Journal of Democracy
Volume 36, Number 4, October 2025
Johns Hopkins University Press
Dean Jackson (bio) and Samuel Woolley (bio)
In 2024, observers worldwide braced for the electoral impact of generative artificial intelligence (AI). With those contests over, attention should shift to the longer-term risks AI poses to democracy. This essay predicts three such risks. First, AI-backed efforts to replace political communication may erode representative democracy. Second, AI may exacerbate trends toward the concentration of wealth and power, preserving only the facade of democracy. Third, economic trends in media and technology threaten to emaciate already weakened sources of trustworthy information. Avoiding these outcomes will require policymakers to reduce their reliance on the perspectives of industry professionals.
The year 2024 opened with predictions that a surge of mis- and disinformation powered by artificial intelligence (AI) would soon be the greatest threat to global stability.1 Pundits claimed that months of “AI elections” lay ahead as campaigns, publicists, and politically motivated individuals around the world experimented with AI.
In fact, AI-powered influence efforts did appear. Chatbots imitated politicians online, and campaigns used cartoon avatars to rehabilitate their candidates’ public image. Female candidates became victims of nonconsensual AI-generated intimate imagery. Social media were flooded by fake newspapers filled with AI-generated “pink slime.” Campaigns and political operatives unleashed AI robocallers and robotexters. Synthetic videos brought national leaders back from the dead to provide endorsements.
Yet by January 2025, it was hard to say whether all this had added up to a bang or a whimper. Despite warnings of undetectable, AI-generated fake imagery designed to trick voters, events vindicated those who argued that the ability to quickly and cheaply proliferate content would not by itself significantly help propagandists.2 Generative AI has made it easy to crank out misleading content, but sheer volume has limited value on an already saturated internet, and technical sophistication is not needed to persuade people of things they already believe. It seems likely that preexisting fears about social media and disinformation were being projected—often mistakenly or without empirical evidence—onto generative AI. And yet, generative AI does appear to be spurring a slow [End Page 139] but steady shift in the behind-the-scenes creation and delivery of political communications. Today’s electioneers are experimenting with using AI systems to augment their data analyses and message targeting.3 This could lead to more refined targeting (for instance, of swing voters) in years to come.
With the urgency of that major election year behind us, we can now think more deeply about longer-term AI-driven risks to political communications—risks that are serious but years or decades rather than weeks or months away. These include threats to democracy stemming directly from how societies choose to develop and implement AI. These fears are distinct from warnings about existential risks to humanity, “runaway” AI, or the arrival of artificial general intelligence (AGI, a vague term that OpenAI uses to refer to “highly autonomous systems that outperform humans at most economically valuable work”).4
Many will continue to focus on these risks, which are real but also often subject to sensationalism (which draws views and clicks). Attention should go as well to more gradual, less dramatic changes. Like good driving, good AI governance requires simultaneous awareness of multiple risks across multiple timelines: the car just ahead as well as the long, downhill curve coming into view in the distance. In the global fight against disinformation, it has taken concerted effort and time to determine which concerns are overblown and which are underrated. The same will be true for those envisioning a world where AI and democracy coexist.
In September 2023, technologists Bruce Schneier and Nathan Sanders described three “deeply divided factions” battling to decide AI policy.5 The first faction, which they called “doomsayers,” feared a future in which uncontrollable AI systems would enslave or even destroy humanity. A second group, the “warriors,” envisioned a zero-sum race among nation-states to develop and control AI, with the winner achieving global leadership for the rest of this century and perhaps beyond.
The third faction, which Schneier and Sanders called “reformers,” wanted to stop the threats to rights that they thought reckless AI adoption would bring, but with a focus on the very near term. Examples of algorithmic, AI-driven rights violations around the globe are already plentiful: Racial biases embedded in criminal-sentencing algorithms cause unequal outcomes in justice systems; antifraud and other vetting [End Page 140] algorithms have erroneously stopped people from receiving public benefits; and an algorithm used to grade U.K. school exams lowered working-class students’ scores and kept many from entering universities.
