August 2 was a major milestone for AI transparency. On the same day, the California AI Transparency Act (CAITA) and the EU AI Act introduced new requirements on products used to generate synthetic content.
While not identical, both laws require companies that make generative AI tools to provide a publicly-available detector that people can use to determine whether a piece of content was made or edited with their systems.
Detection tools are of particular interest to journalists, investigators, and researchers who study the information ecosystem. With that in mind, Indicator and WITNESS conducted the first comprehensive census and evaluation of the available tools.
The number of providers that need to comply with the new regulations is both large and a little fuzzy. Only the 82 signatories of the relevant section of the code of practice associated with the EU’s AI Act agreed to build a detector, but all AI providers that offer their products in the Union must make their content detectable in some way by December 2. Meanwhile, the California law applies immediately to any AI provider accessible in the state with at least 1 million users.
We focused on providers that cleared California’s monthly traffic requirement (according to either Semrush or Similarweb). Several are on the steering committee of the industry-led Coalition for Content Provenance and Authenticity (C2PA) and voluntarily signed the EU code of practice. Others publicly committed to watermarking content only after the California and EU laws kicked in.
We analyzed a total of thirteen companies. The list includes large operators Google, Meta, Microsoft, and OpenAI — but not Anthropic because, despite committing to make its content detectable, the AI lab doesn’t generate photorealistic images or videos. Also on the list are a second tier of tools including Adobe, Grok, Midjourney, and Mistral. Because the bills aren’t exclusive to generalist LLM tools, we included AI avatar companies HeyGen and Synthesia, and AI audio generators ElevenLabs and Suno. Finally, we added TikTok because it’s a massive social media platform with generative AI features.
We found that seven of these providers do not appear to have a dedicated public detection tool available yet, in seeming contravention of current legal obligations. Of the companies that did have a dedicated detector, only one correctly identified all of the images generated by its parent company’s AI tool.
Below the paywall: the results of 243 tests, links to all the provider-built tools we found, and our full analysis. You may also be interested in Indicator’s guide to AI labels and audit of AI labeling on social platforms. An Indicator membership provides access to all of it.
Get the full story — and the data
Members get access to all of the assets identified in our articles.
UpgradeUpgrade now to access:
- The full story
- The spreadsheet with all of the actors we identified in this report
- A walkthrough of how we conducted our investigation
