Gennai OSS: Is Published Code Enough for Deployment?

The cover is an AI-generated conceptual image. It is not a photograph of a government system or office.

What Is Happening

April’s partial release and the planned February 2027 release differ; 180,000 is a trial target.

Separating the April release from the plans published in September

On April 24, 2026, Japan’s Digital Agency released part of Gennai, its generative AI environment for government work. The release covered code and setup instructions for the web interface and selected implementations and development templates for administrative AI applications. Open-source software, or OSS, can be used and modified under its license conditions. This was not an announcement that everyone would receive an account for the government-operated service.

The subsequent September 18 report（外部サイト、新しいタブで開きます） said OSS Ver 2.0 was scheduled for release around February 2027. The government planned to introduce an environment where staff could build and use AI agents around October 2026. AI agents combine tasks such as searching and creating materials in response to instructions. These are separate schedules. Demonstrations of features under development and the holding of a briefing do not establish that municipalities can already obtain the same materials.

The policy page（外部サイト、新しいタブで開きます）, updated on October 9, lists introductory, beginner and intermediate agent-chat videos dated that day. Guidance on using an interface and release of OSS materials to external organizations are separate events. Listing these videos does not by itself establish that the internal government rollout is complete or that municipalities have started using the environment.

Comparing the April announcement, September report and the README (the project’s usage guide), covered-license list and logging documentation in each public code repository shows that “Gennai has been released” can combine available components with future release plans. This article examines the April and September publications alongside the policy page updated on October 9, 2026 to ask what adopting organizations receive and which decisions they must make. It does not report an independent deployment’s performance or costs, or an audit of a fixed version of all the code.

The trial population is separate from actual users

The May 28 announcement（外部サイト、新しいタブで開きます） said approximately 100,000[1] government staff would be able to access it the next day, with gradual expansion to a target population of approximately 180,000[1] across ministries and agencies. The latter is the trial’s target population, not a count showing that everyone used it daily. Availability, actual use and continued use by the same people describe different things.

For example, more people asking a question once would not necessarily establish shorter task times or more accurate answers. Conversely, a smaller user group might find it useful for a particular task. Neither is a result established in this review; they are separate questions for evaluating the trial. Calling the entire trial population users would turn a target population into a count of actual use.

Partial April release

User interface

Web code

Administrative apps

Code/template

Use conditions

Per component

Planned Ver 2.0

Timing

Feb 2027

Agent environment

In progress

Sharing arrangements

Concept stage

Based on the Digital Agency’s April 24 announcement and September 18 report. Planned Ver 2.0 materials are under development and subject to change. The internal government agent rollout planned around October 2026 is separate from release of the OSS materials.

The September report describes the Gennai marketplace as materials for sharing staff-created agent skills across organizations. Skills are instructions and procedures telling an agent how to perform work. The marketplace remains at the planning and concept stage.

Comparisons of headcounts also need matching publication dates and populations. Staff becoming eligible at different times would have different observation periods, so the May availability figure cannot establish October usage or identify non-users by subtraction from 180,000[1]. If municipal staff evaluate adoption, they would need separate records for the population eligible to use it and the time spent checking answers. Records tracking the same population would help distinguish reach from evidence of better work.

Background & Context

Interface, applications, models and cloud differ; in-house parts do not end outside contracts.

Which parts were developed in-house?

Calling Gennai simply AI built by the government can imply that everything, from its interface to its answer-generating model, was independently developed. The developers’ account（外部サイト、新しいタブで開きます） says Gennai Web builds on AWS’s published Generative AI Use Cases application, or GenU, adding functions such as team management and application calls. The interface can call administrative applications built separately through a specified API format. An API is an interface through which systems exchange information.

The staff-facing interface, applications that find task-specific materials, the large language model generating text, and the cloud running them have different roles. For example, a municipality could build an interface for searching documents while using an external model to generate answers. The existence of internally developed components does not establish that external service contracts have disappeared.

