California's new artificial intelligence transparency law, which took effect on August 2, requires AI companies with over 1 million users to publish detection tools that help people identify AI-generated content, but compliance has been spotty across the industry. An independent review found that only 7 of 13 major AI companies evaluated had published the legally required tools, though Suno, an AI music company based in Cambridge, Massachusetts, later shared its detection tool with journalists after initially appearing non-compliant.

What Is California's New AI Detection Law?

On August 2, California enacted a state artificial intelligence transparency law that requires AI companies to mark their photo, audio, and video content with metadata and to make detection tools publicly available to users. A similar law went into effect in the European Union around the same time. The law targets companies with significant reach, as the 13 companies evaluated in the compliance review collectively attracted about 10.6 billion monthly users.

The motivation behind the law is straightforward: as AI-generated images, audio, and video flood the internet, identifying and labeling this content has become critical for educators teaching media literacy, politicians concerned about disinformation, and students being targeted with deepfake harassment. The detection tools are meant to give ordinary users a way to verify whether content was created or edited using AI technology.

Which Companies Are Complying and Which Are Not?

A report published by media outlet Indicator and WITNESS, a human rights organization focused on video content, evaluated AI detection tools from 13 companies. The findings revealed significant gaps in compliance. According to the review, five companies did not have detection tools available at all:

• HeyGen: An AI video platform that did not publish a required detection tool
• Midjourney: A popular AI image generation service without a publicly available detection tool
• Mistral: An AI company that failed to publish the required detection tool
• Synthesia: An AI video platform based in the United Kingdom that did not have a tool available
• xAI: An AI company that did not publish a detection tool

Suno's situation was initially unclear. The original report indicated that the AI music company did not have a tool available, but Suno subsequently shared its detection tool with journalists and stated that it released the tool at the beginning of August in compliance with the law.

Companies that did publish detection tools included Google, OpenAI, Adobe, Microsoft, and Meta, among others. However, the mere existence of a detection tool does not guarantee its effectiveness.

How Effective Are These Detection Tools Actually?

The compliance review uncovered a troubling reality: just because a detection tool exists does not mean it can reliably identify whether content was AI-generated. The authors tested tools from multiple companies and found significant limitations in their accuracy and functionality.

Most tools could identify their own unedited content, but only Google and OpenAI could identify their own output even after it had been edited by users. This is a critical limitation, since real-world AI content is often modified, cropped, or reformatted after generation. When it came to identifying content from rival companies, the results were even worse. Only Adobe and Microsoft could identify another tool's unedited content more than half of the time.

The report noted that these limitations are somewhat unsurprising given that the law does not require detection tools to flag AI content from competing companies. Additionally, three of the tools tested limit the number of scans a user can run. OpenAI's verification tool was particularly restrictive, blocking users after as few as 7 tests in a single session, while Google and Meta capped usage at 10 to 15 tests per day per user.

What Are Companies Saying About These Limitations?

When contacted by journalists, several companies acknowledged the challenges. A Microsoft spokesperson explained that their AI detection tool is designed specifically to identify content generated or edited by Microsoft AI systems, and that "editing or other downstream modifications can affect the availability or readability of those signals".

"Metadata is not foolproof. It can be stripped, lost through uploads and downloads, or broken by transformations like file format changes, resizing, or screenshots," OpenAI stated.

OpenAI, statement to KQED

Synthesia, the UK-based AI video platform, told journalists that they have included the required metadata in their content and are "still seeking clarification on a few technical points" regarding implementation of other parts of the law. Alexandru Voica, Synthesia's head of corporate affairs and policy, emphasized that transparency rules work best when they account for a tool's reach and risk profile, noting that their business-to-business platform has a different risk profile than consumer applications.

HeyGen, Midjourney, Mistral, and xAI did not respond to requests for comment about their compliance status.

What Happens Next for Enforcement?

When asked whether the California State Attorney General's office had begun fining companies under the new law, officials said they cannot comment on potential or ongoing investigations. This leaves the question of enforcement largely unanswered, though the threat of penalties may encourage lagging companies to develop and publish detection tools in the coming months.

The compliance challenge extends beyond California. In July, State Senator Josh Becker indicated that starting in January, large online platforms, including social media companies and search engines, will be required to detect AI content and enable users to inspect it. Meta has already begun tagging ads that have been "created or significantly edited by generative AI" with labels, a move that began earlier in the summer.

In July, State Senator Josh Becker

The European Union's parallel transparency law also requires detection for AI-generated text, not just images and audio. Anthropic, which does not generate photorealistic images or videos, announced that it will be watermarking text in all its new models to comply with EU requirements.

How to Verify AI-Generated Content as a User

While detection tools remain imperfect, users can take several steps to verify whether content was created with AI:

• Check for Metadata: Look for metadata tags or watermarks embedded in images, audio, or video files that indicate AI generation or editing
• Use Official Detection Tools: Try detection tools published by major AI companies like Google, OpenAI, Adobe, or Microsoft, keeping in mind their limitations with edited content
• Look for Editing Artifacts: Be aware that edited or reformatted content may not be detected by tools, so examine files for signs of manipulation or unusual characteristics
• Cross-Reference Sources: Verify the origin of content by checking the source platform or creator directly rather than relying solely on detection tools

The California transparency law represents an important first step toward making AI-generated content more identifiable, but the current state of detection tools reveals that technology alone cannot solve the problem. As AI music generation, image creation, and video synthesis become increasingly sophisticated, the gap between legal requirements and technical capability will likely remain a challenge for regulators and users alike.