For years, governments have been using computers to make decisions about healthcare claims, child-welfare investigations, and housing assistance. Large language models (or LLMs, the software behind services such as ChatGPT) will make this practice both more opaque and more pervasive. As the use of AI spreads, it will only become harder to find and fix the injustices that it will automate across an increasing number of domains. It is also worrisome that AI is being implemented—including by governments—while it is still too crude, and without sufficient vetting or moderation.6 This will cause problems that will be even more acute and difficult to detect.
Inequality and social exclusion long pre-date AI, of course, but technology can add new force and extent to old evils: Consider how much scientific industry has done to heighten the lethality and destructive potential of warfare. Moreover, technology developed for one purpose can often be used for another. In China and the United States, for instance, data gathered from social media, license-plate readers, facial-recognition cameras, government records, and other sources were first collected for use in public safety and counterterrorism before contributing to new forms of predictive policing.7 This encroachment and the existence of this AI-powered panopticon cast their own chill on political activity.
Today’s Choices Will Shape Tomorrow
While sympathetic to reformers’ immediate concerns, we hope to draw attention to the longer-term ways in which AI could strain—or even crack—the foundations of democratic political systems around the world. The beginning of wisdom is to set aside techno-determinism: In shaping the future, human agency matters. Policymakers and business leaders today are choosing how AI will develop and be used in societies around the world for decades to come. In other words, AI is an example of a “normal technology” whose impact will be decided by accelerants and guardrails that humans put in place.8
What if, over the next two decades, concerns about runaway AI fail to bear out but instead, industry lobbying and the pressure of great-power competition lead to AI being adopted with little regulation?9 The technology could change the processes, institutions, and socioeconomic preconditions that allowed democracy to thrive for most of the last century. Many AI “warriors” are openly working toward a world where AI replaces much of the infrastructure that makes contemporary democracy tick. Some in Silicon Valley even argue that the era of democracy has passed, and are resuscitating once-scorned notions of technocracy.10 As the political scientist Henry Farrell warns, the race for human super-intelligence [End Page 141] is leading powerful actors whose material and ideological interests are hostile to democracy to push AI systems as replacements for things that governments do.11
If the current boom in AI leads to its being adopted without sufficient thought, we predict three trends, each of which will be bad for democracy. First, AI-backed efforts to streamline or even replace political communication between political officials (actual and aspiring) and the public will break the feedback loop between governors and governed which is at the center of representative democracy. Second, AI will exacerbate existing concentrations of wealth and power and turn democracy into a hollow shell. Third, the broadscale consolidation of information by LLMs, alongside generative AI’s capacity to appropriate from and overshadow original, human-generated work, will have economic consequences for already dwindling sources of trustworthy news and information. Today’s epistemic crisis will deepen while the tech sector claims even greater control over information and public discourse.
Popular representation and elections have not always been synonymous with democracy. In ancient Athens, citizens filled the Assembly by lot, and the U.S. Founders maintained a distinction between the representative republic they created and “pure” democracy. Today, however, the terms are largely interchangeable.
This obscures the complexity of the processes by which representative government translates popular will into policy. These processes are often ancient, imperfect, and frustrating. Technical changes in district-drawing and vote counting can greatly change outcomes, making parties rise or fall in number and officials more or less prone to become polarized, for example. Reliance on elections—often with low turnout—to gather input from citizens with widely varying levels of political knowledge and education has long dismayed many observers. Political discourse often feels fruitless and even toxic.
