California AI training bill demands the impossible

California's AB 412 demands a complete list of training works.

California lawmakers are revisiting a controversial mandate that would force AI developers to identify and disclose every copyrighted work used to train generative systems. The core idea sounds straightforward: catalog all training data so owners can see how their work is being used. The reality, critics warn, is strikingly different. The Electronic Frontier Foundation has filed an opposition letter with the California Senate Privacy Committee arguing that AB 412 is simply unworkable in practice and would tilt the AI industry toward the largest players.

The central snag is practical, not political. There is no national, machine-readable register of all copyrighted works, and no reliable way to verify ownership across the vast and shifting landscape of online data. The filing notes that many works are registered in limited ways or not at all, while others are licensed or released into the public domain under varying terms. In short, the bill asks developers to perform an ongoing, real-time cross-check against a copyright system that simply wasn’t designed to support this kind of workflow. Even if a company could compile an initial list, keeping it accurate as new data is ingested in continual model updates would be an almost impossible task. The result, according to EFF, is a compliance burden so heavy that it would effectively privilege those with the deepest pockets and largest data footprints.

The stakes extend beyond paperwork. Supporters say requiring transparency about training data would curb copyright misuse and give creators a clearer line into how their works are used. Opponents argue the opposite: the mandate would chill innovation, drive up development costs, and entrench the dominance of the biggest AI incumbents who can afford arduous data-science chores and legal reviews. The EFF’s position frames AB 412 as a policy that, if enacted, would push compliance into a technology stack already prone to opacity. Without a workable mechanism to identify and attribute training data at scale, the rule could end up being more fantasy than feasible guideline.

From a practitioner’s standpoint, the proposal presents several concrete constraints and tradeoffs. First, the data-traceability problem is not just about ownership; it’s about provenance and licensing across data streams that are often noisy, duplicated, or aggregated from third parties. For many developers, the cost of attempting even a rough reconciliation, let alone a fully auditable ledger, could be prohibitive. Second, the bill’s framing risks uneven punishment: larger firms with compliance teams and lawyers could absorb the overhead, while smaller teams would face a disproportionate hit to time to market and reliability. Third, the enforcement question looms large. If there is no reliable, universal registry, how would regulators adjudicate claims of noncompliance or disputes over ownership? And what penalties would attach, when the data landscape itself is contested and constantly evolving? Finally, observers will watch for any legislative revisions. If AB 412 moves forward, expect intense lobbying around carveouts, safe harbors, or alternative transparency mechanisms that aim to preserve usefulness for developers while addressing creator concerns.

The debate underscores a broader policy challenge: how to protect creators without throttling AI development. Proponents push transparency as a tool for accountability; opponents warn it could undermine competition and innovation. For compliance officers and tech leaders, the key takeaway is that if AB 412 advances, the precise scope and enforceability will matter a lot more than the headline ambition. Watch for how the bill handles timelines, exceptions, and the practical means of proving what data was used in training. Until then, the industry remains in a stalemate between copyright protection and the practical realities of training modern AI.