Shopping can be a science. Dozens of seals and symbols supposedly identify sustainable and animal-friendly products. The Nutri-Score is intended to help you find healthy foods with its traffic light colors. How much this really brings is unclear. Some people ignore the labels, others are confused and don’t know what to look for.
This insight can be transferred to social media. “Users are overwhelmed with too much information,” says Ramak Molavi Vasse’i. She leads a research project on AI transparency at the nonprofit Mozilla Foundation and has studied how to label content generated by artificial intelligence. When it comes to food labels, no one can see through them anymore, says Molavi. “The same threatens with AI content. More labeling can also backfire.”
This is one of the key findings from the report “In Transparency We Trust?“, which Mozilla made available to the SZ before publication. The investigation deals with a question to which Facebook, Instagram, Tiktok and other platforms have to find answers as quickly as possible: How can it be possible to distinguish authentic content from AI fakes ?
The challenge is great and urgent. Since Open AI released Chat-GPT at the end of 2022, generative AI has developed rapidly. Within seconds, texts, sounds, images and videos can be created that appear more real month after month. A video in which Vladimir Putin declares nuclear war or a supposedly senile Joe Biden says nonsense at a press conference – this has long been technically possible.
Visible labels, invisible watermarks
Mozilla researcher Molavi and her colleague Gabriel Udoh have divided the existing approaches into two categories. On the one hand, there are clearly perceptible identifiers, such as visible symbols in images and videos or audible clues in synthetic audio content. On the other hand, AI products can be provided with invisible watermarks. These markings are contained in the metadata or are embedded in the content using cryptographic processes.
All visible methods have one major disadvantage: they are voluntary. “Most AI fakes on the internet are created with the intent to deceive, for example pornographic deepfakes that are created and distributed against the will of those affected,” says Molavi. “The makers are unlikely to start labeling this content visibly.” Even if they use tools that mark images or videos as AI-generated, such symbols can easily be retouched or cut away.
Labels also shift some of the responsibility to users. Similar to food, they would have to learn to deal with different forms of cues. “This can cause confusion and even further erode trust in real content,” says Molavi. For certain content, labels simply come too late. If women are to be degraded with sexualized fakes, the damage is already done as soon as others see the material – AI labeling or not.
The Mozilla report has somewhat higher hopes for invisible watermarks that can be read by machines. Cryptographic markings are harder to manipulate or remove entirely without destroying the file. However, this would require uniform technical standards to mark and recognize content across platforms. There are initial approaches to this, such as the Coalition for Content Provenance and Authenticity (C2PA), which includes Google, Microsoft and Adobe. At the same time, tech companies are also working on their own solutions, which could result in a tangle of watermarks.
A huge, involuntary social experiment
“Visible labels or invisible watermarks are not a panacea, but rather just one component among many in the control of AI,” says Molavi. “You can’t make the same mistake as the tech companies and believe that there can only be one technical solution to every problem.” She considers the platforms’ attempts so far to be a small step on a long journey. Unfortunately, voluntary self-regulation alone has never been enough. There is a need for effective regulation that is monitored and enforced.
There are elections in more than 60 countries in 2024, and Donald Trump could become US President again. Even without generative AI, the 2016 and 2020 election campaigns were characterized by lies, propaganda and disinformation. The technology does not exist in a vacuum, says Molavi. “It encounters a social media world in which algorithms favor emotional and extreme content. In such an ecosystem, AI fakes go viral particularly quickly.”
The researcher sees AI companies and platforms as having a duty. The old Facebook motto “Move fast and break things” should not be repeated. “It cannot be the task of researchers and users to point out risks and develop solutions,” says Molavi. “It sometimes feels like we’re part of a big experiment with hundreds of millions of involuntary beta testers.”