Technologists offer AI as a means of improving democratic processes from political campaigning to constituent services to opinion polling.12 In 2024, for example, a candidate for the Pennsylvania General Assembly deployed an AI robocaller whose creator mused that state and local officials might one day use such chatbots to collect and analyze constituent opinions.13 Today, U.S. officials are beginning to use AI in attempts to scrutinize contracts, detect fraud, and deal with recipients of government services. Former Google CEO Eric Schmidt has even predicted that AI will fundamentally remake government, pervading legislative and judicial processes.14
Public comment, an important part of the U.S. regulatory process, is already being manipulated by computer-generated content. In 2017, for example, a Federal Communications Commission call for comments on repealing “net neutrality” drew 22 million responses, about a third of [End Page 142] which were traced to automated accounts created by telecom companies while another third came from a single computer-science student. Generative AI makes this kind of manipulation much easier to scale up and refine. Congress drafted a “Comment Integrity and Management Act” to address this issue in 2024, but it has yet to become law.15 Some observers predict that future pollsters will have AI make inferences to compensate for declining survey-response rates—subbing postulated voters for real ones.
More profoundly, polling and formal models provide none of the benefits of deliberation, which political and cognitive science consider essential to group decisionmaking. Deliberation is central both to public debate and legislative processes. Through it, people challenge biases, question assumptions, solicit information, and forge compromise, reaching more defensible conclusions and durable decisions. When a series of field experiments brought together five-hundred U.S. voters to discuss divisive policy issues from healthcare to immigration, their policy views came closer together and they became less hostile to members of the opposing political party.16
It is reasonable to expect that similar conclusions apply to legislators: The more reliant they are on data to form their position, the more hardened their stance becomes and the more polarized the legislature grows. Politicians who are reliant on polling data also have less incentive to compromise and are more likely to rely on “narrow-cast” messages that target specific groups of voters with wedge issues.17 The result is political communication more akin to slogans shouted into a bullhorn than thoughtful dialogue or debate.
In The Idea of Justice, economist Amartya Sen describes democracy as the process of public reasoning that allows conflicting perspectives to coexist under shared governance.18 Attempts to supplant this process with AI reflect an impulse as old as Plato’s Republic—the hope that perfect technical knowledge can replace deliberative processes that are slow, frustrating, and almost never totally satisfying. And yet those processes have endured for centuries, outlasting many rivals. They succeed precisely because they are slow and deliberative. They cannot be bypassed or replaced through computing power. To do so would render them undemocratic.
Inequality, Technocracy, and Oligarchy
The second risk we foresee stemming from unfettered AI is the more rapid corrosion of democracy by rising economic disparities. This is not a new trend, but AI may make it worse: Scholars largely agree that information-technology gains late in the last century fueled sharper inequality, and expect advances such as AI to do the same in this century.19 The predicted changes to labor markets are dramatic. Dario Amodei, the [End Page 143] CEO of Anthropic, believes that AI could replace half of all entry-level white-collar jobs within five years. At least one startup company, Mechanize, says its core mission is to automate white-collar work as quickly as possible. A recent headline claimed that the world might soon see its first billion-dollar company with a single human employee.
Recent scholarship holds that inequality—whether of wealth or income—is a strong predictor of democratic decline and collapse, even in older democracies, because it drives negative trends such as distrust, polarization, cynicism, and autocratic populism.20 Most analysis focuses on these broader factors, while another common research lens examines how changes in labor markets affect factors such as wage spreads, educational access, and civic and social participation. Writing in the Journal of Democracy, Stephanie Bell and Anton Korinek suggest that if AI widens the gap between rich and poor, democracies might degenerate into corrupt oligarchic systems that could, in a vicious cycle, breed deeper populist reactions among citizens frustrated by insular, unfair elites.21
Governments are prone to capture by well-resourced groups and individuals, and tech multibillionaires have a lot of resources. They may want to retain democracy’s trappings and some of its features precisely because a powerful autocrat would threaten their own independence. Yet today’s antidemocratic, ultra-wealthy AI evangelists go a bridge beyond: While old-school oligarchs wanted to capture the state, the new AI promoters want to supplant it altogether.
It is not a coincidence that some of the world’s richest tech executives are both ideologically opposed to representative government and enamored of a technology that promises to replace people—and their messy demands for justice—with software. In fact, some theorists expect exactly this outcome as the yawning gap between the world-historically rich and everyone else widens. Democracy in a capitalist system is possible because capital-rich elites value an educated workforce enough to tolerate the demands for redistribution that it will make; workers’ bargaining power and participation in organizations such as unions further buttress the system. What happens when elites believe that AI can replace those human workers?