The official AI application repository（外部サイト、新しいタブで開きます） includes a template for deploying a model on Azure, an implementation for consulting legislation on Google Cloud, and a template for searching government reference materials on AWS. Retrieval-augmented generation, or RAG, retrieves material to inform generated answers. Applications for different clouds do not establish that the entire Gennai Web system can be moved unchanged to any cloud.

Access to the published interface and applications can help procuring organizations distinguish existing functions from those their work requires them to add. Providers also have material for deciding whether to rebuild common functions or focus on task-specific documents and procedures. Yet a shared interface does not imply shared materials or permissions: each organization must check who may search whose documents.

Domestic models also require a distinction between evaluation plans and results

The domestic model trial announcement（外部サイト、新しいタブで開きます） describes providing models from NTT DATA, Fujitsu and Preferred Networks on SAKURA Cloud, with repeated comparison experiments planned from September through November. Users would see outputs without model names and indicate their preferences. The announced method needs to be distinguished from results obtained after the experiments.

When a municipality considers changing its model, domestic models and clouds can also be candidates for its configuration. However, user preference, task accuracy, operating cost and portability each require different evidence. A preference comparison alone does not measure quality across every administrative task. The Digital Agency’s policy goal of autonomy should likewise be assessed through records showing which parts can actually be changed.

Reading the Structure

Some files use non-MIT licenses; adopters must check changes, maintenance and data conditions.

More published code does not remove license conditions

This article provides general information, not individual legal advice. Assess compliance with licences and contracts against the actual terms, consulting a qualified legal professional where necessary.

The Gennai Web README（外部サイト、新しいタブで開きます） specifies MIT licensing for software in general and CC BY 4.0 for documentation. MIT permits use, modification and redistribution subject to conditions including retention of copyright and permission notices; CC BY requires attribution and identification of changes, among other conditions. Some files developed with AWS support instead fall under the Amazon Software License, or ASL. Describing everything as MIT omits conditions relevant to migration decisions.

The covered-file list（外部サイト、新しいタブで開きます） explains that those parts cannot run unchanged outside AWS and must be reimplemented for another environment. The corresponding ASL text（外部サイト、新しいタブで開きます） also restricts use in relation to Amazon-provided services and platforms. Whether code can be read, modified or moved unchanged should be checked component by component.

This comparison does not support a binary claim that OSS means freedom while foreign components mean dependency. It supports identifying the components requiring changes and their owners before organizations adopt the same published code. In addition to cloud costs, the adopting organization’s burden depends on who implements, tests and maintains those changes.

Public repositories change, so recording the adopted code and documentation versions together with the responsible staff helps compare the original conditions with later updates. Reviewing differences and test results remains necessary for determining whether existing settings or connections must change, as well as for adding new features.

Conditions connecting government operation with an organization’s own deployment

The official commercial-use explanation（外部サイト、新しいタブで開きます） describes providers offering a service, organizations procuring their own deployments, and the sale of AI applications. These are contemplated models, not established adoption results or exclusive procurement routes. Choosing a service and commissioning a deployment also involve different counterparties when checking incident contacts and update responsibilities in a contract.

The policy page（外部サイト、新しいタブで開きます） describes support for confidentiality category 2 inputs in the environment operated by the Digital Agency. The transport ministry’s explanation describes category 2 as information likely to qualify for nondisclosure, excluding the secret-document category. Placing the public code elsewhere does not establish equivalent conditions for handling the same data. The deployed environment needs checks of what is sent to which model, where it is stored, and who can access it.

The official logging document（外部サイト、新しいタブで開きます） says enabling the function automatically collects and sends application logs, execution histories and authentication-system user information, among other records, every day. A configurable function and the actual storage and transmission behavior of a deployment require separate checks. Organizations would need to define the purpose of using work records for evaluation, along with their storage destination, readers and retention period. Users of a provider’s service should also check contractual terms for returning or deleting data and records when use ends.

