Taiwan's Traditional Chinese AI corpus stalls as copyright law stays unchanged
A TechNews report citing the Taiwanese magazine Business Today (今周刊) says Taiwan's effort to build sovereign-AI Traditional Chinese corpora is constrained more by copyright rules than by technology, with only about 1% of training data coming from the private sector. Taiwan.md, an open knowledge base started by new-media artist Wu Che-yu, has collected 1,123 Traditional Chinese articles drawn mainly from public academic papers, blogs and government material, and the team discards anything with copyright doubt. The report cites a July 2025 case in which a National Taiwan University doctoral student faced a criminal complaint from the Central News Agency over 140,000 unauthorized news items in a shared corpus; the two sides later settled and the corpus was taken down, which researchers describe as a chilling effect. Chinese-language training data for major models is mostly Simplified Chinese, with Traditional Chinese under 1%, the report adds.
It shows that copyright and legal risk, not technical capability, can bottleneck sovereign-AI data work and limit whose language and viewpoint models learn.
Taiwan.mdMinistry of Digital Affairs (Taiwan)Central News Agency (Taiwan)