Rising income inequality and technological advances are changing these structural precursors to democracy. Today, technology moguls see a future in which AI increases returns to capital, replaces educated workers with machines, and weakens labor’s already diminished bargaining power even further. In other words, they hope to make the public and its inconvenient demands for government services and social justice irrelevant.22
If all this is deeply unsettling, consider one final danger: Democracy’s replacement by a synthetic alternative is likely to lead only to even greater inequality and concentration of power, creating a feedback [End Page 144] loop in which democratic decline and inequality accelerate each other. In the worst case, it could take an economic collapse or a civil conflict to break out of such a cycle.
Kings of the Information Jungle
The third challenge we anticipate is that the information landscape—already unsettled by social media’s rise and shifting in response to generative AI—might transform in ways that make democracy less tenable. This prediction is based on two trends: First, important information sources—from newspapers of record to digital outlets—are already struggling economically as LLMs divert web traffic and diminish a revenue stream that social media have been siphoning away for years now. Second, as LLMs become a more common point of departure for consumers seeking information, a small number of tech-sector players will gain more control over public discourse and public opinion.
Despite worthwhile experiments in using AI to supplement journalists’ coverage, human reporting offers benefits that AI cannot replace. As media scholar Courtney Radsch has pointed out, AI is not going to interview witnesses to news events.23 The deeply contextual, human-driven nuance and instinct that reporters provide is not something we can get from a machine.
In newsrooms currently, AI is often used to churn out low-grade content (“churnalism”) that consists of lightly edited or unedited official statements, police reports, sports scores, or PR-agency press releases. This is not sourced reporting, it is dehumanized reporting by way of variables selected by constantly evolving computational processes. These processes are, however, determined by coding, data-tagging, and content generated by people, with their biases and desires often baked into the output in opaque and worrisome ways. The hiddenness of the human role in such activities gives many the misleading impression that AI and algorithms, whether news-oriented or otherwise, are somehow more “objective” than human decisionmaking. In fact, they can be more flawed and less grounded in reliable, properly cited sources than content that is overseen by traditional human editors who follow clear standards for ethical journalism.
If the news industry’s already parlous economic fortunes keep sliding, the tasks of holding authorities to account and helping to make sense of events may go begging, since only humans can perform them. From Wikipedia to newsrooms, knowledge professionals are worried that LLMs are overwhelming their sites with web scrapers, plagiarizing their content, and shrinking traffic from internet searches.24 AI-generated news summaries may also diminish consumer demand for subscriptions to newspapers and websites, shrinking revenue further. As trustworthy, vetted sources of information wither—trained by LLMs even as these take away traffic—news [End Page 145] consumers will have fewer sources of information on recent and especially local events. The internet is already awash in junk content, and AI threatens to make it worse.
Fear of this is at the center of a lawsuit that the education-technology firm Chegg has filed against Google. Chegg’s is an antitrust case which charges that fewer users visit Chegg because Google puts its own AI-generated summaries at the top of search results. Another lawsuit shows how the largest publishers may yet survive: The New York Times is suing OpenAI, the creator of ChatGPT, for using the newspaper’s content to train that model and for the model’s ability (when prompted) to reproduce content from the Times.25 Many expect the case to end in an agreement that OpenAI will pay for access to the newspaper’s material. Meanwhile, publishers and local outlets without the clout and resources of the New York Times are at risk of being left out in the cold.
Ironically, the impoverishment of web publishers threatens future AI development as well—resulting in a dead end on the information superhighway. Today’s LLMs were trained on trillions of words, and if advances are to keep up the pace of improvement set so far, AI will require exponentially more data. Where are these vast oceans of data to come from?
