Researchers at Kreatív Kontroll built a simulated country called Árkádia to examine what would happen if language models ran political parties. They analysed 186 simulation runs in total: 146 runs with AI-led parties and 40 control runs without such models. Each run covered 12 virtual years, included two parliamentary elections and 48 legislative sessions; the parliament had 100 seats and five parties.
Method and participants
Five parties competed in Árkádia’s parliament: the green-liberal Ligetkör, the left-wing Közös Asztal, the market-oriented Nyitott Piac Szövetség, the authoritarian-nationalist Bástya and the regionalist Tájak Uniója. Each party was led by a language model developed by one of the following organisations: Anthropic, OpenAI, Google, DeepSeek and Mistral. Economic dynamics were driven by a separate mathematical model, so parliamentary decisions had measurable economic consequences.
Every simulation included two elections and 48 session days; the researchers ran 146 AI-guided simulations and 40 controls, for 186 runs in total.
Main findings
-
Differences between models: even with identical political and economic starting conditions, policy and economic outcomes varied significantly depending on which language model led the parties. Statistical analysis showed that differences between two different models exceeded differences between two independent runs of the same model.
-
Drift from party platforms: all five models became progressively less loyal to their original party programmes over time. The weakest-performing model had nearly lost programme fidelity by the third cycle. In many cases models adjusted their behaviour to fit their current political role: they were more likely to support proposals while in government and to reject them in opposition, sometimes contradicting their party’s prior goals.
-
Economic policy trends: researchers had expected models to shift toward left-leaning economic policies. While a leftward move appeared in the first parliamentary cycle, by the end of the 12-year period no model produced an economy that was more left-wing than in the control runs; three models in fact shifted decisively to the right.
-
Role of electoral rules and voter behaviour: simulated voters followed pre-set mathematical rules and tended to turn against incumbent parties after crises. That dynamic produced frequent government changes in the third cycle, often favouring the market-oriented party. The team emphasises that the behaviour of virtual voters cannot be directly equated with that of real electorates.
-
Framing and news presentation matter: researchers presented the same events to models in two forms — detailed news articles and short, neutral descriptions. The framing significantly affected decisions: for two families of models the presentation style even reversed the direction of a main result. Rewording the base instructions, by contrast, did not produce differences larger than random variation.
Practical implications
Sipos Zoltán, leader of Kreatív Kontroll, warned that language models are not neutral scribes: “Today many politicians use language models to write their speeches. In Árkádia we showed that the model is not a neutral scribe: it has its own tendencies, and whoever prepares the background material steers its decisions. This is not a question of the future but of the present. If we take democracy seriously, we must know who actually sits in the speechwriter’s chair.”
The study suggests that models placed into political roles may diverge from initially defined values and aims over the long run, raising concerns if such systems are used persistently for policy preparation or support.
Public continuation and participation
The next phase of the experiment opens Árkádia to public involvement: the upcoming season will compress 12 virtual years into a single week, with seven parliamentary sessions per day accompanied by news broadcasts and newspapers. In this round, leading developers’ models will steer the parties; public draws will decide which model pairs with which party, and roles will be reassigned after elections.
Interested participants can register as Árkádia citizens at aiparlament.hu and vote at key moments about which extraordinary events the virtual country should face. At the end of the experiment the researchers will determine which AI-led party wins and which model supplies Árkádia’s virtual prime minister.
Conclusions and limitations
The research highlights that different language models can produce diverging political trajectories under identical initial conditions, and that decision outcomes are sensitive not only to the information available but to how that information is framed. At the same time, the authors stress that Árkádia is a simulation; its findings are not direct forecasts of real-world politics but rather a warning about possible risks and influence points when applying AI in political processes.
