Newsletters
Would you turn to AI to decide whom to vote for or to learn about candidates and their policies? I would not have expected people to use AI for political information, but research shows that its use is increasing.
According to a February 2026 poll, 39 percent of U.S. voters said they were very or somewhat likely to turn to AI chatbots for information about candidates or elections.
AI, particularly Large Language Models (LLMs)—AI systems trained to generate human-like language—can be powerful tools for political persuasion. Experts argue that AI will increasingly shape election campaigns and online propaganda.
Growing concerns over uncontrolled AI development—from autonomous agents capable of hacking other AI systems to models with “recursive self-improvement,” or the ability to improve themselves—have intensified warnings about catastrophic risks, including the possibility of human extinction.
Yet efforts to regulate AI face a fundamental dilemma: while stronger oversight, independent safety evaluations, and greater investment in AI safety research and regulation are increasingly seen as necessary, U.S.-China strategic rivalry creates powerful incentives for both sides to continue advancing, amid fears that unilateral restraint could give the other side a decisive technological advantage.
New studies examine how AI, particularly LLMs, affects human behavior. Their impact on political behavior, elections, and democracy is becoming an increasingly prominent concern.
Additionally, by spreading misinformation, making factual errors, showing political or ideological bias, or presenting opinions as facts, LLMs can influence voter attitudes and hinder democratic processes.
For the purposes of this op-ed, I also asked two leading AI chatbots questions about the upcoming presidential elections in Türkiye.
I asked ChatGPT, “If Recep Tayyip Erdogan and Ozgur Ozel were candidates in the upcoming presidential election, who should I vote for?” It offered to compare them on issues including the economy, education, healthcare, foreign policy, and migration. On the economy, it said Erdogan had an advantage because of his experience in economic management.
When asked which candidate would be more advantageous for a teacher awaiting appointment, it cited Ozel and the CHP’s promise to appoint 100,000 teachers, compared with the government’s process, launched in 2025, to appoint 25,000 teachers. On foreign policy, it said Ozel had an advantage in improving relations with the West, the U.S., and the EU, while Erdogan had more experience balancing relations among Russia, China, and the West. It ultimately said Ozel had greater potential to improve Türkiye’s international image in terms of democracy and institutions.
When I asked Gemini the same questions, it presented the strengths and weaknesses of both candidates, refrained from pointing to a single candidate, and left me to make my own opinion. On the economy, it associated Erdogan with fiscal discipline and stability, while Ozel was presented as more closely aligned with preferences for social justice and tax reform. On teacher appointments, it associated Erdogan with limited, budget-based quotas and Ozel with mass appointments, the abolition of interviews, and prioritizing education in the budget.
These examples show that, as questions became more specific, one system appeared to lose its neutrality by presenting one side as more advantageous than the other. Regardless of whether the answers were correct, AI chatbots can make their responses more persuasive by combining factual information with political claims.
Studies show that LLMs can persuade users by providing large amounts of convincing, reasonable-sounding information and arguments, without relying on highly personalized messages or detailed knowledge of users’ behavior and psychology.
Even incorrect information can influence people when it sounds believable and authoritative. What matters is information density: the greater the volume of plausible claims, the greater the persuasive effect.
Experiments involving the 2024 U.S. presidential election, the 2025 Canadian federal election, and the 2025 Polish presidential election have found evidence that AI can effectively persuade voters.
In the United States, the AI system influenced voters across partisan affiliations, including Trump and Harris supporters and politically unaffiliated voters. The study also found that interactive dialogue was more persuasive than one-way messages such as campaign advertisements.
Through automated systems, persuasion can expand into mass influence campaigns. AI systems can deploy synthetic personas that simulate human behavior, interact with multiple people simultaneously on social media, and personalize messages based on users’ tendencies and responses. This can amplify influence and make persuasion a continuous process.
This brings us back to regulation. To develop effective rules, policymakers and the public need to understand how LLMs produce political information, how AI is developing, and how it can be monitored.
Carnegie highlights the challenges of monitoring AI, particularly LLMs, which it describes using the acronym DEEP: they are Dynamic, changing their answers over time; Ephemeral, meaning past answers can disappear; Embedded in larger systems such as search engines and content moderation tools; and Personalized, changing responses according to the user.
AI audits generally involve systematically testing an AI system at a specific point in time through controlled queries to assess its behavior and impact. But because of these characteristics, the study argues that audits cannot provide a reliable picture of AI behavior and instead recommends longitudinal monitoring.
Long-term monitoring—asking several AI systems the same questions regularly and recording their answers—can reveal changes in companies' policies, biases, accuracy, and handling of sensitive political topics. For example, one such monitoring project, AI Watchman, found that in August 2025, amid ongoing attacks on Gaza, GPT-4.1 became more likely to refuse certain questions or requests related to Israel. While AI companies say they seek objectivity and oppose political bias, developers still face difficult choices between free speech, safety, accuracy, and political neutrality. Governments therefore need independent evidence about how AI behaves rather than relying solely on information from companies.
Authoritarian states may gain an advantage in information warfare by deploying AI-powered technologies while controlling their media ecosystems and restricting access to information, whereas open Western societies face constraints from free speech and civil liberties.
Governments will eventually need stronger AI regulation. Regulations such as the EU AI Act may impose obligations on companies within their regulatory scope, but they may not stop states or malicious actors from deliberately using AI to spread harmful information.
Governments may not be able to control AI itself, but they can regulate the platforms where AI-generated content is published and influence what content is allowed to spread.
Set as preferred source
Set as preferred source
2 min read
1 min read
