AI gave us Jesus hugs and automated CCP propaganda
🤥 Faked Up #34: AI hugging videos are a prelude to more deepfake harassment, DeepSeek parrots CCP talking points, and two new surveys contradict Zuckerberg's "cultural tipping point" on fact-checking
This newsletter is a ~5 minute read and includes 45 links.
There’s so much research on deception that it can be hard to keep track of it all. Paid subscribers get access to my easy-to-digest summaries of 43 studies.
HEADLINES
Human rights activists are worried about a new anti-disinformation law in Pakistan. A deepfake porn video created to raise awareness about deepfake porn was called out as “insulting.” Watch out for reply bots on Bluesky. The EU will “stress-test” social media countermeasures to disinformation ahead of the German election. Presumably, this will include discussion of the Storm-1516 Russian influence operation. Canada’s inquiry on foreign interference found no “traitors” in Parliament but plenty of danger from disinformation. The Onion is blaming its run-ins with AI slop on Shutterstock.
TOP STORIES
DeepSeek propaganda and you will find it
The Chinese AI model DeepSeek has Silicon Valley freaking out because it’s much cheaper to run than ChatGPT but appears to be equally capable. (For more, check out Wired and ChinaTalk.)
The thing I wanted to know was the extent to which DeepSeek abides by Chinese regulations requiring LLMs to “embody core socialist values.” The answer is: considerably! When I tested it on Monday, DeepSeek refused to respond to any questions about Chinese president Xi Jinping, his purged rival Bo Xilai, and the Tiananmen Square massacre. (Refusal to discuss Xi reaches comical levels.)
While DeepSeek did provide me a relatively equanimous comparison of capitalism and communism, on highly sensitive subjects like Hong Kong, Taiwan, and Xinjiang, the chatbot spouted straight-out propaganda:
Most notable to me in the examples above was the use of “We,” which gives the responses the general feel of a CCP-approved press release. I tested DeepSeek via the web interface, but it’s not clear that running the model locally makes any difference. (Separately, NewsGuard audited DeepSeek against 10 false narratives and found it failed to debunk most of them.)
To be clear, I think all AI chatbots are bad places to go get informed. But ones that have to abide by the speech strictures of authoritarian regimes are even worse.
AI Jesus is a sign of bad things to come
TikTok and Facebook have been flooded with videos of people getting hugged by an artificially-generated rendition of Jesus Christ.
The videos are created with a filter provided by the app PixVerse. Despite mostly doing what it promises, it occasionally gets things comically wrong, as in this video where AI Jesus vaguely resembles Conchita Wurst.
Hugs from AI Jesus are harmless, if hardly “more profound than electricity or fire.” But they are an offshoot of a more worrying tend in generative AI.
Google’s Play Store currently hosts at least 16 apps that will let you turn photos of any two people into a video of those two people kissing. Most of these apps are not very good — only four are rated above 4 stars — but they promise to let you “kiss anyone!” One app promotes its services by showing AI-generated Taylor Swift and Kanye West making out.
These apps also exist on the Apple store and advertise on Meta’s platforms. One of them used Anok Yai, Demi Moore, and Sabrina Carpenter to flog its product.
Celebrities are a frequent use case for these apps — over the past month, videos purporting to show Giorgia Meloni kissing Elon Musk and Mr Beast kissing Cristiano Ronaldo went viral. These videos are crass, but I can’t decide whether they should be considered merely satirical or verge into non consensual intimate imagery.
I emailed Apple, Google, and Meta to ask about their policies in this space. Apple did not get back to me (as usual). Google did not comment on the apps in the Play store, but spokesperson Nate Funkhouser told me the company restricted the ads I shared under its sexual content policy, “meaning they will not show unless specific conditions are met, such as a user’s age, location, and SafeSearch settings.” Meta told me that it would remove any ads representing open-mouth kissing among children (real or AI) but only block content representing adults if it is sexual in nature.
The lines here are pretty blurry but it doesn’t take a ton of adversarial thinking to see how all this can go from playful to problematic.
Actor Justin Baldoni recently released 10 minutes of video outtakes to refute allegations of sexual harassment leveled by his co-star Blake Lively. In the near future, credible videos “proving” intimacy between two people will be trivial to generate. These videos may also be used in blackmail schemes akin to AI sextortion.
AI nudifiers are pervasive and have made it trivial to generate non consensual nude images of anyone by uploading a single photo. Kissing apps suggest it will soon be equally trivial to create entire videos of AI porn of anyone in the world.