What happened
South Korea spent years building public datasets to help train AI. But when the country’s state auditor looked across the programme, it found a basic problem. Government bodies were building and publishing training data without a consistent system for checking what already existed or whether the data was good enough to use.
The Board of Audit and Inspection says the government reported spending 1.6328 trillion won from 2017 to 2024 building 908 datasets published on AI Hub, the national platform run by the National Information Society Agency (NIA). The auditor also examined datasets built and published separately by other public bodies.
One problem was duplication. The auditor found there was no process requiring a public body to check whether similar data already existed. It records that a Daejeon district office built a household waste dataset in 2025, even though the ministry had published one in 2020. Of its 45 categories, 26 were similar. The report says wildlife, concrete crack and egg datasets already on AI Hub were also built again by other bodies.
The auditor found three problems and notified them for improvement. Two notifications went to the Ministry of Science and ICT and one to the NIA. These were notifications requiring improvement, rather than sanctions.
What an auditable version would have shown
The auditor found that individual public bodies had “no unified quality standard”, so each applied its own checks. The bodies reported their quality management arrangements to the auditor. Of 26 bodies that had built 313 datasets outside AI Hub, four had no quality standard and nine had no third party check.
The problems were visible in the data itself. At the time of the audit, the auditor found Korea Expressway Corporation’s 2025 road CCTV dataset had 2,680 images without bounding box coordinates. The report says the dataset was “still published as it was, although an AI model cannot learn from it”.
The auditor found another problem in Seoul’s 2020 bulky waste dataset. It records that “538 of 2,303 images, 23.4 percent, do not match their JSON label files, yet the dataset is published without correction”. Seoul confirmed the errors and said it would fix them before publishing again.
The report records that some public bodies focused on scoring well on the Ministry of the Interior and Safety’s evaluation indicator for opening AI friendly, high value data. The auditor found some published data without enough review of whether it could actually be used to train AI.
Where the gap was
Finding errors was one problem. Fixing them was another.
The auditor found that of 1,270 companies that built datasets from 2021 to 2024, 62 had closed by 6 January 2026. Three more were reported as unable to respond to users. Together, those companies built 213 of the 721 datasets examined, or 29.5 percent.
The auditor says the NIA received complaints about errors in some of these datasets and closed some because the builder had gone out of business. The data remained on AI Hub as of December 2025. The report gives examples, including a wildfire dataset with empty files and a telephone speech dataset where a user complained that “much of the transcription does not match the audio”.
The auditor records another gap. After the one year defect liability period ends, other consortium members may avoid joint liability, and the NIA has no means to compel them. That meant responsibility for fixing a dataset could become difficult to enforce even while the data remained available.
A VerificationGate is designed to check an output against a trusted reference before it is released, such as checking that each label matches its image.
A ConductRecord keeps who built each dataset, what checks it passed and how each complaint about it was handled.
VerificationGate and ConductRecord are Headlights designs.
What governance should have looked like
The ministry reported to the auditor that it had no general disagreement with the findings and planned a survey of AI datasets across government. The NIA said it would budget for maintenance and consider letting users propose corrections.
For data published so that others can train AI, best practice would be to check it against a common standard before release, record who is responsible for it, and fix or withdraw it when an error is confirmed, whoever built it.
Failure Pattern: according to the state auditor, public bodies published AI training data under differing quality standards and checks, and the agency running the national platform closed some complaints about defective data because the contractor that built it had closed.
Governance Principle: a body that publishes data for others to train AI should check it before release and keep someone responsible for fixing it for as long as it stays online.
The reference implementation of VerificationGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- 인공지능산업 육성실태 Ⅲ (AI 학습용 데이터 구축사업 중심), audit report and press release (Board of Audit and Inspection of Korea, published 12 August 2026)
- State audit finds duplicate, unusable data in 1.1 billion dollar AI training project (Korea JoongAng Daily, 12 August 2026; headline currency symbol written out)