California's AI Transparency Act Takes Effect with Fines That Compound Daily

A new law in California is putting the spotlight on artificial intelligence (AI) transparency, particularly when it comes to synthetic media. The state’s AI Transparency Act has taken effect, and companies that fail to comply face significant fines. According to the law, any provider of generative AI tools must offer a free public tool for assessing whether an image, video, or audio clip was created or altered by their system. This includes placing a latent disclosure in those files and giving users the option of a visible, manifest disclosure.

The threshold for this requirement is set at more than 1 million monthly visitors or users for a publicly accessible generative AI system in California. If a company’s tool meets these criteria, they must ensure that their synthetic media carries a clear indication of its origin. This can be achieved through watermarking or other methods that provide machine-readable signals around the content.

The law is part of an effort to combat deepfakes and other forms of AI-generated synthetic media that can deceive users. The California Legislature’s bill text emphasizes the importance of transparency in this area, stating that a blog post about safety will not suffice as compliance. Companies must implement working disclosure machinery that allows users to assess whether an image or video was created by their system.

The penalty for non-compliance is steep. Covered providers can face a civil penalty of $5,000 per violation, collected by the California Attorney General, a city attorney, or a county counsel. Each day in violation counts as a separate offense, making it expensive to ignore this requirement. The law’s authors are clear that they expect companies to take proactive steps towards transparency.

The original operative date for this law was January 1, 2026, but it has been amended by AB 853 to August 2, 2026. This change coincides with the European Commission’s implementation of Article 50 of the EU AI Act on the same day. The overlap is significant because companies like OpenAI and Google now face similar requirements in both California and Europe.

Both regimes prioritize provenance and watermarking over relying on users to guess what looks real. In California, latent disclosure must be detectable by the provider’s own tool and consistent with widely accepted industry standards such as C2PA (Coalition for Content Provenance and Authenticity). This standard is backed by major companies like Adobe and Microsoft.

Google recently demonstrated its compliance efforts when it pulled a new AI image-editing feature from Google Earth less than 24 hours after launch. The feature allowed users to transform real-world places in Google Earth, which can create synthetic media that carries borrowed authority from the map beneath it. A fake image over a real coordinate can deceive even the most discerning viewer.

AI researcher Henk van Ess used this feature to create convincing images of refugee camps along the US-Mexico border and a nuclear plant in Iran. Google told users that these images included SynthID watermarking, which could be checked through Gemini or Lens. However, as The Verge reported, Digital Digging was able to fool Hive’s AI detector with an AI-altered Google Earth video.

This incident highlights the challenges of implementing effective disclosure mechanisms. A watermark that disappears when a file leaves a demo environment is not transparency – it’s a weak receipt. Companies must ensure that their systems can survive the messy path from creation to screenshot, upload, compression, reposting, and detection.

The next wave of compliance will hit large online platforms with more than 2 million monthly California users on January 1, 2027. These platforms, including public-facing social media, file-sharing services, mass-messaging products, and search engines, must detect provenance data that follows widely adopted specifications from an established standards-setting body.

The distinction between model makers and platform providers is crucial here. The former can mark a file with metadata, but the latter decides what users see when files move through feeds or search results. If a platform buries a provenance signal in its interface, readers are still left guessing.

Compliance has been uneven so far. OpenAI has published a California training data summary under AB 2013, and Assessed Intelligence reported that OpenAI, Anthropic, and Google posted their disclosures by the January 1, 2026 deadline. However, xAI is notable for suing California Attorney General Rob Bonta over AB 2013 in federal court.

This regulation marks a significant shift towards transparency in AI-generated synthetic media. It acknowledges that users cannot eyeball these files at scale and instead asks companies to implement robust disclosure mechanisms. The law’s authors are clear: if your product makes media that can pass as real, California is now asking whether you can prove what your product made.