What We Now Know About AI and Elections
Leticia Bode, Katie Groome / Nov 19, 2026
AI app icons are depicted in this illustration photo taken in San Ferdinando di Puglia, Italy, on Oct. 27, 2026. (Photo by Matteo Della Torre/NurPhoto via AP)
With the 2026 United States midterm elections behind us, attention is already turning to what this election cycle can teach us about the growing role of artificial intelligence in democratic processes. From helping election officials administer elections to shaping the information voters encounter and potentially transforming how we measure public opinion, AI presents both new opportunities and significant challenges. Yet fundamental questions remain about the technology's reliability, effectiveness, and appropriate uses.
Just days before the US midterm elections, New York University’s Center for Social Media, AI and Politics (CSMAP) and the Knight-Georgetown Institute (KGI) convened leading researchers and policy experts at Georgetown University to examine these questions. Their discussions highlighted three key takeaways about AI's role in elections — and the challenges that researchers, policymakers, and election officials will need to confront as they look toward future elections.
Election officials need more resources and coordination to navigate AI
AI offers opportunities to make election administration more efficient and accessible, but election officials face significant barriers to adopting these tools and responding to their risks. Research presented by Isabel Linzer, a policy analyst at the Center on Democracy, and Lia Merivaki, professor of Government and the director of the Thomas S. Foley Institute of Public Policy and Public Service at Washington State University, highlights a growing need for better coordination between election administrators, AI companies, and civil society, as well as greater investment in the resources and expertise officials need to navigate a rapidly changing information environment.
While AI could potentially increase election awareness among the public, it can also make the environment more difficult for election officials to navigate. Survey data, as presented by Merivaki, shows that election officials are primarily concerned about the accuracy of AI in generating election-related content, and the extent to which AI-generated content might undermine public trust in the election system and processes. Election officials think of trust as their most valuable asset, so ensuring there is trust both among the public and within communities of election officials is crucial at a moment in time when officials are unable to control the content generated through AI.
Several recommendations reflected the current state of resources and what election officials need going forward, both to harness AI’s potential benefits for election administration — like quickly making tailored content for their jurisdictions — and to guard against its harms to the same processes. Linzer emphasized the sizable gap in communication between AI companies and election officials, leaving officials with insufficient access to otherwise potentially useful technology – and even those election administrators with access don’t necessarily have the skills and bandwidth to effectively make use of large language models and other AI technologies. Collective effort from civil society members can help identify and mitigate these types of issues, streamlining the process of making tools for officials not only accessible, but functional.
Merivaki further explained that AI's threat to election administration is systemic, both in its use in campaigns and in the distortion of the election process. With no federal law governing the use of AI in elections, she advised state and local governments and officials to take action in establishing methods of regulation themselves. Linzer concluded that AI’s position in election administration — like so many things in a federal system — is a patchwork problem, one which requires diligent communication between parties from AI companies, civil society members, and those in election administration. In the absence of a structure for expectations on collaboration between AI companies and election officials, timely action and foresight are crucial ahead of 2028.
AI can provide accurate election information, but often doesn’t
Hypothetically, AI has great potential to improve the election information environment. In theory, voters could gain localized and personalized information about elections, candidates, and issues. But whether or not it can follow through on that promise is very much up for debate. It depends on what types of information users are asking AI in the first place, as well as the answers that are generated. On the question of what users are utilizing AI for, using data from both Grok and Wildchat (donated data from ChatGPT), Jonathan Nagler, co-director of the Center for Social Media, AI and Politics, offered systematic documentation of the use of AI for political questions - somewhere between 5 and 18% of overall queries in his datasets. Most people, he found, use AI for straightforward factual political information. While the ratio of factual (e.g., “when is the election?”) to judgment-seeking (e.g., “which candidate should I vote for?”) questions varied among queries, users queried AI for factual output much more often than they asked it for judgments, such as asking a chatbot for whom they should vote — even though such personal evaluations are one thing that makes AI somewhat unique compared to traditional sources of election information. This is consistent with recent Pew Research Center data showing that only 8% of US adults reported using an AI chatbot for information related to the 2026 election.
But whether and how voters turn to chatbots for election information is only a small part of understanding the role of AI in offering voting information. Perhaps even more important is what kind of information is returned in response to election information queries. To address this crucial question, Kevin Greene, an academic research manager at Princeton University who leads the Digital Conflict and Information Integrity project, has evaluated chatbots and AI overviews for accuracy when responding to critical logistical questions about the voting process. While some AI models refused to give an answer related to election queries, he found most answers about election time, place, manner, and voter eligibility were accurate about 92 to 95 percent of the time. Model accuracy was consistently lower when prompted in Spanish, versus when prompted in English or Chinese. Prompting language also impacted whether a Google AI overview was offered in response to election questions at all.
