▲ The Board of Audit and Inspection building
An audit has revealed that some artificial intelligence (AI) training data built by the public sector is of low quality and has been improperly managed.
As Prime Minister Han Sung-sook has repeatedly emphasized the importance of utilizing public data since taking office, attention is focused on whether efforts to improve the situation will gain momentum.
The Board of Audit and Inspection (BAI) released the results of its audit on the "current status of fostering the artificial intelligence industry" containing these findings today (August 12).
According to the BAI, as of November of last year, the Ministry of Science and ICT (MSIT) had built 908 types of training data and made them available on AI Hub, while 26 individual agencies, including the Ministry of Food and Drug Safety, the Seoul Metropolitan Government, and the Korea Expressway Corporation, had built 313 types and released them on their respective portals.
This was intended to allow companies to immediately use the data to help foster the AI industry.
However, it was found that individual agencies pursued projects without checking data that had already been built, resulting in the overlapping production of similar data.
For example, regarding four types of image data such as wild animals, household waste, concrete cracks, and eggs, the Ministry of Science and ICT spent approximately 7.3 billion won to build training data, after which five agencies, including the Seoul Metropolitan Government and the Korea Expressway Corporation, newly built similar data.
Similarly, for three types of data related to pills, oral cavities, and autonomous driving, the Ministry of Science and ICT spent 23.3 billion won between 2021 and 2023 to build training data, while three agencies, including the Ministry of Food and Drug Safety, separately built similar data during the same period.
In addition, because there were no common quality control standards applied to individual agencies, concerns were raised over degradation in quality, such as the creation of data missing information essential for AI training.
Among the top 20 agencies in terms of training data construction performance, nine allowed deliveries without third-party quality verification when executing projects.
The Ministry of Food and Drug Safety and Korea Dongseo Power did not require third-party quality verification nor even self-inspection by the project-executing organizations.
The "road-driving CCTV training dataset" built by the Korea Expressway Corporation last year lacked vehicle location information in the annotation data essential for AI learning, and the "large waste training data" built by the Seoul Metropolitan Government in 2020 had mismatched file names and images in 538 out of 2,303 files, raising concerns of errors.
The BAI also pointed out that the National Information Society Agency, which is entrusted by the Ministry of Science and ICT with data construction and post-management duties, neglected public complaints regarding data errors citing the closure of businesses that supplied the training data.
The BAI noted, "Public sector agencies are making efforts to support AI companies, but complaints from businesses about the lack of usable data continue. Improvements in both the quantity and quality of data are urgent to leap forward as an AI powerhouse."
(Photo: Yonhap News)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.