The US Copyright Office is soliciting public input regarding potential new regulations concerning the use of copyrighted content by generative AI. Major AI companies, including Meta, Google, Microsoft, Adobe, Hugging Face, StabilityAI, and Anthropic, as well as Apple, have shared their perspectives. While they have some differences in their approaches, the common theme among most of these companies is their opposition to the idea of having to pay for using copyrighted materials in training AI models.
The Copyright Office initiated the comment period on August 30th, with a deadline for written comments by October 18th. The proposed changes include aspects such as the use of copyrighted data for AI model training, the potential copyrightability of AI-generated content without human involvement, and the issue of AI copyright liability. In recent times, there have been numerous copyright-related legal disputes, involving artists, authors, developers, and corporations alleging violations in various cases.
Here are summarized responses from each company:
- Meta contends that establishing a new licensing system after the fact would be chaotic, as it would necessitate identifying numerous rightsholders for little financial benefit due to the small royalties involved.
- Google likens AI training to reading a book and argues that if training could be done without making copies, it would not raise copyright concerns, as it aligns with the purpose of copyright law.
- Microsoft is concerned that requiring consent for using accessible works in AI training would hinder innovation and disproportionately affect smaller AI developers who lack the resources for licensing.
- Anthropic suggests that the existing copyright law can accommodate the diverse interests at stake, promoting AI’s benefits while addressing concerns.
- Adobe invokes the Sega v. Accolade case to argue that intermediate copying of software for functional purposes can constitute fair use, benefiting creative expression.
- Hugging Face maintains that training on copyrighted material for creating AI models serves a broadly beneficial purpose and is generally considered fair use, although some circumstances may present tougher calls.
- StabilityAI points out that various countries, including Singapore, Japan, the EU, Republic of Korea, Taiwan, Malaysia, and Israel, have established safe harbors for AI training similar to fair use.
- Apple emphasizes that AI-generated code, when under the control of a human developer who makes expressive decisions, can have sufficient human authorship to be copyrightable.