Introduction

Generative artificial intelligence (AI), and especially large language models (LLMs), is playing an increasingly substantial role in political discourse—for example, discussing political topics, answering legal questions, and relating election information to users. LLMs produce content that regularly reaches billions of people through chatbots and search engines, and a growing number of people are using them for political information. In the week before the 2024 UK election, 13 percent of eligible UK voters used LLM chatbots for election information. Just two years later, a March 2026 poll found that 46 percent of Americans used AI to get the news at least occasionally, and 39 percent said they had used AI to understand politics at least rarely. A February 2026 poll of U.S. voters found that 39 percent were very likely or somewhat likely to use AI chatbots to learn about candidates or elections. A recent study found LLM use already composes 2.9 percent of all time spent online.

As we demonstrate in this article, a growing body of scientific evidence shows that LLMs are not just ubiquitous, but persuasive and influential when discussing political issues. Based on their reach and influence, we see their impact on political discourse as a critical new frontier of modern politics. Today, the world is on the precipice of a new information ecosystem, roughly analogous to the start of widespread adoption of social media two decades ago. Ensuring that the public and policymakers understand the implications of this transition and are informed to make key decisions will require extensive independent research, just as social media has. 

We see LLMs’ impact on political discourse as a critical new frontier of modern politics.

Like social media and search engines, LLMs are often proprietary, personalized, and non-deterministic—they don’t give the same answer every time or to every person. They also change frequently and produce ephemeral content, but researchers typically can’t see into the past. If a person asks ChatGPT, “Am I eligible to vote in my state’s primary election?,” there is a real possibility they will see an answer that is different from the one they would have seen the prior week or month. If scholars want to study how past LLM responses affected some societal outcome, like voting, after the fact, they won’t be able to.

To thoroughly understand the role and impact of LLMs in political discourse, then, researchers, governments, journalists, and civil society organizations need to look beyond one-off AI audits and individual studies to a new paradigm of research: comprehensive longitudinal monitoring. Longitudinal monitoring entails building systems that repeatedly (for example, daily or weekly) probe LLMs on a wide range of topics and track their responses, creating a thorough record. This approach enables tracking trends in LLM function, discovering undisclosed changes in LLM behavior, and making rigorous comparisons across LLM providers to power an entire field of essential causal research on the impact of LLMs. While longitudinal monitoring is costly and complex—especially over time, on many topics, and across many LLMs, geographies, and languages—both the technical qualities of the models and their burgeoning impact on political discourse demand this more comprehensive approach.

Longitudinal monitoring entails building systems that repeatedly probe LLMs on a wide range of topics and track their responses, creating a thorough record.

LLMs Can Be Politically Impactful

LLMs are already having a meaningful impact on public discourse, including about political issues, politicians, and elections. It is also clear that LLMs will be increasingly integrated into a range of other tools and systems—including search, content moderation, and fact-checking—which will extend the impact of LLMs on political discourse.

An emerging body of evidence demonstrates that LLMs are already influential and persuasive on key political issues. One series of experiments found political messages generated by LLMs to be as persuasive as messages written by lay humans on policy topics, such as tax policy or paid parental leave. Study participants noted that the LLM-generated messages’ use of facts, evidence, logical reasoning, and a dispassionate voice were all factors in the persuasiveness of the messaging. Another project tested nineteen different LLMs on over 700 political issues and found a similar result—a high density of factual claims drove the persuasiveness of LLMs. This project noted that post-training and prompting significantly improved persuasiveness of LLMs, although it also made the LLMs less factually accurate.

That LLMs can be persuasive is important, but not inherently good or bad—the net effect will depend on the information quality relative to alternative sources. One experiment found that LLM use (GPT-4o, Claude 3.5, or Mistral) was as beneficial for political knowledge as using Google Search when researching issues including climate change, immigration, criminal justice, and COVID-19 policy. Notably, research about retention has suggested that learning through LLM summaries may lead to shallower knowledge than learning through web search. The availability of LLMs has also changed online forums like Reddit—an analysis of the largest political forum on Reddit found that the release of ChatGPT resulted in increased ideological divisions, but notably less toxicity and affective polarization.