Use a provider’s service

Incidents and updates

Contract

Inputs and storage

Service terms

Records at termination

Return/delete

Organization’s deployment

Incidents and updates

Assign roles

Inputs and storage

Test locally

Published code changes

Check changes

These are the article’s proposed checks based on the contemplated adoption models and published specifications. They are not findings from examining a particular contract or deployment. Both models require checking each component’s license conditions.

Issue reporting and maintenance commitments

The April release announcement explicitly says necessary fixes will continue for the time being, while permanent maintenance is not guaranteed and publication may end. The README accepts only critical issue reports, does not guarantee responses to every issue, and does not accept Pull Requests, which submit proposed code changes. The briefing FAQ (frequently asked questions) describes accepting bug and security reports and having made fixes, differing from the README’s description of its reporting scope. Beyond reporting to the publisher, adopting organizations need their own repair responsibilities and procedures for checking a change’s effects.

In the briefing’s FAQ（外部サイト、新しいタブで開きます）, the Digital Agency also says execution without an internet connection has no settled development plan and connections via networks such as LGWAN remain under consideration. LGWAN（外部サイト、新しいタブで開きます） is an administrative network interconnecting local governments’ internal networks. An answer saying OSS users are unrestricted does not establish a finished connection for the organization.

Accessibility concerns whether people can use a service regardless of disability or other circumstances. The FAQ says expert checks have been conducted, while formal policy development and public testing remain ahead; the README notes remaining issues in part of the post-generation image controls. Sharing code and design components does not establish that an organization’s users can operate added functions. Assigning responsibility for checks that include screen readers and keyboard operation, alongside document search and answer use, can clarify where users report difficulties and who fixes them.

Checks should extend beyond initial adoption. Assigning someone to test again when documents or models change can help the adopting organization make decisions while awaiting the publisher’s response. Sharing conditions for suspending use and checks required before resuming also helps prepare explanations for users.

Related guide

Because starting with task needs also matters for municipalities, see the guide to AI adoption for NPOs（このサイトの記事）.

Reference Books

For related reading, see AIガバナンス入門: リスクマネジメントから社会設計まで (ハヤカワ新書) (Introduction to AI Governance: From Risk Management to Social Design)（外部サイト、新しいタブで開きます）. The publisher’s introduction（外部サイト、新しいタブで開きます） describes Hiroki Habuka’s book as addressing AI benefits and risks, social design, and AI literacy. This article checked that public introduction and does not quote the book’s body or chapter conclusions. The official documents above support its account of Gennai’s specific conditions.

The English title follows the author’s Kyoto University profile（外部サイト、新しいタブで開きます）; Hayakawa Shinsho identifies the publishing series. The book is in Japanese.

Remaining Questions

Future releases, evaluation of use and work quality, and maintenance changes still need evidence.

Gennai’s release enables adopting organizations to read code tailored to administrative work. The next materials to examine for planned releases are the license conditions and operational documentation published with them. Recording the adopted version and maintaining dated comparisons between the environment used first within government and the materials available to external organizations would help users decide what to prepare and when, including after schedules change. The Ver 2.0 announcement should not be read as an inventory of features already obtainable.

Evaluating the trial requires more than its target population: which staff used it, for which tasks, for how long, and what they checked before adopting an answer all matter. A count of questions records operations; it does not establish answer accuracy. Can an organization match the population and period to its evaluation questions, and include time spent correcting errors and reasons for non-use in its decision to continue?

If the publisher’s maintenance policy and external service contracts change separately, can adopting organizations choose to maintain their deployments or move elsewhere? That decision requires knowing which components can be changed and how data can be retrieved, and assigning responsibility for checking access rights and records during migration. After comparing the publisher’s work with its own responsibilities and assessing the staff and funding needed for routine repairs and migration, can the organization explain suspension and resumption conditions even when personnel change? Where can users report difficulties, and how would those reports inform later updates? Who would retain the decision records so that successors can trace the same evidence? These questions concern the adopting organization’s continuity, beyond the scope of the released code.

