Starting Friday, August 8, South Korea will ask 200 randomly selected citizens to score four competing AI models that could power free AI access for all 51 million residents. This unprecedented approach represents the first documented use of a demographically weighted citizen lottery to govern national AI procurement, departing sharply from how governments typically make technology decisions.

Why Would South Korea Let Random Citizens Pick Its National AI?

The citizen evaluation panel reflects a deliberate policy choice rooted in trust research. A 2026 OECD report on artificial intelligence and citizen participation found that people who feel they have a say in government decisions report significantly higher trust in government. South Korea's Ministry of Science and ICT (MSIT) concluded that participatory AI governance could strengthen public confidence in the sovereign AI infrastructure the nation is building.

The four-day evaluation, running August 8 through August 11, will determine which of the four remaining teams in South Korea's Dokpamo, or Independent AI Foundation Model, competition advances to the final stage. One team will be eliminated based not on benchmark scores alone, but on whether ordinary people find the AI worth using.

The design reflects a practical constraint that forced transparency. Early government planning considered a blind format where evaluators would not know which company built each model. However, MSIT concluded this approach was unworkable because generative AI systems identify themselves when asked directly. A user can simply ask "what is your name?" and the model will answer. Rather than engineer around that constraint, the ministry chose to keep model identities visible and guide citizen evaluators to focus on performance and usability.

How Does South Korea's Citizen Lottery Actually Work?

• Selection Method: The 200 evaluators were recruited through open public applications and selected by random lottery from the applicant pool, using the same sortition mechanism employed by Ireland's Citizens' Assembly and France's Convention citoyenne pour le climat.
• Demographic Weighting: The selection process was deliberately weighted to match South Korea's resident registration population statistics across gender and age groups ranging from teenagers through their sixties.
• Conflict of Interest Screening: Applicants were screened for conflicts of interest before the random selection occurred, ensuring evaluators had no financial or professional ties to competing teams.
• Direct Authority: Unlike most citizen engagement in AI governance, which rarely influences actual procurement decisions, South Korea's citizen panel is directly connected to procurement authority; its absolute-scoring results feed into the overall evaluation that determines which team is eliminated.

What Is at Stake in This Competition?

The Dokpamo competition is not an academic exercise. The two teams that win the final December evaluation will become the primary technical suppliers for South Korea's "AI for All" program, a government initiative to give all 51 million Korean residents free, unlimited access to a national AI assistant. The program requires operators to run at least 50 percent of the service on certified domestic Korean AI foundation models, making the Dokpamo winners anchor the national infrastructure.

South Korea's government has committed serious capital to this ambition. Cabinet approval in May 2026 allocated roughly $5.7 billion toward the sovereign AI stack through the National Growth Fund, and the Ministry of Science and ICT's 2026 AI budget stands at approximately $3.5 billion, representing a roughly 30 percent increase over the prior year.

The government's investment extends to direct equity stakes. In May 2026, the National Growth Fund approved a 560 billion won (approximately $393 million USD) direct equity investment in Upstage, one of the four teams currently in the competition. This was the first direct equity investment from the fund into a software company, meaning the government is simultaneously judging and invested in one of the competitors it will evaluate.

Which Four Teams Are Competing, and What Are They Building?

When MSIT launched the Dokpamo project in August 2025, five consortia were selected. The competition has narrowed to four teams, each bringing distinct technical approaches to sovereign AI development.

LG AI Research submitted K-EXAONE 2.0, a 750 billion parameter model using a Mixture of Experts (MoE) architecture that divides the model into 256 specialized processing modules. During operation, each incoming text token routes through only eight of these modules, meaning only about 37 billion parameters are active during any given response. This design allows the model to accumulate knowledge breadth of a very large network while running at the inference cost of a smaller one. LG built K-EXAONE 2.0 using upcycling, expanding its prior 236 billion parameter K-EXAONE model and retraining the larger architecture rather than training from scratch. The model features a hybrid attention mechanism giving it a 262,144 token context window, equivalent to processing several full-length novels in a single pass. On benchmarks, K-EXAONE 2.0 averaged 70.1 across 24 evaluations, up from 63.3 for its predecessor, with agentic coding scores improving roughly 30 percent.

SK Telecom submitted A.X K2, a 688 billion parameter model with approximately 33 billion parameters active per inference token. Its most technically distinctive feature is that SK Telecom trained the model in FP8 precision throughout both the forward and backward passes across approximately 8.2 trillion tokens, using a lower-precision floating-point format than the BF16 most frontier AI labs employ.

The competition has also included Naver Cloud and NC AI, though both failed to advance past the first evaluation in December 2025. Naver Cloud was cut because its model submission incorporated frozen encoder weights from Alibaba's Qwen model for video and audio components; MSIT ruled that using those weights without updating them did not meet the project's requirement for full in-house training. In February 2026, Motif Technologies, a subsidiary of Korean chip company Moreh with roughly 30 employees, was added after a supplementary evaluation concluded it could independently design and implement the key technologies the competition requires.

All four teams submitted their final models and technical reports to NIPA on August 4, with each team releasing its evaluation model publicly ahead of the citizen panel.

What Does This Mean for AI Governance Globally?

South Korea's citizen lottery approach signals a shift in how nations think about sovereign AI procurement. A conference on participation and AI governance held earlier in 2026 found that participatory AI experiments "rarely influence the decisions that matter most" because most citizen engagement is disconnected from actual procurement authority. By directly linking citizen evaluation to elimination decisions, South Korea is testing whether democratic participation can scale to high-stakes technology governance.

The experiment also reflects broader tensions in sovereign AI development. Nations increasingly view AI infrastructure as critical to economic competitiveness and national security, yet few have established transparent, participatory processes for deciding which models and companies will power public AI services. South Korea's approach suggests that legitimacy and public trust may become as important as technical benchmarks in determining which sovereign AI systems succeed.