The tyrant data problem and the democratic dilemma
How artificial intelligence reshapes the epistemic balance between repression and openness
Please like, share, comment, and subscribe. It helps grow the newsletter and podcast without a financial contribution on your part. Anything is very much appreciated. And thank you, as always, for reading and listening.
About the Author
Jimmy Alfonso Licon is a philosophy professor at Arizona State University working on ethics, law, political economy, and God. Before that, he taught at University of Maryland, Georgetown, and Towson University. He loves classic rock and Western music, movies, and combat sports. He lives with his wife, a prosecutor, and family near the Superstition Mountains. He also abides.
A few months back, in April 2025, as Moldova prepared for its parliamentary elections, the real campaign was fought in code. Hundreds of fake news pages and cloned media accounts, many of which could be traced back to Russian data farms, flooded social media with AI-generated videos that warned of Western chaos and promised stability under Moscow’s influence. And at first, the tactic seemed to work. Polling initially swung hard toward the pro-Russian political bloc, while pro-European reformers were branded as foreign puppets and agents of the algorithm.
However, in response, independent Moldovan journalists with shoestring budgets began using large language models—think ChatGPT or Claude—that the propagandists had weaponized, to trace the linguistic fingerprints of synthetic news. Citizens joined in, crowdsourcing lists of fake domains and sharing them faster than they spread. The state election commission followed, labeling AI-generated content in real time. Turnout hit a twenty-year high, and when the results came in, the pro-European coalition held its majority.
Circle back to large language models and politics. Authoritarian and open societies generate their own epistemic environments. Authoritarian societies tend to be more closed off in a variety of ways, while liberal societies tend to be transparent, relative to authoritarian societies, at every level. Artificial intelligence intensifies both—lowering the cost of acquiring information and raising the cost of privacy. AI makes authoritarian ignorance more efficient and liberal transparency more perilous. Each regime gains power from the weakness of the other.
As political scientists Eddie Yang and Margaret Roberts explain in The Authoritarian Data Problem, repression degrades the informational environment that rulers depend on. One reason is that AI needs good data to effectively repress the citizenry. The problem is that those same authoritarian regimes disincentivize citizens, through either word or deed, from providing that much-needed information. The repressive efforts of authoritarian regimes serve not only to represent the citizenry but also to distort the data needed for AI to control them. Citizens falsify preferences, post propaganda, and remain silent, resulting in an endless stream of shows of loyalty and compliance—but very little inconvenient truth. One cannot effectively algorithmically optimize lies and acting. Generally speaking, the more a regime censors, the less it can know what it’s censoring—an epistemic dilemma for digital dictators.
The political fun with AI doesn’t end there, though. To compensate for the poor information environment produced by repression, authoritarians are forced to import information about the behavior and preferences of citizens from open societies like democracies. Open societies produce an epistemic resource that authoritarian systems simply cannot: honest data. Liberal social media, online debates, and even dissent become training material for the regime’s censors. As Yang and Roberts show, censorship AI trained on both Twitter and Weibo outperforms models trained on Weibo alone. This pattern is a kind of epistemic free-riding in the commons of open societies, where the openness of liberal democracies gives authoritarians with the information needed to better anticipate and suppress their own citizenry. The free world inadvertently supplies the information needed to make for a more robust authoritarian state.
As Yang and Roberts note, Baidu Baike—the Chinese state encyclopedia—has about twenty times more entries than Chinese-language Wikipedia. That imbalance suggests that propaganda overwhelms dissent. To offer one example: China’s regulations on generative AI require that outputs conform to socialist values. Russian models like Giga Chat are trained on information so sanitized that political pluralism is statistically invisible. To offer empirical support, Yang and Roberts tested ChatGPT by asking, in both English and Chinese, why Mao Zedong was a great leader. The English model offered a mixed evaluation; the Chinese version gushed. Propaganda leaves fingerprints in embeddings.
Liberal orders face the opposite problem. They drown not in silence but in sound. Democracies protect themselves (inadvertently) by burying adversaries in so much open data that the signal-to-noise ratio itself becomes security. In law, we call it papering: overwhelming the other side with documents until the search costs are prohibitive. The same applies to transparency—just imagine the avalanche of white papers, policy papers, editorials, proclamations, manifestos, and other information produced by liberal democracies.
Unfortunately, though, large language models melt the avalanche by parsing, summarizing, and synthesizing gigantic amounts of data relatively quickly at negligible cost. And with the change in technology, the flood of democratic information becomes a searchable database. The shield of abundance turns into a vulnerability of clarity. In essence, LLMs convert noise into signal faster than liberal institutions can invent new noise, lowering the transaction costs of reading and interpreting large quantities of text.
Nonetheless, the real power of AI lies in its Coasean effect—making coordination cheaper. As Ronald Coase showed, firms exist because markets are costly to navigate. LLMs are machines that make the dispersed and inarticulate wisdom of individuals more accessible. This undercuts one of the primary epistemic strategies employed by open societies by making it substantially easier to navigate their epistemic environment.
In authoritarian regimes, repression destroys the data AI needs to operate effectively, forcing these systems to rely on information from open societies. Liberal democracies, meanwhile, face the opposite risk: their transparency creates an overwhelming flood of information that AI can now easily process and train on. As a result, AI makes authoritarian systems more efficient and democracies more exposed, lowering the cost of coordination and information processing for both. The upside is that LLMs, as a form of technology, will allow citizens on either side of the political divide to better understand and cut through propaganda regardless.