German public broadcaster ZDF recalled its U.S. correspondent in late February after AI-generated footage appeared in a report about U.S. Immigration and Customs Enforcement (ICE) officers.
advertisement
advertisement
The station apologized for using footage depicting children clinging to their mother as they were arrested by ICE agents. The video has a Sora watermark, clearly indicating that it was created by OpenAI’s video tools.
Euronews’ fact-checking team, The Cube, identified the video as one of thousands of AI-generated clips of ICE officers flooding social media. Many of them are not clearly labeled and have collectively amassed millions of views.
consistent pattern
Data shared with The Cube by AI detection company TruthScan shows more than 200 short clips depicting ICE agents being chased by teachers, brawling in bars, and being arrested by the NYPD on the subway.
TruthScan announced it had scraped over 200 videos from a single Instagram account. The videos are short, dramatic and emotional.
The Cube found that the account continues to post a large number of AI-generated clips of ICE, some of which have garnered thousands and even millions of views.
There are various clues as to why the video is not real. In one example, steam appears to rise from a solid surface during a confrontation between ICE officers and a cafe employee. This is a subtle sign that this clip is a composite.
According to TruthScan, each video is exactly 10 or 15 seconds long, strongly suggesting that it was generated using OpenAI’s Sora 2 video model, which provides users with a 10- or 15-second video option.
Unlike many AI-generated videos that contain watermarks, the majority of these ICE clips do not contain watermarks.
Similarly, last November, 404 Media reported that an AI-generated clip of an ICE agent from a Facebook account went viral, with one clip alone racking up 4 million views. The video has since been deleted.
Increasing risks to public trust
Experts say the risks extend beyond the mislabeling of individual clips to the widespread dissemination of fake content on sensitive political issues, which could undermine public trust.
“There are significant risks, both negative and positive, in collective responses to law enforcement actions,” said Ari Abelson, co-founder of Open Origins, a media authenticity company that detects deepfakes.
“It’s a concern for law enforcement agencies that want to maintain credibility, but it’s also important for agencies that want to be held accountable when they go too far,” Abelson told The Cube.
Similar to ZDF broadcasts, AI-generated ICE content is not limited to anonymous accounts. Sometimes it even reaches the upper echelons of government offices.
In January, the White House posted a digitally altered image of Nekima Levy Armstrong, an activist and lawyer who was arrested at an ICE-related protest in Minnesota.
Analysis revealed that Armstrong’s skin was darker in the altered image than in the original image, and he appeared to be clearly sobbing.
In response to questions about the doctored image, the White House responded to a post by Deputy Secretary of Public Affairs Kaylan Dole: “Once again, to those who feel the need to reflexively defend the perpetrators of our nation’s heinous crimes, I send this message: Law enforcement continues. Memes continue.”
Urmston said she was “disgusted” by the White House for posting the image.
According to Abelson, concerns arise when there is a lack of guardrails, or safeguards, to clearly indicate that content is fake.
“Political cartoons have always been used in politics, and they’re used in very similar ways,” he said. “The main difference is how surreal the videos are.”
“The social and information space requires the ability to prove that photos and videos are real, so we can enjoy fakes as fiction and feel the gravity of real events,” he added.