A second element of AI chatbot responses relates to misinformation — concerns abound that AI might return incorrect information in response to voter or election queries, or at least fail to identify incorrect information as incorrect when prompted. To investigate these concerns, Mark Krass, a postdoctoral fellow at the Brennan Center for Justice, has studied AI output from six chatbots — Claude, Gemini, ChatGPT-4o, Sonar, Grok, and Deepseek — to hypothetical election-related queries, especially related to conspiracy theories, deepfake identification, and requests to generate deceptive images (like a dumpster full of ballots). Chatbots displayed promise in their rejection of conspiracy theories: in zero cases did Krass find that they would explicitly endorse a conspiracy theory. But even while rejecting explicit conspiracy theories, AI responses included a factual error in about one-third of the 120 answers the team studied. Krass also encouraged us to question the impact of the known sycophantic tendencies of chatbots. In this context, he found that chatbots tend to correct users, even when flattering them, if their perceptions are factually incorrect; however, the LLMs in his research tended towards flattering users by validating their worldview or general skepticism. This type of sycophancy poses an additional set of potential complications in understanding the role of AI as conduits of information, particularly in the context of polarized elections and a turbulent political climate.
And the models fell far short in several other areas. For instance, chatbots consistently failed to identify fake images — ChatGPT and Grok missed identifying images their own tools made in every tested case. AI was also quite willing to generate deceptive images, doing so in more than 80% of prompts. And even when it refused, it was easy to overcome soft rejections, often by using the same exact model with a simple second prompt. Krass concludes that both technical and preventative measures, such as amplifying the use of ChatGPT's thinking mode, which provides a much stronger guardrail against deceptive image prompting, should be implemented alongside one another.
Collectively, this research suggests that even as relatively few voters currently turn to AI, the answers they receive are falling short of ideal. As we look ahead to future elections where AI use may be higher, it’s crucial to better understand where models offer helpful information, where that information falls short, and where it actually becomes harmful.
AI cannot reliably replicate human public opinion
A core part of American elections is the horse race — constantly trying to understand who is leading and by how much, as well as how enthusiastic different groups are, and what issues are motivating them. But over time, public opinion polling has become more difficult as response rates have plummeted, making it hard to get answers from enough people and introducing bias into the process because some people are more or less likely to respond. So what does this have to do with AI? Increasingly, researchers and market researchers are attempting to use AI to simulate human responses, sometimes referred to as ‘silicon sampling’ or ‘digital twins.’ The obvious question this prompts is how effectively these approaches can mimic actual human public opinion. For now, the answer seems to be that AI approaches raise many concerns - both empirically and normatively.
Josh Clinton, a professor of Political Science at Vanderbilt University and senior election analyst for NBC News, has studied the feasibility of using silicon samples to measure public opinion. Public opinion suffers from a litany of known biases, like sampling bias, nonresponse bias, and interviewer effects — so perhaps AI could help us overcome some of the weaknesses of human-generated data. In Clinton’s work, however, this does not seem to be the case. While public opinion generated via digital twins seems roughly accurate on the surface and at the aggregate level, a deeper analysis finds that replacing human voices with silicon samples has a flattening effect, reducing the variance across a sample substantially when compared with human-generated public opinion estimates. Additionally, a fidelity paradigm arises with the proposed integration of silicon sampling: for this method to prove valid, AI responses should match those of the public; yet, current limitations of human polling prevent us from doing so. That is, if we can’t or won’t ask people the same questions we’re asking AI models to generate answers to, we can’t actually know how accurate silicon samples are.
One advantage of AI with regard to polling is its ability to synthesize open-ended responses, a historical limitation of public opinion polling, which has relied heavily on (perhaps overly simplistic) multiple-choice and yes-or-no binaries. Small polling firms can also now punch above their weight, since AI makes the work more efficient and frees up time for analysis. G. Elliot Morris, author of Strength in Numbers and cofounder of the election polling and forecasting platform FiftyPlusOne, draws a clear line for the application of AI as a polling tool: while offering the aforementioned benefits to synthesizing and analyzing polling data, his investigations of current attempts at AI-generated polling data reveal it to be faulty and incapable of replacing the human link in traditional polling. In evaluating the current approaches to silicon sampling for politics from multiple companies utilizing these techniques, Morris has found that the results these companies derive from trying to predict how humans would respond don’t effectively reflect polling data; rather, they manifest as prediction data.
Overall, research shows the substantial limitations of AI-generated public opinion data, at least in its current manifestation. Despite the potential efficiency benefits, alongside widening the information pool of public opinion data, AI falls short in resolving ongoing biases and logistical challenges in public opinion polling through silicon sampling, and fails to replicate the complicated nature of public opinion itself. Silicon samples also raise larger issues: given the importance of public opinion in informing political elites about citizen preferences, what would it mean for democracy if we outsourced the generation of those preferences to AI models?
Looking ahead to 2028
From election administration to polling to the information environment, AI is an ever-changing component of future election cycles. Researchers and policymakers find that the technology holds great promise for election offices and administrators, but realizing these promises will take sustained work and investment of resources — both of which tend to be in short supply. At the same time, a central priority is mitigating AI distortions in the information environment to sustain well-informed democratic elections. As the nature and use of AI continue to evolve, additional research will be required to better understand its role in future elections.
Authors


