AI-generated fake copies of real videos circulate on TikTok : NPR

Tiktok’s researchers and users say that there is another type of deception to seek on the successful video application: Deepfake videos that copy the exact words of a real creator but with a different voice. In this case, the screenshot on the left is the original Tiktok video; On the right, a copy using a character apparently generated by AI.

Bronson Arcuri / NPR / @ aliunfilted_

hide

tilting legend

Bronson Arcuri / NPR / @ aliunfilted_

Millions of Tiktokkers watched a version of a video during last week, wrongly declaring that “they install incinerators at Alligator Alcatraz”, referring to an internet conspiracy theory that ovens, which were largely divided despite proof of immigration managed by the State in Florida Everglades, which are largely propagated.

One of the videos circulating in the rumor attracted nearly 20 million views. This stimulated a conversation on Tiktok, with creators weighing with their own sockets and, in a handful of cases, trying to demystify the baseless theory.

But there is an account whose tactics are distinguished in this familiar cacophony of disorderly online virality: a realistic tiktokker giving a direct description to the camera of the plot of the incinerator. The speaker’s image and voice seem to have been created with artificial intelligence tools, according to two NPR media experts consulted. The Twist: The words pronounced in the video are exactly the same as those of another video published by another Tiktok account a few days before. The copied version attracted more than 200,000 views to Tiktok.

Deepfake researchers – images and videos generated by the IAI are so plausible that they can encourage people to think that they are real – say that this duplication seems to represent a new way in which AI is exploited to deceive: I have recite the real words of a real person, up to triples, “UMS” and “UHS”.

Ali Palmer, a creator in Dallas, Texas, who publishes on Tiktok like @aliunfilted_, made a video on a father who jumped from a Disney cruise ship to save his child who was scammed using his exact words for a video made with AI.

The copy on Tiktok is endemic, she said, but generally spam accounts that make him republish all his video. The accounts fed by AI reciting his words by a person generated by AI are new, she said.

“It’s overwhelming. It looks like a violation of privacy,” said Palmer, 33. “I could never imagine copying someone and then making money on it. Just feels dirty.”

With all types of copying, Palmer reported it to Tiktok, but nothing happens. “It’s incredibly frustrating.”

Hany Farid, professor at the University of California in Berkeley who studies digital criminalic, said what was new here, is that the words of an average person are stolen.

“All the time, we see the identities of people being co -opted to do things like Hawk Crypto Scams or push false remedies against cancer, but he is generally a famous person or influencer,” said Farid.

Farid used a digital forensic tool to analyze the video of the Copied incinerator, the Disney Cruise video and other videos published by the same account, at the request of NPR, and concluded that they were AI products.

“This is the kind of thing that would be super easy to do with today’s IA tools, and something that would easily pass through content of content moderation,” he said.

Although the use of AI to copy videos does not seem to violate Tiktok’s policies, the platform says that it forces users “to label all the content generated by AI which contains realistic images, sound and video”. NPR copied videos identified are not labeled as generated by AI.

Deepfakes has become more sophisticated in recent years, in addition to being more and more deployed for malicious purposes.

The technology has been used to pretend to be politicians, including Secretary of State Marco Rubio, former president Joe Biden and Ukrainian President Volodymyr Zelenskyy. The rise of deep “nudification” tools has prompted Congress earlier this year to adopt a federal law to combat the spread of non -consensual intimate images, including false nudes generated by AI.

More and more fiction is this last approach to try to capitalize on a viral moment of Tiktok by having a fictitious creator recite to recite the words of a real creator.

It is difficult to assess how common the practice is on Tiktok, which is used by more than a billion people worldwide. NPR could not identify people or motivations behind the accounts reproducing the words of the creators. The accounts did not respond to requests for comments. Tiktok either, nor most of the creators whose words were attached.

Two accounts identified by NPR using the words of other creators and apparently using images and voices generated by AI, have similarities: they each have around 10,000 subscribers but do not follow anyone. Many videos published by the two accounts Represent black characters who, according to Berkeley, seem to be “low quality depths” made with AI. And each account has piloted the words of another creator of Tiktok on various viral subjects, ranging from a woman who received a lifting in Guadalajara to a woman wearing a dog necklace in Meghan Markle dancing in a delivery room in the hospital.

In Markle’s video, the creator has taken British accent. In others, the voice assumes a completely different register.

“The biggest revealer is when you take a step back and watch the account as a whole, when you watch a video compared to another video, it is clear that the character’s voice goes from video to video,” said Darren Linvill, communication teacher and co -director of the media crisis of the Clemson University. Linvill also examined the videos at the NPR request and concluded that they had been manufactured with the help of AI tools.

Together, the two accounts have rang millions of points of view by entering viral stories which combine more towards fodder of tabloids than on political drama. But researchers who follow the information war sponsored by the State say that the test of new strategies for the virality of the juice is something that the actors supported by the government are also doing regularly.

Linvill studies how nations, including China, Russia and Iran, use digital tools for propaganda. He says that the creation of IA characters, such as false news anchors, is a tactic also used by the influence operations supported by the State. While NPR has found no indication that the accounts he has identified are part of such a campaign, there is often overlapping of the tactics used by Depottive States actors and accounts seeking to obtain a commitment to social media platforms.

“Theft of content on social networks is as old as social media,” said Linvill. “What we see here is that AI does what we have seen being really good and again, and it is systematizing things, which makes the processes cheaper, faster and easier. And AI is really good to make it faster and easier to steal content.”

Alex Hendricks scrolled Tiktok this week when he saw two back to back videos on the plot of Florida Incinator. It is normal to see many creators weigh on the same subject, but these videos have struck him as Unusual because the monologues were strangely identical, just to different voices.

Hendricks therefore made a tiktok video pointing this. However, he barely had views.

“I know there are a lot of false news on Tiktok, and I was skeptical because of the AI, but this kind of replication was new and crazy,” said Hendricks, 32, who works in retail in Montana. “This is why I say to everyone: I don’t think you see anything. And reference everything you see before sharing it on Tiktok,” he said. “But I’m not sure they’ll listen to.”