A year ago, Will Smith devoured spaghetti and people laughed. The fact that artificial intelligence turns text input into moving images was impressive at the time, but the result was reminiscent of a horror comic. The actor is clearly visible, but obviously not real. In the short clips, the face appears distorted, almost monstrous, with misshapen noodles growing out of the nose and disappearing into the ears.
Eleven months later, no one is laughing anymore. On Thursday evening, Open AI showed what AI is now capable of. The text robot ChatGPT is now following the video generator Sora. The model creates videos of up to one minute that have nothing to do with Spaghetti-Smith. If you don’t look very closely, you could mistake the computer-generated clips for authentic recordings.
There’s a woman walking through Tokyo at night, the wet asphalt reflecting the neon-colored lights of the city, the streets reflected in the lenses of her black sunglasses. Three Golden Retriever puppies play in the snow, crystals and flakes fly through the air in slow motion, every hair of their fur moves with them. A drone circles over the rugged cliffs of California, in the golden light of the sunset the waves roll majestically against the rocks and burst into white foam.
They are graceful, almost cinematic videos – with one crucial flaw: not a single pixel is real. Sora combines the technologies of ChatGPT and the image generator Dall-E. The model puts together video snippets as if they were words. From the training material, Sora deduced how light and shadow relate to each other, what makes seemingly real movements and textures, and which physical laws need to be observed.
You can find mistakes with a magnifying glass
None of the dozens of clips Open AI shows on its homepage is perfect. If you know you are watching a synthetic video, you can quickly find errors. The woman walks a little too floaty, the people in the background are reminiscent of characters from a computer game. The dispersing snow changes direction in the air, the waves glide over the rocks rather than breaking against them.
It can also be assumed that Open AI has selected particularly successful and impressive examples. Sora is not generally accessible; currently only a select few security researchers and artists can play around with it. It also remains unclear what material the model was trained with and how long it takes to create a video.
Nevertheless, it would be wrong to focus exclusively on Sora’s weaknesses. Technology will never be as bad as it is today. Sora is the Japanese word for heaven and is intended to illustrate “limitless creative potential,” say its creators. You don’t need a vivid imagination for this. AI development is rarely linear, but if you remember how ridiculous Will Smith looked, you can vividly imagine what could be possible in a year.
Sora will be abused
This is fascinating and frightening at the same time. Anyone who wants to tell visual stories will love tools like Sora; there may soon be no technical limits to creativity. But the risks are just as obvious and drastic. For one thing, AI video generators could put out of work many of the people whose work they were trained to do: motion designers, illustrators, commercial filmmakers, drone pilots, cameramen. Sora won’t win an Oscar, but even mediocre, uninspired videos are good enough for some purposes – and they cost a fraction.
Secondly, Sora will be abused just like any other technology before it. It was only on Wednesday that it emerged that Chinese, Russian and Iranian hackers had tried to use tools from Open AI and Microsoft for criminal purposes. The same thing awaits Sora. In nine months the USA will elect a new president, social media will be flooded with lies and propaganda – possibly also with synthetic videos.
If Open AI is to be believed, the company is trying to minimize the dangers of the technology. Certain text entries should be blocked automatically, such as depictions of violence and sexuality or possible copyright violations. With the Dall-E image generator this works unreliably; many protective mechanisms can be overridden with a little skill. In theory, the sexualized AI fakes of Taylor Swift that went viral at the end of January shouldn’t have existed at all.
In any case, the researchers at Open AI apologize as a precaution. Despite all the precautions, we cannot predict all the ways people will misuse our technology. says the announcement. One question will probably have to be asked more often in the future: Is this real?