On a September day in 2024, Dario Amodei, CEO of Anthropic, sat on stage at the Amazon Web Service’s Public Sector Summit. His message: “AI needs to empower democracies and allow them to function better and remain competitive on the world stage.”
At a time when much of the conversation about AI and democracy still revolves around deepfakes and synthetic fakery, Amodei’s statement tapped an alternate vein. What if artificial intelligence could supplement democratic processes and institutions, lowering barriers to participation and making them more inclusive and representative?
It’s not an idle question; researchers and entrepreneurs are already designing tools and products with this premise in mind. Deployed thoughtfully, they just might deliver. If adopted carelessly, though, we could be on the verge of a civic-tech cash grab that harms where it promised to heal.
The Promise of AI-Facilitated Deliberation
It’s not just Amodei. Just days before his remarks, 23 scholars released a paper titled “How Will Advanced AI Systems Impact Democracy?“ Their expectations are not rose-tinted; they warn of both opportunity and the potential for serious disruption of democratic processes and institutions.
The key throughline is language, the domain in which large language models (LLMs) excel. “Language,” the authors write,
allows information about candidates and policies to be shared with to [sic] voters, lawmakers to create legislation, and citizens and representatives to collectively discuss, deliberate and decide on which course of action to pursue. In liberal democracies, elected leaders use oratory to explain and justify their decisions and actions to the larger public, interest groups use persuasive messaging to lobby for their preferred policies, and the general public engages in debate, either informally or through organized events such as town halls and citizens’ assemblies.
These forms of communication and, more crucially, deliberation are the stuff of democracy. Research from cognitive science tells us that groups make better decisions when they deliberate. Research from political science shows that the process of deliberation doesn’t just change the decisions people make, it also changes how they feel about the deliberative outcome and one another.
Given this understanding, the authors’ recommendations also make sense. They write about “bridging algorithms“ to encourage constructive dialogue rather than online conflict. They also refer to “new avenues“ through which government officials and campaigning politicians can gain insight into the public’s views and concerns.
Major voices in the tech industry are especially bullish on this latter concept. In a recent essay in the MIT Technology Review, Andrew Sorota and Josh Handler—who work on AI policy for Eric Schmidt, former CEO and chairman of Google—note that “several states and localities are already using AI-mediated platforms to conduct democratic deliberation at scale, building on research showing that AI mediators can help citizens find common ground.”
Case Studies in Collective Deliberation
Sorota and Handler are right: there are several case studies of governments using LLMs to open the gates for broader democratic input. They often do so using tools called “collective dialogue systems,” or CDSes, which can facilitate democratic deliberation in several ways. For instance, they might take in and summarize a range of viewpoints that surface interesting ideas or areas of broad agreement for decision-makers; provide specialist or technical insights during group discussions; or serve some other function.
In a December 2024 paper, more than two dozen scholars explain,
In formal decision-making contexts, CDSes are typically not stand-alone exercises, but instead operate as part of a larger process that frequently includes in-person deliberations among stakeholders and decision-makers. Sometimes these technologies are coupled with citizen assemblies, where a “jury” of participants is selected to deliberate in person. In these cases, a CDS is often used jointly with citizen assemblies to crowdsource perspectives from the broader public.
This use of AI has a proven track record. The city of St. Louis used the GoVocal CDS to collect public feedback on how it should allocate more than $250 million from a settlement with the owner of the former St. Louis Rams. GoVocal served as a platform for more than 16,000 citizens to share ideas, discuss proposals, and express preferences. It then made all of that discourse legible and available to decision-makers.
Similar tools have found success internationally. In 2015, Taiwan, which has long been at the vanguard of digital innovation for democratic governance, used a tool called Pol.is to solicit citizen input and draft recommendations for the regulation of online alcohol sales. Since then, Taiwan has experimented with integrating Pol.is into other processes. In 2023, Carl Miller, cofounder and research director of the Centre for the Analysis of Social Media at Demos, said that LLMs have made the process of using Pol.is for public dialogue simpler and less expensive to run.
