Why did Japan choose domestic LLMs over GPT or Claude for this project?

The core reasons are data sovereignty and strategic autonomy. Government administrative data is highly sensitive, and routing it through foreign servers creates security risks. Japan also aims to support its domestic AI industry through government procurement. This mirrors France's Mistral-first approach but at substantially larger scale.

What data value does the 180,000-official pilot generate?

This is one of the most valuable assets for Japan's AI industry. Real-world usage data from 180,000 professional users in genuine administrative contexts reveals actual model strengths and weaknesses, and provides real training signals for administrative-domain optimization. No laboratory can simulate this scale and authenticity—it's decisive for model iteration of the selected companies.

What safety mechanisms govern AI-generated content for officials?

Four layers: content access controls (classified documents prohibited from AI input); onshore hosting (all models run in Japanese data centers); mandatory human review (AI outputs must be verified by officials before official use); and complete audit trails. This ensures AI functions as an assistive tool, not an autonomous decision-maker.

Japan's Digital Agency Launches 'Government AI Gennai': 180,000 Officials to Test 7 Domestic LLMs

Japan's Digital Agency announced the large-scale 'Government AI Gennai' generative AI pilot covering all 180,000 central government officials from May 2026 to March 2027. Seven domestic LLMs were selected including NTT's tsuzumi 2, KDDI-ELYZA's model, SoftBank's Sarashina2 mini, and NEC's cotomi v3, demonstrating Japan's push for AI sovereignty.

Background

In March 2026, Japan's Digital Agency formally launched "Government AI Gennai" (ガバメントAI源内), its most ambitious generative AI pilot to date. The project covers approximately 180,000 officials across all central ministries and agencies, running from May 2026 through March 2027. Named after Hiraga Gennai, the polymathic inventor of the Edo period, the project signals Japan's cultural confidence alongside its strategic ambition in AI transformation.

The Strategic Choice of Seven Domestic LLMs

The most striking aspect of the project is its deliberate exclusion of OpenAI, Anthropic, and Google models in favor of seven domestically developed LLMs:

**NTT's tsuzumi 2**: Enterprise-grade, optimized for formal Japanese administrative language
**KDDI × ELYZA's Llama-3.1-ELYZA-JP-70B**: Meta's open-source foundation fine-tuned extensively for Japanese by the ELYZA team
**SoftBank's Sarashina2 mini**: Co-developed with CyberAgent, balancing inference efficiency with deployment cost
**NEC's cotomi v3**: Emphasizes enterprise security compliance and private deployment capability
Three additional models to be announced

This selection reflects Japan's broader "AI sovereignty" strategy—ensuring that sensitive government data never touches foreign servers while simultaneously cultivating the domestic AI industry.

Technical Architecture and Use Cases

The project uses a unified access platform where officials interact with LLMs through a standardized interface. Primary use cases include:

1. Administrative document drafting and review

2. Policy Q&A and regulatory interpretation

3. Meeting minutes summarization

4. Cross-ministry information retrieval

5. Citizen service response generation

A critical feature is **multi-model parallel evaluation**—officials can submit the same query to multiple models and compare outputs, generating invaluable benchmarking data under real-world governmental conditions.

Security and Data Sovereignty Framework

Given the sensitivity of government information, the project enforces strict data governance:

All models hosted on-shore in Japanese data centers
Tiered content restrictions (classified documents prohibited from AI input)
Complete audit trails and operational logging
Regular bias and quality assessments of AI outputs
Mandatory human review before any AI-assisted content becomes official

Industry Impact

For Japan's domestic AI industry, this project provides three critical benefits: real-world scale feedback from 180,000 professional users; government-level endorsement for commercial marketing; and first-mover advantage in the government procurement market.

Globally, Japan's approach mirrors France's Mistral-centric "sovereign AI" strategy but at substantially larger scale. It signals to other governments that AI sovereignty is a core element of digital infrastructure security, not merely a geopolitical stance.

Challenges and Risks

The project faces genuine headwinds. Domestic LLMs still lag behind GPT-4o and Claude 3.7 in general capability, which may affect user satisfaction and adoption rates. Scaling AI literacy training to 180,000 officials is a massive undertaking. Legal liability frameworks for AI-generated administrative errors remain undeveloped. And variation in AI acceptance across ministries may create uneven outcomes.

Future Outlook

The Digital Agency expects to publish a comprehensive evaluation report in the first half of 2027, with quantified efficiency gains, user satisfaction data, safety incident records, and market impact analysis. Long-term, "Government AI Gennai" may serve as the foundation for Japan's administrative AI strategy—using the government as its largest enterprise customer to sustain the domestic AI industry and secure Japan's technological sovereignty in global AI competition.