By Shelley Shan / Staff reporter
The government is exploring ways to compensate authors and other content providers for materials used to train artificial intelligence (AI) models, but has yet to determine how content should be valued, licensed or paid for.
Democratic Progressive Party Legislator Michelle Lin (林楚茵) raised the issue earlier this month during a meeting of the legislature’s Transportation Committee.
She asked whether the Ministry of Digital Affairs would provide content free of charge for AI model training following efforts to obtain voluntarily licensed material from Taiwanese book publishers.
“AI Artificial Intelligence” is displayed over a keyboard in an illustration photograph taken on Monday last week.
Photo: Reuters
She also asked whether the ministry was developing a mechanism for AI developers to pay news outlets for the use of their content.
Minister of Digital Affairs Lin Yi-ching (林宜敬) said the ministry is collecting four types of training content through its voluntary initiative: works whose copyright protection has expired and that are in the public domain, book descriptions provided by publishers, book preview material and works that authors have voluntarily licensed.
The ministry has not sought AI training material from Taiwanese news outlets, he said.
“We are aware that several international news media outlets have signed content authorization agreements with large AI firms, and we would work to facilitate such partnerships in Taiwan,” Lin Yi-ching said.
“Our ultimate goal is for large AI firms, particularly those based in the US, to pay for the content they use to train their models,” he said, adding that the ministry is considering only public-domain and voluntarily licensed material to avoid delaying Taiwan’s AI development through potentially protracted legal disputes over content use.
The ministry began building a corpus of high-quality traditional Chinese content for Taiwan’s sovereign AI models late last year, initially using materials provided by central and local government agencies. On Sept. 15, it announced rules governing the participation of private publishers and e-book platform operators, including provisions for voluntary participation, content authorization and the removal of works by authors who opt out.
As of the middle of this month, the corpus had accumulated 3.3 billion tokens — units of text that AI models use to process and learn language.
The government is also studying ways to encourage AI developers to pay for training materials, but has not specified how payments would be calculated, how any revenue would be shared or when a mechanism might be introduced.
At a separate meeting convened in July by the Taiwan Intellectual Property Office (TIPO), representatives of magazine publishers and broadcast media proposed measures to ensure AI companies pay for the use of their content.
The government should establish a licensing platform to allow media companies to register content for AI companies to use in exchange for payment, CommonWealth Magazine legal affairs director Chiu Ai-lun (邱愛倫) said.
She also urged AI companies to disclose the content they license and the sources of their training data.
Some international AI companies continued to scrape media Web sites even when the owners have explicitly prohibited such activity, Satellite Television Broadcasting Association secretary-general May Chen (陳依玫) said.
She also called for a mechanism to connect content owners with AI companies seeking to license their material.
TIPO’s draft Copyright Guidelines for Generative Artificial Intelligence (生成式人工智慧著作權指引) could give the impression that obtaining authorization was the only way to use copyrighted material, Taiwan Innovative Software and Services Association vice chairman Chang Jung-kuei (張榮貴) said.
Fair-use provisions, which permit the use of copyrighted material without the rights holder’s permission under certain circumstances, should remain available when data sources are lawfully accessed and the AI-generated output does not replace original works in the market, he said.
He also proposed exploring corpus-matching or collective licensing mechanisms.