Researchers and engineers continue to improve these systems in the era of widespread consumer access to LLMs. When OpenAI announced a $1 million grant program for democratic inputs to AI in 2023, many of the recipients iterated on approaches like Pol.is.
Astroturfing: The Peril of AI
If every cloud has a silver lining, every moon has a dark side. So it is for AI as a tool for democratic processes. In an article for the Journal of Democracy, scholar of direct democracy David Altman warns:
Generative AI will act as a powerful accelerator for direct democracy, making its practice cheaper, easier, and more frequent than ever before, but also systematically undermining the civic foundations that make direct-democracy mechanisms both legitimate and stable. . . . We are heading toward a future of more direct democracy, but of a poorer, more volatile kind that will ultimately lead to a legitimacy crisis.
That future may be even closer than Altman suspected when he published those words in January 2026. Just a month later, a Los Angeles Times investigation revealed that an “AI-powered platform” called CiviClick had orchestrated “at least 20,000” emails to an obscure California agency, the South Coast Air Quality Management District, in opposition to a proposed rule phasing out gas-powered heat for residential homes. The rule failed in a 7-5 vote, an extraordinarily unusual outcome.
CiviClick describes itself as offering “personalized grassroots advocacy software.” According to the Times, however, some of the people whose names were used in CiviClick’s emails said they were unaware the company was sending emails on their behalf. As a result, the outpouring of opposition looks less like grassroots and more like astroturf. If the Times‘s allegations—and CiviClick’s characterization of its letters as personalized—are true, then the company appears to have weaponized the personal data of thousands of people to influence public policy on behalf of private interests, all while cloaking itself in the language of democracy. The case for this argument becomes even stronger when considering that CiviClick worked on behalf of a PR consultant.
CiviClick is a well-documented example, but not a lonely one. From political polling to election campaigns, political technologists are exploring how to remove people from more and more processes. When politics ceases to connect people to movements and instead replaces them with numbers and LLMs, it stops being democratic and becomes something else.
Lessons for the Future
We can take at least three broad lessons from the observations above:
First, the potential impact of AI on politics covers a broad range of positive and negative consequences. This was also true for the printing press, telegraph, radio, television, and the internet. Communication is fundamental to democracy and new communications technologies interact with it in complex ways, making it difficult to predict or even measure in real time. The key to harnessing the positive effects while avoiding the negative ones lies in implementation—a process that is under human control and not inherent to the technology itself.
Second, the beneficial use cases were overseen by public servants who were methodical and transparent in their approach. In the St. Louis case study, officials were clear in their reflections that the process of using AI for collective dialogue required a purposeful culture shift as well as technical know-how. It was not magically efficient; it required patience. Ultimately though, it can yield new insights and, handled appropriately, build public trust.
Finally, and perhaps most controversially, democracy requires friction to function. While gatekeeper is often thought to be a pejorative term by those concerned with democracy, institutional gatekeepers, which include advocacy groups and media organizations, play important roles in preserving the quality and authenticity of democracy (though they should also seek to include more viewpoints from society’s margins). Circumventing those roles can short-circuit deliberative processes and make them easier for well-resourced, private interests to manipulate.
For some things, the process is the point, even if the process takes time and effort. Public deliberation is one of those things. AI can enhance it, but it must not replace it.
Dean Jackson is a senior fellow at the University of Pittsburgh's Communication Technology Research Lab and at Tech Policy Press. In 2022, Jackson was an investigative analyst with the Select Committee to Investigate the January 6th attack on the US Capitol.
Resilience & Resistance is a Charles F. Kettering Foundation blog series that features the insights of thought leaders and practitioners who are working to expand and support inclusive democracies around the globe. Direct any queries to globalteam@kettering.org.
The views and opinions expressed by contributors to our digital communications are made independent of their affiliation with the Charles F. Kettering Foundation and without the foundation’s warranty of accuracy, authenticity, or completeness. Such statements do not reflect the views and opinions of the foundation which hereby disclaims liability to any party for direct, indirect, implied, punitive, special, incidental, or other consequential damages that may arise in connection with statements made by a contributor during their association with the foundation or independently.