As California lawmakers return to session in August, companies using artificial intelligence (AI) to support ordinary business functions should pay close attention to California Assembly Bill 412 (AB 412). As drafted, it would require companies to answer questions that, in many cases, they simply cannot.
No Practical Way to Comply With AB 412
AB 412 appears straightforward. If a copyright owner believes a work was used to train an AI model, they can submit a request to the company that developed or substantially modified the model. The company must investigate and respond within 30 days. In practice, however, the bill asks companies to determine something they often have no practical way to determine. A request under AB 412 may identify a work only through a copyright registration, preregistration, or index number, which does not provide the underlying work or a searchable dataset. It generally consists only of metadata such as the title, author, claimant, registration number, and effective date. While sufficient to confirm that a registration exists, it does not allow a company to compare the work against any AI training data the company may maintain.
The scale of the problem is significant. The US Copyright Office issued more than 441,000 registrations in 2023 alone, while the open web contains billions of works. There is no comprehensive, searchable, machine-readable registry that links copyright registration numbers to the underlying content they represent.
As a result, a company receiving an AB 412 request often has no reliable way to locate the underlying work — assuming it remains publicly available at all — and compare it against training data it has retained. The company is effectively left with two impossible choices: Explain that it cannot verify the claim because it lacks access to the underlying work, or attempt to locate and assemble copies of potentially thousands of works from scattered sources to evaluate a single request.
AB 412 Reaches Companies Across Every Sector of California’s Economy
The bill’s reach extends far beyond companies developing frontier AI models.
Today, businesses across every sector of the economy build AI systems by adapting or fine-tuning existing foundation models for their own operations. A hospital system may customize a model to assist with clinical documentation. A financial institution may adapt a model for fraud detections. A manufacturer may tailor a model for quality control or predictive maintenance. These organizations did not develop the underlying model and typically never receive the original training datasets or records used to create it.
AB 412 would require companies to respond to requests about the data they use to modify AI models created by other companies. As AI adoption expands across healthcare, finance, manufacturing, retail, logistics, and countless other industries, the bill’s compliance obligations create a significant burden for companies across California’s economy that incorporate AI tools. It reaches a wide range of businesses that use and adapt AI as part of their everyday operations.
The bill’s obligations are also extremely broad. While it only requires companies to disclose information about data they use to train or modify a model, that requirement to establish a rights holder mechanism extends to all “uses” of data, not just training.
AB 412 Seeks Information and Records That May Not Exist
The bill also assumes that companies possess historical records about training data that, in many cases, were never created or no longer exist.
AI models are developed over long periods of time. During that process, datasets are routinely cleaned, deduplicated, updated, replaced, and rebuilt as part of normal development. Training data is transformed into tokenized representations, rather than maintained as a library of original source documents. Vendors that supplied data may no longer exist, and companies may also be required under privacy laws to minimize or delete certain datasets over time. For example, California’s privacy law prioritizes the principle of “data minimization,” which encourages companies to delete personal data when it is no longer needed — an important practice for protecting data.
Despite these realities, AB 412 effectively requires companies to prove whether a particular copyrighted work was used during model development, even when no practical means exist to reconstruct that information.
The bill also places no meaningful limits on the scope of requests. A rights holder could submit claims identifying thousands — or even hundreds of thousands — of works at once, while companies would still be required to respond within 30 days under the threat of escalating penalties.
Nor is the bill limited to future AI development conducted under new recordkeeping practices. It reaches backward to models developed years ago and expects companies to produce information that was never recorded in the first place.
AB 412 Even Creates Legal Exposure for Its Intended Beneficiaries
AB 412 has been described as intended to benefit the creative community without regard (as noted above) for its impacts on California-based manufacturers, growers, retailers, or other service providers. But this poorly designed bill may even create collateral damage for California’s artistic community. Why? Because artists of all types are incorporating AI into their creative processes, and therefore could be sued by class action attorneys targeting Californians under this law.
AI provides creators with new tools to enhance their craft — from special effects in film, to sound mixing, architectural planning, and vehicular styling and design. Artists, architects, designers, and other creators who use creative tools could just as easily be targeted under AB 412 as any of the other unintended victims of this legislation.
The Bottom Line: AB 412 Creates an Unworkable Compliance Obligation
Transparency in AI development is an important policy objective. But transparency requirements only work when companies have a realistic ability to comply.
AB 412 fails that basic test. It requires companies to answer questions about information they often do not possess, cannot obtain, or cannot reasonably reconstruct. And it imposes those obligations not only on developers of foundation models, but on businesses across California that customize and deploy AI as part of their normal operations.
Rather than creating a workable transparency framework, AB 412 creates an impossible compliance obligation backed by escalating legal and financial penalties.