Some observers have suggested that the next generation of LLMs might be trained on synthetic data—in short, machines will produce content that trains other machines. Yet what if human-produced content proves hard to replace? Some researchers suggest that AI systems trained on synthetic, machine-produced data revert to the mean over time in a process called “model collapse”: When models built from statistical aggregates are layered on top of each other, their outputs can become more homogenous.26 Other researchers hope that so-called reasoning models will yield more insightful results without the need for exponentially more training data, but a new paper from researchers at Apple suggests these models overthink simple problems while quickly giving up on complex ones.27
If LLMs become a leading or even primary source of news and information, they could deepen the already troubling ability of technology companies to act as information gatekeepers and molders of opinion. Google, Facebook, TikTok, and Twitter currently decide what pieces of user-created content billions of people see. Yet at least those services relay (most of the time) real content created by real human beings. Generative AI, by contrast, comes up with its own content via complex combinations of user prompts, decisions made during model training, and safeguards and instructions that companies add when preparing models for consumer use. If AI chatbots start supplying large chunks of the media that people use, the influence of tech executives on public discourse will become even greater and less transparent.
A few moguls with their own economic and political interests should [End Page 146] not be trusted with greater power over the flow of information. Recent incidents in which Elon Musk’s AI chatbot Grok responded to user prompts with unrelated conspiracy theories about white genocide in South Africa and a deluge of antisemitic posts demonstrates the risk that such systems might be used for political manipulation.28 In China, similar dangers are apparent from the Chinese AI model DeepSeek’s refusal to answer queries about the 1989 Tiananmen Square massacre.29 As for AI providers without strong ideological motives, they are too exposed to political pressures to be trusted with outsized power over the flow of information.
Avoiding Synthetic Democracy
Policymakers can ward off any or all of these outcomes. A more critical eye on AI procurement and deployment by governments would be a reasonable starting point. Slower, more considered adoption, like that modeled by a pilot program for state-government workers in Pennsylvania, could reap AI’s potential efficiency gains without incurring its harmful consequences. Another common recommendation is to make risk and human-rights audits standard in AI services used by government agencies. Inequality could be addressed through numerous and specific tax, spending, and labor policies. Likewise, recommendations for improving the health of news media—through public support, philanthropic models, and citizen initiatives, for example—are common. The challenge is not a lack of policy levers or ideas.
Rather, democracies today suffer from a lack of political will. Policymakers are enamored of AI’s economic promises or fear the threat of rival governments’ winning the “AI race” and using that victory to assert dominance. Cautions regarding unequal economic outcomes, growing climate risks (AI data centers require staggering amounts of electricity), and overinvestment in unproven technologies have failed to lessen the sway that venture capitalists and tech industrialists have over policymakers. This cannot continue if the outcomes we fear are to be avoided. Yet democracy advocates start from a place of deep disadvantage against AI moguls who hold both the commanding heights of the economy and a formidable grip on elite opinion.
Two points must be driven home: First, AI is not an ineluctably dominant force grinding forward independent of human will and reason, but a technology of uncertain utility whose implementation is up to people to decide. Second, the processes of scientific investigation, government regulation, and public debate may take time, but this does not make them inferior to government by machine (meaning government by those who own the machine). Efforts to improve governance and expand knowledge may not always be improved when sped up by computers; rather, as we have seen with social media, increases in the speed and volume [End Page 147] of information can decrease the quality of dialogue and decisionmaking. Replacing ancient institutions such as universities, legislatures, and courts with technological quick fixes will concentrate power and information so intensely that the conditions which make democracy possible will be destroyed.
The social sciences and humanities are well placed to offer insights on these points and can provide important counterweights to the perspectives of industry professionals. The discourse surrounding AI is one more arena in which technical knowledge has become overvalued compared to other forms of expertise. Unfortunately, political attacks on universities and other sources of knowledge today are eroding the collective voice of professionals concerned about AI and the future of democracy. They must speak loudly now, while they still can.
Dean Jackson is a nonresident fellow at the Atlantic Council’s Digital Forensic Research Lab and the principal of Public Circle, LLC, a research consultancy focused on democracy, technology, and media.
Samuel Woolley is associate professor of communication and holds the William S. Dietrich II Endowed Chair in Disinformation Studies at the University of Pittsburgh.
