A survey of 14,208 Korean civil servants asked one question: “Can AI read the documents your agency produces?” 91.1% said no. The documents are trapped in HWP — a domestic Korean word processor format — and scanned PDFs, the kind machines cannot even open. Meanwhile, 68.9% of those same officials have already used generative AI, and 67.5% of them are running foreign AI tools on outside networks because the inside ones can’t help. In the age of AI, government documents ought to be a first-class data asset. Instead, the machines can’t read them. This is the paradox of a country that has topped the OECD Open, Useful, Reusable Government Data Index multiple times.
The United States works differently. The Office of the Federal Register accepts documents in Word, then runs them through the Government Publishing Office’s typesetting system, which produces XML, HTML, and PDF versions for distribution. The key is the split between the tool you write in and the format the public receives. The GPO Style Manual, in continuous use since 1894, treats clear, content-first prose as the norm; it does not specify table-border colors or pixel-perfect spacing, the way Korean practice does. GovInfo, the federal platform, serves hundreds of thousands of documents — including the Federal Register and the U.S. Code — in machine-readable form.
The 2010 Plain Writing Act made clear government communication a legal requirement. The interesting twist is that the law has no teeth in court — it explicitly bars judicial review. Instead, a private nonprofit publishes annual report cards grading each agency from A to F, the Office of Management and Budget issues implementation guidance, and each agency posts its own annual report online where the press can pick it apart. Transparency and public shaming, both soft levers, are rewriting document culture without a single lawsuit. The structured documents that accessibility rules already required for disabled users have, as a bonus, turned out to be exactly what AI needs.
The real bottleneck in Korea isn’t the HWP format. It’s the culture of decoration. Documents begin with an outer border, then fill the page with a table, then nest tables inside tables inside tables. In the early 1990s, when Microsoft Word couldn’t yet handle Korean input, Hangul Word Processor’s flawless Korean support standardized the military, the schools, and the public sector all at once. Three decades of path dependence later, public documents are still treated like desktop-publishing artifacts. Moving off HWP doesn’t fix this; carrying the decoration culture into DOCX produces the same mess in a different wrapper.
The rest of the world is already moving. The UK has adopted Markdown-based Govspeak as its standard for government web content, mandated the Open Document Format for document exchange, and in 2026 recommended machine-readable formats — CSV, JSON, XML — for AI-ready data. Norway and Portugal have also mandated ODF, and NATO has designated ODF as one of its required standards. Korea’s Lee Jae-myung administration has placed AI at the center of its agenda and enacted an AI Framework Act, but the underlying document-format bottleneck remains unaddressed.
The way out is a layered approach. Keep authoring in the existing tools, but at the publication stage produce machine-readable formats — Markdown, JSON, XML — in parallel. Just as the U.S. writes in Word but publishes in XML, the trick is to separate writing from distribution. For the billions of HWP documents already in existence, a national conversion project should be designed to start with the highest-value targets — statutes, policy reports — and turn them into structured data. At the same time, Article 14 of the Korean Language Basic Act should be strengthened with designated officers in each agency, mandatory public annual reports, and a private-sector evaluation body, layering compliance pressure from multiple directions. Mandating prose-centered forms instead of nested tables is a more fundamental fix than any format conversion. The U.S. experience proves the point: transparent reporting and outside scrutiny — soft power, not court orders — can shift how a bureaucracy writes.
In the AI-government era, competitiveness will be decided by data quality, not GPU count. And that starts with documents. Korea already has the capacity — first in the OECD on open data — and the institutional foundation, in the AI Framework Act. What’s left is breaking a thirty-year habit. The first step isn’t to adopt a new AI model. It’s to start writing documents that AI can read.