References

Partial open-source release of Government AI Gennai — Digital Agency (2026). Digital Agency

Report on the nationwide briefing on Gennai OSS Ver 2.0 plans — Digital Agency (2026). Digital Agency

Launch of the large-scale Gennai trial targeting approximately 180,000 government staff — Digital Agency (2026). Digital Agency

Gennai Web development overview — Digital Agency (2026). Digital Agency

Gennai AI applications repository — Digital Agency (2026). Digital Agency

Trial of domestic foundation models on a domestic cloud in Gennai — Digital Agency (2026). Digital Agency

Gennai Web README — Digital Agency (2026). Digital Agency

ASL-covered files — Digital Agency (2026). Digital Agency

Amazon Software License — AWS (2008). AWS

Contemplated commercial uses of Gennai OSS — Digital Agency (2026). Digital Agency

Government AI Gennai policy page — Digital Agency (2026). Digital Agency

Instructions for Enabling Log Transmission — Digital Agency (2026). Digital Agency

Gennai OSS Ver 2.0 briefing FAQ — Digital Agency (2026). Digital Agency

MIT License for Gennai Web — Digital Agency (2026). GitHub

Attribution 4.0 International (CC BY 4.0) deed (license version release year) — Creative Commons (2013). Creative Commons

The CC BY 4.0 year identifies the license version’s release, not the update date of this deed page.

Topics – Compliance Pocketbook — Ministry of Land, Infrastructure, Transport and Tourism, Kyushu Regional Development Bureau. Kyushu Regional Development Bureau

Press release: Creative Commons launches Version 4.0 of its license suite — Creative Commons (2013). Creative Commons

The CC BY conditions are available in Japanese（外部サイト、新しいタブで開きます） and English（外部サイト、新しいタブで開きます）.

Statistics cited in this article

1. 1Digital Agency, announcement of the large-scale Gennai trial（May 28, 2026） Open source

Related Content

October 16, 2026NEW

Can majority support alone bring about Security Council reform?

The September 2026 G4 statement, proposals to expand membership, and the adoption and ratification of Charter amendments

International AffairsNational SecurityGovernance

August 7, 2026

Before Restricting Children's Social Media by Age — How to Establish Age Is Still Sitting in the Open-Questions Column

In a paper issued in April 2026 by Japan's Children and Families Agency, whether to require age verification from social media operators still appears as a question. The methods listed are age data supplied by mobile carriers and estimation by AI, both at the stage of assessing feasibility. A rule that draws a line at an age needs, before the line is drawn, a means of telling which side someone is on. This piece reads the debate through where duties sit and how identity is established.

EducationDigital & AIPolicy Analysis

August 6, 2026

Fraud Losses in Japan Rose 98% in a Year While Cases Cleared Stood Still — Most Reports From People in Their 30s, Most Money From Those in Their 70s

Confirmed figures published by Japan's National Police Agency in May 2026 put losses from remote-contact fraud in 2025 at 142.31 billion yen, up 98.0% on the previous year, close to double. Cases cleared came to 6,575, down 0.0%. Victims aged 65 and over fell to 51.3% of reported cases, down 14.0 points. This piece reads how the choice between counting cases and counting money changes where effort is sent.

Social IssuesPolicy AnalysisDigital & AI

August 4, 2026

Toward a Right Not to Have Your Voice Used — Japan's Copyright Act Protects the Performance, Not the Voice

On 27 July 2026 an expert panel at Japan's Ministry of Justice broadly agreed a draft report extending the publicity right to voice as it already covers likeness. But the Copyright Act protects a performance rather than a voice, and the publicity right has no statutory basis at all. When generative AI imitates a voice without recording any performance, no branch of law catches it squarely. This piece asks who gets to draw the outline of a right that is not written down.

AILegal & RegulatoryPolicy Analysis

Get new columns by email

1-2 social structure analysis columns per week. Free to subscribe.

Website

Subscribe→

Join ISVD's activities?

Sign up to receive the latest research and activity reports. Feel free to reach out about collaboration or project participation.

Join→Contact Us→

Support ISVD’s research and activities

Support ISVD’s research and publications with JPY5,000 per year. There is no automatic renewal.

Become a supporting member→About supporting membership→