NOTES
1. World Economic Forum, Global Risks Report 2024, January 2024, www.weforum.org/publications/global-risks-report-2024.
2. Sayash Kapoor and Arvind Narayanan. “How to Prepare for the Deluge of Generative AI on Social Media,” Knight First Amendment Institute at Columbia University, 16 June 2023, https://knightcolumbia.org/content/how-to-prepare-for-the-deluge-of-generative-ai-on-social-media.
3. Dean Jackson and Meghan Conroy, “Apocalypse Later? The Real Impact of AI on the 2024 Elections,” Atlantic Council, 17 October 2024, www.atlanticcouncil.org/content-series/the-big-story/apocalypse-later; Zelly Martin et al., “Political Machines: Understanding the Role of Generative AI in the U.S. 2024 Elections and Beyond,” University of Texas at Austin Center for Media Engagement, May 2024, https://mediaengagement.org/research/generative-ai-elections-and-beyond/.
4. Yoshua Bengio, “AI and Catastrophic Risk,” Journal of Democracy 34 (October 2023): 111–21; Tom Davidson, “The Danger of Runaway AI,” Journal of Democracy 34 (October 2023): 132–40; OpenAI, “OpenAI Charter,” 9 April 2018, https://openai.com/charter.
5. Bruce Schneier and Nathan Sanders, “The A.I. Wars Have Three Factions, and They All Crave Power,” New York Times, 28 September 2023, www.nytimes.com/2023/09/28/opinion/ai-safety-ethics-effective.html.
6. Samantha Shorey, “AI and Government Workers: Use Cases in Public Administration,” Roosevelt Institute, 15 July 2025, https://rooseveltinstitute.org/publications/aiand-government-workers.
7. Prithvi Subramani Iyer, “How Big Data Can Bolster Autocratic Legitimacy (Via the Rhetoric of Safety and Convenience),” Policy Brief no. 137, Toda Peace Institute, September 2022 (September 2022), 4, https://toda.org/assets/files/resources/policy-briefs/tpb-137_big-data-autocratic-legitimacy_iyer.pdf; Dia Kayyali, “Ask the Experts: AI Surveillance and U.S. Immigration Enforcement,” Tech Policy Press, 22 April 2025, www.techpolicy.press/ask-the-experts-ai-surveillance-and-us-immigration-enforcement.
8. Arvind Narayanan and Sayash Kapoor, “AI as Normal Technology,” Knight First Amendment Institute at Columbia University, 15 April 2025, https://kfai-documents.s3.amazonaws.com/documents/c3cac5a2a7/AI-as-Normal-Technology—Narayanan—Kapoor.pdf.
9. Eric Schmidt, “AI, Great Power Competition and National Security,” Daedalus 151 (Spring 2022): 288–98, www.jstor.org/stable/48662042.
10. Barton Gellman, “Peter Thiel Is Taking a Break from Democracy,” Atlantic, 9 November 2023, www.theatlantic.com/politics/archive/2023/11/peter-thiel-2024-election-politics-investing-life-views/675946; Jill Lepore, “The Failed Ideas That Drive Elon Musk,” New York Times, 4 April 2025, www.nytimes.com/2025/04/04/opinion/elon-muskdoge-technocracy.html.
11. Henry Farrell, “Should AGI-Preppers Embrace DOGE?” Programmable Mutter, Substack, 18 March 2025, www.programmablemutter.com/p/should-agi-preppers-embrace-doge.
12. Bruce Schneier, “Ten Ways AI Will Change Democracy,” Harvard Kennedy School Ash Center for Democratic Governance and Innovation, 6 November 2023, https://ash.harvard.edu/articles/ten-ways-ai-will-change-democracy.
13. Dean Jackson and Zelly Martin, “Forget Deepfakes: Social Listening Might Be the Most Consequential Use of Generative AI in Politics,” Tech Policy Press, 18 June 2024, www.techpolicy.press/forget-deepfakes-social-listening-might-be-the-most-consequential-use-of-generative-ai-in-politics.
