A trade group including members Apple, Google, and Meta sent a letter to the US Copyright Office on Monday warning that measures proposed in its recent inquiry on artificial intelligence policy issues would “handicap” domestic AI development and encourage innovators to take their efforts overseas.
TechNet argued that using copies of copyright-protected material to train AI models doesn’t qualify as the kind of copying that violates copyright law. A statutory licensing scheme to govern machine learning would create an “intractable economic problem” and incentivize technology companies to take their billions of dollars of investment capital to “more innovation friendly” jurisdictions, the industry group said in a statement. It added that required disclosure of training materials would force companies to disclose trade secrets to foreign AI developers.
The letter comes on the heels of President Joe Biden’s executive order on AI that outlines a broad approach to AI regulation and calls on the Copyright Office to provide recommendations about the use of protected works in AI training, an issue that has led to multiple lawsuits against Meta, OpenAI, and Stability AI. The trade group noted that its comments reflect the majority position of its members but “do not indicate unanimity.”
The group pressed the Copyright Office to preserve law provisions that offer “technology-neutral safeguards,” according to the statement. ” Generative AI “is a development copyright law should celebrate and encourage, not restrict.”
AI technology supports a “democratic creative economy,” TechNet said, and is no different than previous developments like the invention of the camera. “That the technology enables the creation of new works that might compete with the works of today’s artists is not a reason to distort copyright into a tool of protectionism,” the letter said.
Companies have poured “tens of billions of dollars of investment” into AI by relying on case law that “intermediate copying” of works as part of a process to create new works is a “transformative” fair use, TechNet said.
The group also argued that AI developers’ shouldn’t be liable for users’ wrongdoings because doing so “makes no more sense than imposing liability on the creator of a word processor simply because someone used it to draft a work of fiction that infringes on a copyrighted book.”
While the Copyright Office’s notice posed the idea of licensing regimes to pay copyright owners and creators for the use of their works in training AI models, TechNet said such a scheme would be “impossible” to administer due to the billions of works used to train models. It would fail to provide “meaningful” compensation to creators, it added.
The group said the potential requirement to disclose training materials is similarly impractical because it’d “unduly burden” developers and has no justification. “Should musicians be required to disclose every poem read on the web, every mp3 sent in an email, or every Instagram cover art inspiration that played a role in developing a hit song?,” it asked.
Materials used for training models are a “highly curated set of data” and forced disclosure would undermine incentives to innovate while hindering companies’ ability to compete with foreign adversaries.