A number of countries have created exceptions in their copyright laws which allow text and data mining by AI companies to train LLMs. In Asia, Singapore created such an exception in 2021. Now Hong Kong and Indonesia are planning the same. In one China case: iQiyi vs. MiniMax, iQiyi sued MiniMax for using its copyrighted video materials to train AI models without permission, marking China's AI video LLM infringement case. Several publishing cases have been started in India, as publishers claim LLMs are being trained on scraped data including their works.
The issue is more delicate than the question of simply ingesting and training LLMs, as the Indian cases and the narrow Singapore law provisions show. Some IP owners specifically restrict the accessing and use of their copyright works, others do not consent to such access and reproduction. Many have business models predicated on licensing. And in many cases the training can happen in the cloud, and that raises jurisdiction questions. So the issue is more about how the LLM secures its training data and if and how it has to pay for it.
A number of US copyright organizations have now objected to countries creating statutory exceptions for LLM training in copyright laws, without authorization or compensation. The Digital Creators Coalition, a coalition of nearly 50 trade associations and industry groups has filed comments to the USTR, in time for the annual USTR Special 301 review, listing all the countries that have or propose such exceptions.
But the issue is still live in the US. OpenAI which owns ChatGPT, submitted an open letter to the White House Office of Science and Technology asking for the right to scrape everything, under fair use principles. At the same time it advocates not allowing foreign LLM developers to do the same under US export policies.
Expect more debate and argument as 2025 follows and more LLMs appear around the world.
コメント