加州AB 412法案要求AI开发者完成不可能任务
SOURCE / EFF-Updates · California’s AB 412 Still Demands Developers Do The Impossible
原文
California’s AB 412 Still Demands Developers Do The Impossible
完整原文
Español California lawmakers are again considering A.B. 412, a bill that would require AI developers to identify and disclose copyrighted works used to train generative AI systems. The problem this year is the same as last year: it’s practically impossible to comply with this law. The bill demands information that often does not exist, and cannot realistically be obtained. EFF submitted an opposition letter to the California Senate Privacy Committee explaining why we continue to believe A.B. 412 is simply unworkable. To the extent developers do follow this law, it will have the effect of locking in the power of the largest companies in AI. A Burden That Can’t Be Met A.B. 412 sounds simple: just have AI developers create and keep a list of all the registered copyrighted works they use in AI training. That may seem straightforward. In practice, it’s anything but. There is no machine-readable “list” of copyrighted works at the U.S. Copyright Office. And many copyright holders can get a copyright without even depositing a publicly viewable sample of the work—for example, software companies may register copyright on proprietary code without revealing it to the public. And on the open internet, copyright information is often incomplete, unavailable, or impossible to verify. One image may be registered with the copyright office, while the next is licensed under a free Creative Commons license (like the images that EFF creates), and the next is public domain. A message forum user might post an original story, photograph, or poem without any indication of ownership or registration status. The bill effectively asks developers to continuously cross-reference massive batches of online data against a copyright system that simply wasn’t designed to do so. If California passes A.B. 412, its impact will go far beyond the large AI companies we read about in the headlines. Not Just Big Tech Supporters often frame this bill as a way to help creative workers have some leverage against Big Tech, but the bill reaches much further than the big AI companies. Its definition of “developer” extends to anyone who makes a generative AI model available to Californians. That includes indie developers tinkering with an existing model, open-source initiatives, nonprofits, and other non-commercial efforts. Recent amendments added exemptions for universities and government entities, which is important, but that still leaves out a vast swathe of non-commercial tech work that’s done by people without full-time jobs in government or academia. Large companies will hire compliance teams and lawyers to navigate these requirements. Smaller organizations and independent developers usually can’t. The result will be fewer opportunities for startups and new entrants. Faced with this massive compliance burden, some won’t even try. Courts Are Already Deciding These Questions The bill is premised on the idea that copyright owners currently don’t have good remedies if they’re mistreated by AI companies. That simply isn’t true. And the growing wave of federal court filings in this space prove it. Content companies that want to sue tech companies, large or small, have no problem doing so. Those courts are still working through important questions about fair use and transformative use. Some courts have already concluded that many AI training activities qualify as fair use. Others continue to evaluate the issue. California lawmakers should not rush to impose new state regulation while those questions remain unresolved. This is why copyright is governed at the federal level: both creators and fair users benefit from a single set of nationwide rules. At this point, the bill remains a solution in search of a problem. Rights holders already have powerful tools to protect their interests under existing federal law. What this bill adds isn’t clarity or transparency, but a costly and essentially impossible compliance burden that will discourage small developers and researchers. California has been able to support both artistic creativity and tech innovation for decades now. But A.B. 412 does not strike the right balance. If you are a California resident and interested in speaking out about this bill, you can find and contact your representatives through this website.
归纳
加州再次审议AB 412法案，要求AI开发者识别并披露用于训练生成式AI系统的受版权保护作品。然而，该法案面临实施难题：美国版权局没有机器可读的版权作品列表，许多版权信息不完整或无法验证，开发者难以合规。电子前哨基金会（EFF）提交反对信，认为法案不切实际，且将巩固大型AI公司的市场地位。法案对“开发者”的定义广泛，涵盖独立开发者、开源项目和非营利组织，而大型公司可通过合规团队应对，小实体则难以承受。此外，联邦法院已在审理AI训练中的合理使用问题，加州不应仓促立法。
点评
法案要求AI开发者披露训练数据中的版权作品，但因缺乏机器可读版权列表而难以合规，且定义过宽将加重小实体负担，可能固化大公司市场优势。
法律视角点评
AI 生成 · 人工审核核心关切
法案要求AI开发者披露训练数据中的版权作品，但因缺乏机器可读版权列表而难以合规，且定义过宽将加重小实体负担，可能固化大公司市场优势。
法律依据
美国版权法合理使用原则（17 U.S.C. §107）正在联邦法院审理中。
实务启示
中国法律人应关注训练数据版权合规，避免类似披露义务的立法趋势，并善用合理使用抗辩应对AI版权纠纷。