California's A.B. 412: Why Mandatory AI Training Disclosures Remain Unworkable

• California's A.B. 412 is under consideration again, proposing that generative AI developers disclose all copyrighted works used for training.
• EFF opposes the bill, arguing it is practically impossible to implement because no machine-readable database of copyright exists and much online data lacks verifiable ownership.
• The bill risks centralizing power in Big Tech by imposing compliance burdens that only well-funded companies can afford, while potentially stifling startups, non-profits, and independent developers.
• Existing federal law and ongoing court cases regarding fair use already provide mechanisms for copyright holders to address grievances, making state-level regulation redundant and preemptive.

Technical and Practical Hurdles

• There is no unified, machine-readable registry at the U.S. Copyright Office for developers to cross-reference.
• Copyrighted works often lack public samples or clear metadata, particularly with proprietary software or informal online content.
• The requirement to continuously verify massive, unstructured datasets against an incomplete copyright system is considered unworkable.

Economic Impact

• The bill's broad definition of "developer" includes individuals, small organizations, and open-source initiatives.
• Compliance costs will naturally favor large corporations that can afford dedicated legal and compliance teams.
• Stricter regulations may force smaller innovators to exit the market, reducing competition and diversity in AI development.

Legal Context

• Federal courts are currently adjudicating whether AI training constitutes fair use, making it premature for states to impose conflicting or additional regulations.
• The EFF advocates for maintaining copyright governance at the federal level to ensure consistent nationwide standards for both creators and technology developers.