📊 A mountain of public opinions that once took government officials over a month to sort through— processed by AI in just 10 minutes. Fujitsu's large language model "Takane" has successfully automated the classification and summarization of approximately 120,000 characters of public comment data in a proof-of-concept with a Japanese central government agency. Japan's government AI adoption is shifting from "exploration" to "real-world deployment."
What Are Public Comments? The Bridge Between Government and Citizens
In Japan, the public comment system (formally known as the "Opinion Solicitation Procedure" under the Administrative Procedure Act) requires government agencies to publicly disclose draft regulations—such as cabinet orders and ministerial ordinances—and solicit opinions from the general public before finalizing them. It serves as a critical mechanism for ensuring transparency and fairness in policymaking.
However, the system has long faced operational challenges. Topics of high public interest can generate thousands or even tens of thousands of submissions. Government officials must read every opinion, classify them by topic and stance, draft individual responses, and consider how to reflect them in policy. The entire process often takes over a month before results are published.
What Fujitsu's "Takane" Actually Achieved
On February 3, 2026, Fujitsu announced the results of a demonstration experiment conducted in collaboration with a central government agency, applying its LLM "Takane" to public comment operations.
Takane is an enterprise-grade LLM co-developed by Fujitsu and Canadian AI company Cohere. It has achieved world-leading performance on the JGLUE (Japanese General Language Understanding Evaluation) benchmark and is designed for use in secure private environments—critical for sectors like government, finance, and R&D where data confidentiality is paramount.
The proof-of-concept, conducted in 2025 using actual past public comment data (approximately 120,000 characters) from the participating agency, yielded the following results:
- Automated classification and summarization: Tasks such as sorting opinions by support or opposition and generating summaries—previously done manually—were completed in approximately 10 minutes
- Cross-referencing with draft legislation: When both the draft law and individual opinions were fed into Takane, the system correctly identified the relevant legal clauses for over 80% of the submitted opinions
- Freeing officials for higher-value work: By reducing time spent on organizing and tabulating opinions, officials can redirect their efforts toward substantive tasks like evaluating opinion content and incorporating feedback into policy
Alignment with Japan's Digital Agency Initiatives
This experiment is part of a broader movement driven by Japan's Digital Agency, which has identified public comment aggregation and analysis as a key area for generative AI application. The initiative also ties into EBPM (Evidence-Based Policy Making) efforts, where efficient digital analysis of citizen feedback is seen as essential for improving policy quality.
Building on these results, Fujitsu is developing a generative AI service that goes beyond public comment processing to support the entire spectrum of policy formulation and legislative drafting, with a target launch by fiscal year 2026. Future plans include constructing AI workflows that systematically integrate appropriate AI models throughout the legislative process, as well as developing AI agents capable of autonomously supporting complex research and coordination tasks.
Why a Japanese-Specialized LLM Matters
Most leading LLMs have been developed primarily for English. Japanese, however, presents unique challenges: mixed character systems (kanji, hiragana, katakana), frequent subject omission, and complex honorific expressions. Processing public comments that contain administrative jargon and legal terminology demands strong Japanese contextual understanding.
There is also a practical constraint: general-purpose cloud-based LLMs are often unsuitable for handling sensitive government data that cannot leave secure environments. Takane's design for private deployment gives it a significant advantage in government adoption scenarios.
Does AI-Powered Public Comment Processing Strengthen Democracy?
The original purpose of the public comment system is to reflect diverse citizen voices in policymaking. In practice, however, the sheer volume of submissions can overwhelm officials, leading to concerns about the system becoming a formality rather than a genuine feedback mechanism.
If AI-driven classification and summarization become standard practice, officials could be freed from organizational busywork to focus on genuinely engaging with the substance of public opinion. There is also the potential for AI to surface minority viewpoints and novel perspectives that might otherwise be buried in massive datasets—potentially improving the quality of democratic processes.
At the same time, there are legitimate concerns. AI summarization could strip away nuance from individual opinions. Classification biases could subtly influence policy decisions. The principle that final policy judgments must remain with human officials is essential, with AI serving strictly as a support tool.
Japan is steadily advancing government digitalization and AI adoption. Does your country use AI technology when gathering citizen input on policy? How do public participation mechanisms like public comment periods work where you live? We'd love to hear your perspective.
References
Reactions in Japan
I honestly didn't know it took over a month to process public comments. No wonder the system feels like a formality. If AI can classify them in 10 minutes, there's no reason not to adopt it. It's a better use of taxpayer money too.
What scares me is having AI 'summarize' opinions. If dissenting views get rounded down to 'some opposition was noted,' the whole point of submitting comments is lost. Who verifies the accuracy and fairness of the summaries?
Former government ministry employee here. Tallying public comments was absolute hell. Reading and classifying each one until late at night. If this lets young bureaucrats spend more time on actual policy work, I genuinely welcome it.
