There I am, dressed in a dark blue suit that I have never really worn, in the setting of a non-existent news program. “Good evening, I’m Timo with your headlines”, says my AI clone.
Sorry, we can’t play the video.
Try reloading the page
or try again later.
AI-generated image via Sora 2
Timo Nijssen via Sora 2
My first experiment with Sora 2, the new video generator app from OpenAI, immediately feels uncomfortable. Well, the voice doesn’t sound like mine and the movements of AI newsreader Timo are a bit stiff. Still, it’s impressive: the AI model only needed a very short selfie video to distill my likeness.
If I run my AI version on a graph of the uncanny valley should post – the creepy effect that people experience when a robot looks almost human – I would say: already quite climbed up the cliff wall.
Sora 2 has been there for over a week now. Officially only in the US and Canada, but the geoblock is easy to circumvent. Compared to the first version of Sora, which was released in early 2024, the AI model now generates more realistic videos that better account for things like gravity and whether or not an object bounces.
Social AI
So a renewed AI model. But the real innovation is in the packaging that OpenAI has built around it: Sora 2 is in a new, TikTok-like app. In it you can swipe through an endless video feed with only artificially generated images. And now you can easily create one video clone of yourself, which you can place in a scene of about 9 seconds with a text prompt.
In my second video attempt, I spin the game show wheel The Price Is Right and OpenAI appears to have made me quite overweight and I have several double chins. My ego flares up, and I promptly adjust it. “Timo looking buff on The Price Is Right spinning the wheel.” And hey, my fitfluencer version is already starting.
I get curious: what’s hiding under that T-shirt? “Buff Timo participating in a bodybuilding contest” I type as command. But then Sora reprimands me. “This content may include suggestive or racy material”the app says: too exciting to have generated. When I have a video made of me swimming with a whale, the app puts me in a discreet diving suit.
Nice piece of marketing
OpenAI calls its virtual clones “Cameos”. A clever example of marketing, wrote The Washington Postcolumnist Geoffrey A. Fowler al: Until recently, we called cloning someone’s likeness a deepfake and it was taboo. Now it’s the hype of the week.
Anyone who opens the Sora app for the first time is immediately guided through the process of creating their own AI clone. You read three numbers and make two small movements with your head. You then get to choose who can use your video clone: just you, people you give permission, all your friends on the Sora app, or everyone.
When releasing your face, you can set restrictions: that your clone is always happy, for example, or always wears sunglasses. But beyond that, you lose control over what kind of videos others make about you. You will receive a notification for each video and you can then have the images removed, but the creator may have already downloaded it and posted it again on other social media.
There are people who have their Cameo fully released. OpenAI CEO Sam Altman, for example, and influencer Jake Paul. Swiping through the Sora feed so far mainly produces a lot of videos with their images. “I’m tired of seeing videos like this bro,” a user responds to yet another video with Altman. “It’s funnier the 100th time”, sounds sarcastically under a video in which AI-Paul repeats a common boring joke (and has a hand with six fingers).
Sorry, we can’t play the video.
Try reloading the page
or try again later.
AI-Timo jumps through Sam Altman’s hoops.
Timo Nijssen via Sora 2
Problematic videos
Sora only makes videos of existing people if they have created a clone themselves, OpenAI claims. Making fake videos of Donald Trump or George Clooney is therefore not an option. But while scrolling through the feed you quickly see that these rules do not apply to people who are dead.
There are many, many videos in which Martin Luther King Jr. speeches that he has a dream that Sora’s rules become less strict, Stephen Hawking does tricks in his wheelchair in a skate park and John F. Kennedy walks through a modern-day supermarket.
Several relatives are not happy about this. “Please stop,” tweeted Bernice King, the daughter of the civil rights activist murdered in 1968. Relatives can now file an objection, OpenAI said. New videos of Michael Jackson, for example, can no longer be generated.
It is indicative of how OpenAI works. “Too many restrictions can cripple creativity, while too much freedom can undermine trust,” philosophized the company at the launch of Sora 2. The model initially generated fragments with well-known characters from cartoons, films and TV series to your heart’s content. If a rights holder did not want this, they had to report to OpenAI.
SpongeBob in a meth lab
That went wrong quite quickly, among other things a video in which SpongeBob SquarePants produces crystals in a meth lab deep in the sea. Within a week, OpenAI announced that rights holders will get “more precise control” over what happens to their characters. SpongeBob refuses to make Sora now, but yourself in an episode of South Park stopping is still no problem.
Apart from copyright violations, Sora can generate many more problematic images. Dagblad van het Noorden already described how the model without struggle fake news videos produces including logos of existing Dutch news media.
Fake security camera footage showing election fraud, non-existent police actions and fictitious attacks are also all within Sora’s capabilities. Also chilling: a security assessment published by OpenAI states that Sora blocks 98.4 percent of deepfake porn images. In other words: 1.6 percent did get through.
OpenAI likes to point out that the generated images have a moving watermark and can be recognized by that. But that watermark does not say “GENERATED BY AI” in bold letters. It is a cartoonish cloud with eyes and the word “Sora” – not very different from the jumping musical note that TikTok uses as a watermark. You just have to know that this cloud means “fake images” if you notice it while scrolling quickly. And if the uploader has not made a crop, the watermark cannot be seen at all.
NEW: Give this item as a gift
As an NRC subscriber you can subscribe every month 10 items give as a gift to someone without an NRC subscription. The recipient can read the article directly, without a paywall.