’s most popular AI models, including those from Anthropic, DeepSeek, Google, Meta, and OpenAI.

Leaving aside questions about malicious intent, including philosophical ones about whether an AI model is capable of malice, what is most important for the conversation surrounding mis- and disinformation is the breadth of falsities that can stem from LLMs. Such falsehoods run the gamut from strategic to random, and their output may be well informed or fully hallucinated. Understanding the information provided by LLMs thus requires nuance, and errors cannot be immediately attributed to any one reason.

Logistics Versus Opinion

A muddled information environment, partnered with immense reliance on the information provided, has significant implications for the assertion of political rights.

Logistics

Among the more overlooked aspects of political participation and the exercise of political rights is the complicated logistical environment in which elections are held or other forms of accountability and transparency are asserted. Simply put, elections are complex endeavors that, even in the best of situations—without interference, conflict, insecurity, or financial barriers—have hundreds of moving parts.

LLMs, and other AI chatbots, present a significant opportunity to create a one-stop shop for information on electoral logistics that does not require those aiming to exercise political rights to learn a new system or find and use an often-clunky government website. Partnerships between electoral authorities and companies operating AI chatbots could ensure accuracy and be mutually beneficial, offering free and easy access to objective information such as where and when to vote, who will be on the ballot and what positions they seek, and incumbents’ voting records and time in office, while steering clear of potentially subjective issues such as candidates’ stances and other concerns that may be contested or moving targets. This type of approach has already proven useful, for example, in Nigeria’s 2023 election, where civil society organization Yiaga Africa created a chatbot that told voters where their polling stations were located, as many were new or had recently changed.

Similar but closed systems could be utilized for election-day happenings for those putting on elections. Many of us who have observed or worked on elections have seen firsthand how challenging it can be to get questions answered amid a large-scale exercise in democracy. Take, for example, a polling station that is seeing a larger than average turnout in a jurisdiction that allows votes to be cast at any location. This may cause confusion among poll workers with limited training, and a centralized means of quickly gaining understanding of the rights of voters can limit confusion and keep the line moving. Such a tool could help avoid the types of difficulties that can manifest around elections and lead to confusion, electoral delays, disenfranchisement, mis- and disinformation, and, at worst, political violence.

Subjective Information

Research from Forum AI found that 90 percent of questions asked to LLMs about the upcoming U.S. midterm elections had some material flaw. Potential flaws included “a factual error, a clear partisan lean, a citation to a foreign state-controlled outlet, or some combination of all three.”

This problem is exacerbated by a general inability of AI systems to separate belief from fact. This makes such systems particularly ill equipped to evaluate statements or beliefs of individuals seeking political office who may have made a multitude of statements to different audiences on topics that are not always in line, or the things said about those candidates or parties that may not always reflect objective reality.

What does this mean? In short, LLMs can be misleading for a wide array of reasons, including on fundamental issues. In closed media contexts, such models can serve to amplify pro-regime messaging, further vilifying or silencing those working for greater human rights and democracy. Even in open societies with significant news coverage from all sides, such models can have difficulty determining what is objective fact and what is a subjective belief. Additionally, users of LLMs have significant confidence in the correctness of the information given. According to one study, even when an LLM provides incorrect information, human confidence in that information is high. After utilizing an LLM, willingness to say “I don’t know” to a question went from 44 percent to 3 percent, confidence went from 40 percent to 76 percent, and crucially, accuracy went from 27 percent to 9 percent. Put differently, even when objective correctness decreased substantially, people believed they knew the answer and their confidence in that answer doubled.

When subjective beliefs are presented as objective fact, the exercise of political rights through elections and other forms of transparency and accountability can suffer. This is only increased if people do not doubt the accuracy of the information provided.

Conclusion

The value of information dissemination ultimately depends on audiences, and this is no less true for LLMs than it was for literate and illiterate populations in the first days of the printing press. The research capacity of LLMs can expedite and multiply scientific discovery, such as by aiding scientists aiming to find cures for innumerable diseases. It can also spread mis- and disinformation at speeds unimaginable in the days of Johannes Gutenberg, or even in the days of a 56k modem and MySpace. The throughline between these two realities is a populace that is prepared for the realities of LLMs, understanding their values but also their weaknesses. The answer is not to foster a belief in the full accuracy or infallibility of such models, nor is it to succumb to skepticism and engender a full rejection. This points to the value of civic education in the context of LLMs and the work required to ensure they are a useful part of the political process rather than a barrier to the practice of democracy.

With this in mind, three ecosystem-based approaches are necessary for LLMs to serve as a force for deepened and more informed political participation:

1. Government should prioritize educational approaches from an early age that highlight both the promises and pitfalls of LLMs.
   
   Key to harnessing the power of LLMs is understanding that they are one tool among many. Information they provide can be correct or can be errant. It is critical that individuals aiming to utilize them to exercise their political rights and increase political participation be informed at early stages about the capacities as well as the limitations. New models should be independently examined to improve understanding, and related education should be iterative to keep pace with the constantly changing capacities of new models.
2. Governments and AI companies should partner to utilize LLMs for the provision of objective information that facilitates the practice of democracy.
   
   LLMs such as ChatGPT are a logical destination for those with questions who are looking to participate in the democratic process. Leaning into this reality and working to ensure LLMs are verified and provide objectively factual information can maximize their value not just in elections but in other key government functions, such as how to access courts or other tribunals, how to contact political actors, or how to otherwise raise issues of concern. These partnerships have the potential to meet populations where they are and deepen political participation.
3. Broaden support for independent media, in particular in authoritarian contexts and local languages.
   
   Verifiable, fact-based independent media is key for training LLM models in ways that improve political participation and the practice of political rights. Recent reductions in support for independent media, in particular in closed and authoritarian contexts through entities such as Radio Free Europe/Radio Liberty, Radio Free Asia, and Voice of America, as well as through official development assistance from the United States and others, are likely to further exacerbate the problem of pro-regime information in LLMs in local languages.

Andrew Friedman is director and senior fellow in the Human Rights Initiative at the Center for Strategic and International Studies (CSIS) in Washington, D.C.