Processing 120K characters in 10 minutes sounds impressive, but the Administrative Procedure Act explicitly states that the 'content' of opinions matters, not the quantity. I'm concerned about AI overweighting majority views.
Nice name, Takane—reminds me of Japanese mountain peaks. Being JGLUE top-scorer on a Cohere base means it's pretty reliable for Japanese processing. I'm curious if it handles the unique phrasing of administrative documents well.
I bet officials will just rubber-stamp whatever AI classifies. I can already see a future where they say 'the AI determined this.' Efficiency and quality of democracy are separate issues.
I work in municipal IT and our local government struggles with public comment processing too. If this proves successful at the national level, it should trickle down to local governments. My biggest question is the implementation cost.
Over 80% accuracy in matching opinions to legal clauses—honestly, shouldn't it be higher? What if the remaining 20% contains critical opinions? For production use, I think 95%+ is necessary.
In the context of EBPM, this is a hugely important step. If public comment data can be structured by AI, it becomes possible to analyze which demographics are commenting on which issues. It could fundamentally change how we evaluate policy.
What we want AI to do is 'organize' opinions, not 'judge' them. Classification and summarization are fine, but if AI starts deciding 'this opinion should be reflected in policy,' that's the end. Drawing the line is crucial.
The key point is that it runs in a private environment. You can't feed government data into ChatGPT. In that sense, Takane's design philosophy is well-suited for government use.
Shouldn't the priority be making it easier to submit public comments in the first place? The Digital Agency's e-Gov portal is hard to use and awareness of the system is low. Streamlining processing with AI is meaningless if the entry point stays narrow.
Interesting that Fujitsu partnered with Cohere. NTT has tsuzumi, NEC has cotomi—the domestic LLM competition is finally being tested in actual government operations. Who wins will depend on real-world results.
Aiming for service launch within FY2026 is pretty fast by government decision-making standards. Could be proof that the Digital Agency is seriously pushing Government AI initiatives.
If we're processing public comments with AI, the process itself needs to be transparent. What logic was used for classification? What's the basis for summaries? Pursuing efficiency while keeping it a black box is dangerous and incompatible with government accountability.
As someone involved in federal rulemaking in the US, this is noteworthy. FCC public comments can reach millions, and manual processing has completely broken down. Ironic that Japan is leading with AI here, but there's a lot we should learn.
In France, the Grand Débat National also took enormous time to organize citizen opinions. AI automation makes sense, but the French would strongly resist the idea of 'letting machines handle democracy.' Cultural acceptance is a huge barrier.
Korea's e-People platform faces the same processing delays. But Korean AI development is led by platforms like NAVER and Kakao, unlike Fujitsu's system integrator approach. The platform-led vs. SI-led difference in government AI adoption is fascinating.
Taiwan's vTaiwan and Join platforms facilitate citizen policy discussions, but AI use is still limited. Japan's pilot results are informative, though 120,000 characters may not be enough for Taiwan's referendum-level opinion volumes.
In the UK, Fujitsu has recently been in the news for rather different reasons (the Post Office scandal). The tech might be excellent, but there are brand image issues. Might not matter domestically in Japan, but international expansion could face headwinds.
Mid-sized countries like Poland have limited AI development resources. Being able to apply a domestic LLM to government work like Japan does is somewhat a privilege of larger nations. We need EU-level common tool development.
India has 22 official languages, making LLM-based citizen opinion processing far more complex than Japan's case. This Japanese-specialized model's success is a good example of why language-specific models are needed. Very relevant for multilingual nations.
Denmark has advanced e-government but AI adoption proceeds cautiously due to GDPR. Japan's private LLM approach could be a model that aligns with EU data protection requirements. Worth watching closely.
In Mexico, digitalization of citizen participation itself is still lagging. Before AI processing of public comments, we need basic infrastructure for citizens to submit opinions online. The digital gap with Japan is stark.
Australia's federal government has set up an AI taskforce too. But there'd be significant political pushback against AI processing constituent opinions. 'Your opinion was read by AI' isn't something voters want to hear.
Japanese-Canadian here. Canada is also aggressive on AI strategy. Interesting that Cohere is Canadian but the government success story is in Japan. Shows how technology circulates globally—a Canadian-built foundation powering Japanese governance.
As a German administrative law researcher, whether AI summaries legally constitute 'consideration of opinions' is a critical legal question. How Japan's Administrative Procedure Act handles this needs rigorous legal examination.
In Vietnam, government use of AI for citizen opinion analysis raises surveillance concerns. Japan's open and transparent pilot experiment could serve as a good roadmap for Southeast Asian countries considering AI adoption.
Singapore's GovTech actively pushes AI adoption. What's interesting about Japan's approach is the insistence on a domestic LLM. Singapore flexibly uses OpenAI and Google models. The strategic difference is quite pronounced.
As the first country to appoint an AI Minister, the UAE is aggressive on government AI. Japan's case is a useful concrete use case for AI in governance. Though since we don't have a public comment system per se, direct application is tricky.