That LLMs can be persuasive is important, but not inherently good or bad.

It is also possible that intentional design choices can improve the role of LLMs in the information ecosystem. When tailored to do so, LLMs were able to marginally reduce conspiratorial thinking in one study by providing specific facts contradicting false election denial beliefs. Further, while standard models performed poorly as fact-checkers, LLMs with access to curated high-quality context (a database of fact-checking claims) demonstrated significantly improved responses; other technical interventions can also improve fact-checking capacity.

Emerging Information Risks from LLMs 

In addition to potential positive effects of LLMs, we must also consider their significant risks, including viewpoint and ideological biases, misinformation, and censorship. It should be no surprise that LLMs demonstrate viewpoint biases (an issue that has plagued prior algorithmic systems), and that these vary across different LLMs. In 2023, a paper used an existing opinion poll to compare the opinions of the U.S. populace to the responses of LLMs. This analysis found significant variation between the responses of LLMs and the beliefs of the general U.S. population, and larger gaps for some groups, including older and widowed Americans. More generally, LLMs display ideology that is similar, both geographically and by language used, to that of their developers.

In general, LLM truthfulness remains a key challenge. A late 2025 Newsguard analysis of ten leading LLMs found that when asked about a false news-related claim, the models asserted the false claim was true 35 percent of the time. Earlier in 2025, BBC gave its own news content to OpenAI’s ChatGPT, Microsoft’s Copilot, Google’s Gemini, and Perplexity AI, then asked the models news-related questions, ultimately finding that 19 percent of the answers introduced factual errors. This pervasive issue also interacts with algorithmic discrimination—LLMs have been shown to provide less accurate information, more frequently refuse prompts, and even use more condescending language for users with lower English proficiency, less education, and non-U.S. origins.

LLM truthfulness remains a key challenge.

In contrast to the threats LLMs pose for content they do produce, there are also risks in the other direction: excessive censorship. When examining LLMs used for content moderation, one of our research projects found that speech related to identity categories (such as different races, genders, or religions) is more likely to be incorrectly suppressed than other speech (for example, disability-related content being flagged for self-harm risk). LLMs may also be aiding authoritarian information control. Researchers have observed that prompts written in languages of countries with less press freedom tend to exhibit stronger pro-regime bias, and that Chinese models like DeepSeek display much higher refusal rates, averaging around 33 percent, on topics deemed politically sensitive in China.

This existing research demonstrates that the emergence of LLMs for finding or synthesizing political information, political messaging, fact-checking, content moderation, and related applications will have consequential effects on political beliefs and discourse. For the public and policymakers to have a meaningful awareness of, and influence on, how LLMs produce this critical information, longitudinal monitoring is essential.

The DEEP Necessity of Longitudinal Monitoring

The potential political impacts of LLMs demand rigorous evaluation. To inform the design of effective approaches for doing so, this article calls attention to several key challenges involved. LLMs are hard to study well because of a set of attributes we describe with the mnemonic “DEEP”: they are Dynamic, Ephemeral, Embedded, and Personalized. Taken together, these attributes necessitate ongoing longitudinal monitoring, rather than one-off or intermittent evaluations.

Dynamic: LLMs are always changing, even between model updates. Most obviously, this is because LLMs are inherently stochastic—that is, they inherently have some degree of random variation in their outputs. This randomness can cause significant changes in, for instance, what speech an LLM finds to be toxic. But the dynamism of LLMs goes beyond this: LLMs that leverage tools to interact with databases or the internet may give responses based on that content, which can change. While many LLM companies do note changes in the models themselves (partially inspired by researchers finding drift in ChatGPT’s performance), it’s always possible that some unannounced fine-tuning or other model adjustments are causing changes. There is also an inconsistent culture of disclosing changes to system prompts, which are the default instructions set for an LLM by its developers and cause changes in outputs. Collectively, this range of possibilities means that LLMs can offer dynamic outputs even over short periods of time, which users may not realize.

