California’s new AI transparency law took effect this week, but it defines “transparency” in a way that may not help many consumers. The law, which passed in 2024, sets up a robust transparency infrastructure for AI-generated content, but it does not guarantee that the real “story” of a piece of content is shown when it matters most: at the user’s first impression.
What California’s law gets right
That story is told through the provenance data embedded within AI-generated media. The transparency law takes a thoughtful and technically sophisticated approach to making that data available to journalists, platforms, investigators, researchers, and courts.
The law requires AI companies that make widely used AI generation tools to include two types of provenance data with the content their tools produce. The first is metadata, accompanying the content file that identifies who created the media, when, and with what tool. The other is a smaller set of the same kind of information hidden within the content itself (for example, within the pixels of a generated image).
AI companies must provide a free public detection tool that can identify either kind of provenance data. The law also requires AI creation tools to offer users a way to attach a visible or audible label indicating that the content was AI-generated, but it does not require the tools to affix such a label to everything they generate.
The authors of the law were aware that establishing and preserving provenance data might be of limited use if actual content consumers never see it. So, starting next year, the law will require content distribution networks, such as social networks, to at least alert users to the availability of the data. A platform can provide a link to the provenance data, allow the user to download it, or provide a special interface for displaying it. Even if the platform chooses that last method, however, there is no language in the law specifying that the interface must appear adjacent to the content.
Some platforms might decide to affix an “AI-generated” label directly to the content. Others might display a text link reading “content credentials available,” or a small icon. It is up to the platform. The effectiveness of the law could therefore depend on the platform operator’s (and its lawyers’) design choices rather than on the technical provenance system the law establishes.
Two kinds of “transparency”
The way the law is written illustrates that the AI industry and its regulators are working with two different definitions of “transparency.” One means that AI companies’ tools establish and preserve the technical “story” of a piece of content. The other refers to an assurance to content consumers that images, videos, or audio recordings really are what they purport to be, and that they can be taken at face value.
The California law mainly focuses on the first definition. But provenance data only sets the stage for real user transparency. The law does not require that the data be part of the content’s initial presentation to a user.
The first-impression gap
In the attention economy, that initial presentation matters a lot. One look at a generated video, for example, can seed the user’s mind with an opinion or reinforce one that is already there. A fake product video that goes viral could quickly damage a brand. A generated CEO statement could sink a company’s stock. Synthetic audio of a politician propositioning a minor could tank a campaign. Some pieces of content cannot be unseen, even if they are debunked by provenance data just a few minutes later.
Because many users grew up with the assurance that “seeing is believing,” they are vulnerable to a new technology that short-circuits the maxim. And because confirmation feels better than cognitive dissonance, users might prefer to believe what they see and move on, ignoring the opportunity to drill down into the authentication data.
California was right to require that provenance data be available. But availability is not enough. For some kinds of content, it is crucial that the data accompany the presentation of the content.
Even though California’s law may serve as a useful model for other states, nobody is saying that it is meant to be the last transparency law the state will need. While the law does venture into regulating content distribution, its main thrust is establishing a robust authenticity infrastructure for provenance in the AI age. Future legislation may focus more on the “last mile” problem of meeting users with the “story” of a piece of content when and where they need it most.