14. Eric Schmidt, “Can Democracy Survive Artificial Intelligence?” Deseret News, 5 January 2025, www.deseret.com/opinion/2025/01/05/artificial-intelligence-and-democracy-eric-schmidt.
15. Sarah Kreps and Doug Kriner, “How AI Threatens Democracy,” Journal of Democracy 34 (October 2023): 122–31; Adam Mazmanian, “House Bill Targets AI Generated Comments in Rulemaking,” NextGov/FCW, 8 May 2024, www.nextgov.com/artificial-intelligence/2024/05/house-bill-targets-ai-generated-comments-rulemaking/396419.
16. James Fishkin et al., “Is Deliberation an Antidote to Extreme Partisan Polarization? Reflections on ‘America in One Room,'” American Political Science Review 115 (November 2021): 1464–81; Hugo Mercier and Dan Sperber, The Enigma of Reason (Cambridge: Harvard University Press, 2019).
17. D. Sunshine Hillygus, “The Evolution of Election Polling in the United States,” Political Opinion Quarterly 75, no. 5 (2011), 962–81, https://sites.duke.edu/hillygus/files/2014/06/HillygusPOQpolling.pdf.
18. Amartya Sen, The Idea of Justice (Cambridge: Harvard University Press, 2011).
19. David Rotman, “Technology and Inequality,” MIT Technology Review, 21 October 2014, www.technologyreview.com/2014/10/21/170679/technology-and-inequality.
20. Eli G. Rau and Susan Stokes, “Income Inequality and the Erosion of Democracy in the Twenty-First Century,” Proceedings of the National Academy of Sciences, 122, no. 1 (2025), e2422543121, www.pnas.org/doi/10.1073/pnas.2422543121.
21. Stephanie A. Bell and Anton Korinek, “AI’s Economic Peril,” Journal of Democracy 34 (October 2023): 151–61.
22. Moritz von Knebel, “When We Are No Longer Needed: Emerging Elites, Tech Trillionaires, and the Decline of Democracy,” Tech Policy Press, 8 May 2025, www.tech-policy.press/when-we-are-no-longer-needed-emerging-elites-tech-trillionaires-and-thedecline-of-democracy.
23. Courtney C. Radsch, “Can Journalism Survive AI?” Brookings Institution, 25 March 2024, www.brookings.edu/articles/can-journalism-survive-ai.
24. Casey Newton, “How AI Bots Are Suffocating Wikipedia,” Platformer, 1 April 2025, www.platformer.news/wikipedia-ai-bot-traffic-costs-plan; Nic Newman and Federica Cherubini, “Journalism, Media, and Technology Trends and Predictions 2025,” Reuters Institute, 9 January 2025, https://reutersinstitute.politics.ox.ac.uk/journalism-media-andtechnology-trends-and-predictions-2025.
25. Michael M. Grynbaum and Ryan Mac, “The Times Sues OpenAI and Microsoft over A.I. Use of Copyrighted Work,” New York Times, 27 December 2023, www.nytimes.com/2023/12/27/business/media/new-york-times-open-ai-microsoft-lawsuit.html.
26. Ilia Shumailov et al., “AI Models Collapse When Trained on Recursively Generated Data,” Nature 631, no. 8022 (2024): 755–59, www.nature.com/articles/s41586-024-07566-y.
27. Parshin Shojaee et al., “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity,” Apple, June 2025, https://machinelearning.apple.com/research/illusion-of-thinking.
28. Kylie Robison and Paige Oamek, “Elon Musk’s Grok AI Can’t Stop Talking About ‘White Genocide,'” Wired, 14 May 2025, www.wired.com/story/grok-white-genocideelon-musk.
29. James T. Areddy and Isabella Simonetti, “DeepSeek’s Chatbot Works Like Its U.S. Rivals—Until You Ask About Tiananmen,” Wall Street Journal, 30 January 2025, www.wsj.com/tech/ai/deepseek-chatgpt-tiananmen-square-efcd9938.
Copyright © 2025 National Endowment for Democracy and Johns Hopkins University Press