Ephemeral: The outputs of LLMs often do not persist, but rather are ephemeral. For instance, AI-generated answers in search engines, like Google’s AI Overviews, disappear after the user closes the page. Questions asked to ChatGPT by a user who is not logged in similarly vanish. Subscribers to an LLM chatbot, meanwhile, may have access to a longer record of their history, but this too is solely in the hands of the LLM provider. And in all cases, researchers cannot retrospectively know what an LLM would have answered last week or last year to any question that was not asked at the time. Although there have been efforts to record and save LLM conversations, these are isolated and time-limited efforts; to answer many important questions, researchers need access to consistent records over longer periods of time.

Embedded: LLMs are rarely deployed on a stand-alone basis, but instead are embedded into more complex software systems that may significantly affect their outputs. Most central to this are the LLM guardrails that filter prompts into and responses from LLMs. Tools including Llama Guard, NeMo Guardrails, and Guardrails AI attempt to detect harmful prompts or other misuse, and then either alter input instructions to the LLM or refuse to accept the input. Online services may advertise that they are powered by a specific LLM, but if there are differences in these guardrail systems or other supporting software infrastructure, the overall service will perform differently. Hardware plays a role too. LLMs running on-device, such as on a smartphone, are likely to be smaller and less capable (for example, Gemini Nano) than LLMs running in the cloud. They also may have more limited ability to use other software tools, reduced file access, and smaller context windows than an LLM running on a desktop computer. The outputs of LLMs can’t be assumed to be consistent across different environments—the models themselves are integrated into complex larger systems.

Personalized: Lastly, LLMs are becoming progressively more personalized, differing in their inputs and settings between users. A notable turning point was OpenAI’s 2024 implementation of ChatGPT’s memory feature, which was later mirrored by Anthropic, Mistral, Google Gemini, Google Search, Perplexity, Microsoft, and xAI. These memory features enable the storage of user-specific information across all chat sessions. This information can include background information about the user (such as their name, employment, or location), their preferences (for example, that one prefers brief, fact-dense responses, or that another prefers the programming language R), as well as prior chat histories or even other personal files on the user’s computer. Observing the practices of most technology companies, it’s safe to assume LLM-based tools will continue collecting and storing increasing amounts of personal information, and that personal information will have an increasing impact on LLM behaviors (including advertising). This may significantly complicate the evaluation of LLM-generated content and its impact on individuals and groups of users.

Looking Beyond One-Off AI Audits to Longitudinal Monitoring

A highly skilled research community has developed around the goal of auditing AI systems—a field that we have contributed to extensively and that has informed our policy analysis and government work. Auditing entails the use of computational methods to repeatedly query an AI system with controlled inputs and then analyze its outputs to understand its behavior and impact without direct insider knowledge or access. These studies are generally conducted at a single point in time, but may be repeated at specific intervals. AI audits have been extremely informative and remain essential to the modern, evidence-based understanding of AI, including LLMs.

The research community must move beyond one-off or periodic AI audits to develop infrastructure for longitudinal monitoring.

We argue that the DEEP challenges, also present in other AI systems but even more relevant for LLMs, require the research community to move beyond one-off or periodic AI audits to develop infrastructure for longitudinal monitoring. While an AI audit is typically a point-in-time evaluation, longitudinal monitoring is ongoing, enabling detection of subtle shifts over time and identifying recurring patterns or specific inflection points.

Our AI Watchman project is one instantiation of this mission: a longitudinal monitoring system that automatically runs consistent queries about hundreds of socially salient topics on a weekly basis over long periods of time. Our preliminary results demonstrate the need for this work. For instance, AI Watchman indicated that in August 2025, amid ongoing attacks on Gaza, GPT-4.1’s rate of refusal may have substantially increased on content related to Israel. A month later, as the Texas legislature debated a bill restricting the accessibility of abortion medication (which it later passed), our measurements indicated an increase in GPT-5’s rate of refusal on prompts related to abortion, returning to baseline a week later. Another four-month monitoring project identified similar sudden changes—including changing responses to questions about the 2024 U.S. presidential election just a few months beforehand. A 2024 analysis of ChatGPT performance on a broad range of tasks found significant variation over time, and in some cases substantial performance deterioration over time (for example, on code generation and mathematical problem-solving). While these data points cannot prove a definitive causal link, they may be indications of a company’s internal policy decisions affecting model outputs—changes that would otherwise go unnoticed. Longitudinal monitoring is necessary for identifying these changes as they occur, as well as inferring why and at what scale.

