last week, The Republican National Committee ran a video ad against Biden with a small disclaimer in the top left of the frame: ‘Built entirely with AI images’. Critics question the reduced size of the disclaimer , suggesting that its value is limited. Especially since the ad marks the first substantial use of AI in political attack ads. As AI-generated media becomes more mainstream, many argue that text-based labels, captions, and watermarks are key to transparency.
But do these labels actually work? Maybe not.
For a label to work, it must be readable. Are the letters large enough to read? Are the words accessible? It should also provide viewers with meaningful context about how the media was created and used. And at best, it also reveals intent. Why was this medium introduced to the world?
Journalism, documentary media, industry, and scientific publications have long relied on disclosures to provide audiences and users with the context they need. Journalistic and documentary films commonly use overlay text to cite sources. Warning labels and tags are widely used on products, foods and medicines. It is essential in scientific reporting to disclose how the data and analysis were obtained. But labeling synthetic media, AI-generated content, and deepfakes is often seen as an undesirable burden, especially on social media platforms. It’s an afterthought to the slap. Boring compliance in the age of misinformation/disinformation.
As such, many existing AI media disclosure practices such as watermarks and labels can be easily removed. Even when they’re there, the audience’s eyes now seem to be trained to rapid visual input. can not see Watermark and Disclosure.For example, in September 2019, the famous Italian satirical TV show stricia la notizia Posted low fidelity face swap videos Former Prime Minister Matteo Renzi sits at his desk insulting then-coalition partner Matteo Salvini with exaggerated hand gestures on social media. stricia The watermark and clear text-based disclaimer led some viewers to believe the video was real.
This is called context shift. Even with labels and watermarks, when media is distributed to closed, politicized social media groups, creators have control over how it is assembled, interpreted, and shared. lose. As the Witness and MIT joint research study found, when satire mixes with deepfakes, confusion often arises, as in this case. stricia video. Simple text-based labels of this kind can create the additional misconception that unlabeled things are not being manipulated, which in fact may not be the case.
Technicians are working on ways to quickly and accurately trace the provenance of synthetic media, including encrypted provenance and detailed file metadata. As for alternative labeling methods, artists and human rights activists offer promising new ways to better identify this kind of content by reframing labeling as a creative act rather than an addition. doing.
If the disclosure is embedded in the media itself, it cannot be removed. In fact, it can be used as a tool to help viewers understand how and why media was created.For example, in the David France documentary Welcome to Chechnya, vulnerable interviewees were digitally disguised with the help of ingenious synthetic media tools such as those used to create deepfakes. appeared. This is a clue that the images viewers were seeing were manipulated and that these subjects were taking great risks in sharing their stories. Also, Kendrick Lamar’s 2022 music for his video “The Heart Part 5”, where the director uses his technology to deepfake Lamar’s face Will his Smith, OJ Simpson and Kobe Bryant changed to both deceased and living celebrities such as. The use of this technology is written directly into the song’s lyrics and choreography, such as when Lamar uses his hand to swipe his face, clearly demonstrating deepfake editing. The resulting video is Meta’s commentary on the deepfake itself.
