After California enacted a state artificial intelligence transparency law earlier this month, an independent review of 13 major companies that generate AI images and audio found that only 7 — with an eighth identified by KQED — published the legally required tools to help users decipher whether a piece of content was made or edited with their technology.
As AI-generated images flood the web, identifying and labeling this content has become a key concern for educators teaching media literacy, for politicians worrying about disinformation and for students being sexually harassed with deepfakes, among others.
On Aug. 2, a law requiring AI companies with over 1 million users to mark their photo, audio and video content with metadata and to make detection tools publicly available went into effect in California — a similar law went into effect in the European Union.
The report, published last week by media outlet Indicator and WITNESS, a human rights organization focused on videos, evaluated these AI detection tools from 13 companies. About 10.6 billion users visited the companies monthly, according to the report.
The authors found that HeyGen, Midjourney, Mistral, Synthesia and xAI did not have tools available.
While the original report indicated that Suno, an AI music company based in Cambridge, Massachusetts, did not have a tool, it shared its tool with KQED and said that they released it at the beginning of the month in compliance with the law. Synthesia — an AI video platform based in the United Kingdom — told KQED that they have included the required metadata in their content and that they are “still seeking clarification on a few technical points” when it comes to implementing other parts of the law.
“Our wider view is that transparency rules work best when they track a tool’s or platform’s reach and risk,” Alexandru Voica, Synthesia’s head of corporate affairs and policy, wrote in an email. Voica said that as a business-to-business platform, their risk profile differs greatly from the standard consumer application.
HeyGen, Midjourney, Mistral and xAI did not respond to KQED’s request for comment.
When asked whether or not they had started fining companies under the new law, the State Attorney General’s office said that they can’t comment on potential or ongoing investigations.
While California’s transparency law aims to make AI content more identifiable, the report found that just because a tool exists does not mean it can reliably tell you whether or not a piece of content was AI-generated. The authors of the report found that most tools could identify their own unedited content, but only Google and OpenAI could identify their own output even after it was edited.
They also found that only Adobe and Microsoft could identify another tool’s unedited content more than half of the time.
The report noted that these results are somewhat unsurprising given that the law does not require detectors to flag AI content from rival companies.
Three of the tools, the report said, limit the number of tests a user can run.
“OpenAI’s verify tool is particularly aggressive, blocking us after as few as 7 tests in one session; Google and Meta capped out at 10 to 15 tests a day per user,” the report states.
A spokesperson for Microsoft said that their AI tool is designed 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.”* *
OpenAI said in a statement that while including these tools is an important foundation, “Metadata is not foolproof. It can be stripped, lost through uploads and downloads, or broken by transformations like file format changes, resizing, or screenshots.”
While the California law focuses on photo, video and audio, the EU law also requires detection for AI-generated text. Anthropic, which doesn’t generate photorealistic images or videos, announced that it will be watermarking text in all its new models according to a recent post. In July, State Sen. Josh Becker told KQED 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 started tagging ads that have been “created or significantly edited by generative AI” with labels earlier this summer.
Last week, some Instagram marketers noticed that Meta’s filters appear to be flagging content made with Canva’s background remover with the “AI info” label. Though some creators expressed concern over what this change could mean for online businesses, others said they would rather see “made with AI labels ‘go too far’” — than not far enough.