Learning from the Social Media Era

Researchers can take lessons from social media platform research as inspiration for longitudinal monitoring. One key lesson: Two decades in, it is clear that the research community cannot rely on private technology companies to make necessary data readily available. The evolving rules and restrictions for access to social media data have demonstrated intermittent and half-hearted commitment to collaboration with external researchers from technology companies. Today’s AI companies face many of the same incentives and pressures as social media companies, including the risk that providing data to researchers will enable public criticism of LLM deployment practices.

The research community cannot rely on private technology companies to make necessary data readily available.

In our view, however, the AI companies stand to benefit as well. LLM monitoring is an essential transparency mechanism to inform the public. Rigorous and robust evidence on LLMs will improve public debate, leading to more accurate discussions—including both praise and criticism—about the societal role of LLMs. It will also lead to crowdsourced advice and techniques to improve LLM function. These functions are critical at a time when the Stanford Foundation Model Transparency Index shows meaningful declines in transparency. Ultimately, we believe that AI companies that meaningfully engage with researchers stand to improve their reputation and their products.

AI Companies Are Deciding Their LLMs’ Politics

We have argued that LLMs are both politically impactful and challenging to study, factors which necessitate longitudinal monitoring. It is also essential to understand that AI companies have meaningful agency: The political outputs of LLMs are neither accidental nor unconnected from the normative perspectives of their developers. Rather, the LLM companies are making distinct choices based on their leadership’s normative goals, as well as the prevailing political headwinds. For instance, Anthropic states that it is seeking “political even-handedness” through its (open-source) automated evaluation framework for Claude models. Alternatively, OpenAI states that ChatGPT “shouldn’t have political bias in any direction,” and operationalizes this primarily through seeking “objectivity.” Meta’s Llama team has stated, “Our goal is to remove bias from our AI models and to make sure that Llama can understand and articulate both sides of a contentious issue,” and the newer Meta Superintelligence team has stated similar goals. Google has been more restrained, originally restricting Gemini’s responses related to the 2024 elections in the United States and India, and more recently taking a “both sides” approach to political questions, according to a Washington Post analysis.

The political outputs of LLMs are neither accidental nor unconnected from the normative perspectives of their developers.

Normative goals are critical for setting the future direction of LLMs, and academic researchers, civil society organizations, and the public must play a role. To do so, these stakeholders must be informed by the kind of comprehensive and independent evaluations of LLM behavior that longitudinal monitoring provides. For instance, while the companies and some academics have argued for various forms of political neutrality, it is not clear that this approach is beneficial to the general public, given an information ecosystem with highly asymmetric challenges, where echo chambers and misinformation are more pervasive in right-aligned political discourse. Similar research shows that optimizing LLMs for “truthfulness” can result in a left-leaning bias, suggesting that attempts to make “unbiased” models could come at a cost to truthfulness.

Judgments pertaining to LLM refusal pose a similar challenge; responsible developers must balance potentially competing interests including freedom of expression, factual accuracy, nondiscriminatory speech, and child safety. A report from the Future of Free Speech evaluated models on their support for free expression, finding that xAI’s Grok 4 was the strongest LLM tested in terms of free-speech culture. But that same model, launched immediately following Grok 3’s infamous MechaHitler incident, is also accused of generating tens of thousands of naked images of children. Balancing freedom of expression and access to information on one hand, and safety guarantees and harm mitigation on the other, is an ongoing challenge.

AI Companies Are Implementing Norms

These normative trade-offs—between free expression and safety; between political neutrality and accuracy or relevancy of information—do not have “objective” answers or costless solutions. Regardless, LLM developers and deployers today are actively making choices, intentionally or not, about how to approach these thorny questions.

LLM companies are actively curating ever-larger training datasets, and there is evidence that training data plays a significant role in an LLM’s political responses. Other research has demonstrated that exposing LLMs to diverse viewpoints in post-training (that is, when adjusting the model after its initial development), can lead to the models summarizing multiple viewpoints instead of taking a specific stance. More simply, the design and writing of system prompts have also been demonstrated to substantially impact LLM outputs. Lastly, the user interfaces through which people access LLMs can also significantly influence their outputs and interpretation. For instance, how LLM interfaces cite and display underlying source materials, especially journalistic content, is another lever that LLM companies can use to influence their users. Across all these interventions, LLM companies have options that enable them to implement their normative goals, although many of these interventions are notably opaque to the public.

Normative trade-offs—between free expression and safety; between political neutrality and accuracy or relevancy of information—do not have “objective” answers or costless solutions.

Longitudinal Monitoring Is Challenging, but Societally Necessary

The evidence is clear that LLMs will play a significant and prominent role in the information ecosystem and in political discourse. Further, the technical attributes of LLMs—especially their dynamic, ephemeral, embedded, and personalized nature—warrant not just single-time-point measurements, but comprehensive longitudinal monitoring. While uncommon in the study of technology systems, the collection of this type of systemic, ongoing, and routine data is common in other research fields of significant societal interest. This includes education, family socioeconomics, and the labor market, fields where data is often captured over time from human subjects known as panels. Longitudinal monitoring of LLMs is in its infancy, revealing the relative immaturity of this style of research in computing, compared to social science fields where such work has been conducted over decades of research and supported extensively by federal agencies and other sources of research funding.

There remain significant barriers to longitudinal monitoring worth considering. Most immediate is cost. Longitudinal monitoring requires frequent prompting of many LLMs across many topics, many languages, and many device types. Further, the probabilistic nature of LLMs means an LLM response will at times differ even between two identical prompts of the same model at the same time—our preliminary research demonstrates that the same prompt should be repeated fifteen to twenty-five times to get a reliable result. In many cases, this requires running up a large bill for the market cost of using those LLMs for many, many prompts.

We encourage LLM companies to enable more cost-effective research options, and similarly urge philanthropic foundations and academic institutions to collaborate on developing more efficient shared longitudinal monitoring infrastructure. A future presidential administration should also revive the National Science Foundation’s social science directorate and build a new program for studying the information ecosystem, including longitudinal monitoring of LLMs. It’s also worth noting that if the United States ever creates a technology regulatory agency, this type of research would be essential as part of that federal oversight.

Another foreseeable challenge is that some more adversarial AI companies may take legal action to discourage public-interest research, as some social media companies have done in the past. Legislators should codify safe harbor protections for LLM auditing, monitoring, and user studies, and research organizations like universities should be prepared to defend researchers’ right to conduct such work. Such protections are necessary, since longitudinal monitoring can and should play a role in informing technology policy and regulatory efforts, both in guiding their design and monitoring their effects.

The technical attributes of LLMs—especially their dynamic, ephemeral, embedded, and personalized nature—warrant not just single-time-point measurements, but comprehensive longitudinal monitoring.

The broader question of when and how to govern LLMs goes beyond the scope of this analysis. However, it is worth noting that policymaking has started to touch on the political repercussions of LLMs. Last year the administration of U.S. President Donald Trump issued a misguided and harmful executive order against equity in LLMs procured by the federal government, but more reasoned policymaking is still in its infancy. A Brennan Center review of state legislation related to AI and elections found some proposed (but no enacted) laws addressing election misinformation from LLM chatbots. The European Union’s AI Act gets closer—under the AI Code of Practice, developers of large LLMs must mitigate the potential for harmful manipulation, which could include manipulation of politically relevant LLM outputs or using LLMs for creating propaganda. We believe the election and political outputs of LLMs will be subject to legislative efforts—both reasoned and ill-designed—in the coming years. For instance, Center for Democracy and Technology’s CEO Alexandra Givens recently foreshadowed laws preventing LLMs from offering information about reproductive rights and healthcare.

These debates will continue to heat up, and we expect to be debating the role of LLMs in political discourse for decades to come. For the public and policymakers to have a meaningful say in how they work, both normatively and technically, researchers and governments need to invest in the science and infrastructure of longitudinal monitoring.