Will Smith has become the benchmark of AI-generated video. (Photo by Kevin Winter/Getty Images for AFI)
Getty Images for AFI
AI-generated video just took a great leap forward with Veo 3, Google’s new AI video generation model, and the internet quickly administered the “Will Smith Eating Spaghetti” test.
What Is The “Will Smith Eating Spaghetti” Test?
About two years ago, AI-generated videos of Will Smith eating spaghetti went viral, with commentators describing them as “demonic.”
Those videos were nightmare fuel, featuring shape-shifting inhumans who simply could not mimic the subtleties of eating pasta.
Today’s chatbots can comfortably pass the Turing Test, and could be easily mistaken for an annoyingly upbeat, affirmative human.
The Will Smith Spaghetti test is something of a Turing Test for video generators, a baseline for testing how powerful AI models really are.
Veo 3 is capable of producing terrifyingly lifelike videos, containing AI-generated sound, and one X (Twitter) user quickly put the model to the Will Smith test.
I wouldn’t say that Veo 3 passed with flying colors—visually, it’s mildly uncomfortable, but with the audio playing, it crosses the line into disgusting (why is his spaghetti so crunchy?).
However, it’s infinitely better than the previous models. Many users began posting short videos generated by Veo 3, and at first glance, they are extremely difficult to differentiate from real footage.
Hallucinations are still there, but have been minimized—most of the subjects have the correct amount of fingers, and their body language can pass as human—rubbery limbs and grotesque oddities have been largely ironed out.
A video of a comedian telling a “dad joke” is one of the most convincing—the tells are becoming more subtle, difficult to spot during a quick scroll of the timeline.
Currently, Veo 3 videos are capped at 8 seconds, and I suspect that generating a video longer than that would quickly unravel, exposing the illusion. Otherwise, these short clips function surprisingly well, largely conforming to physics.
Clues can still be found in the uncanny movements of background characters, but even those have greatly improved.
That being said, progress doesn’t come cheap—access to the feature is locked behind Google’s $249.99-a-month AI Ultra plan.
More weaknesses and hallucinations will surely be uncovered as users test the limitations of the software, but the improvement is striking—Will Smith’s spaghetti eating has advanced from “demonic” to merely “gross.”
One has to wonder what effect the advanced wave of AI-generated videos are going to have on the internet.
Social media users will have to train their eyes to spot the fakes, but not everyone is going to be able to do so, and scrolling culture doesn’t encourage careful examination.
AI is also being used for face filters, altering real videos into deep fakes, making the deception all the more difficult to spot.
I don’t think video generators can come close to matching a strong, thoughtful performance from a good actor, but the technology is good enough for a quick video, good enough for a disposable TikTok or Instagram Reel.
Montages, music videos and ads—anything with quick cuts can now be faked.
Previous attempts at AI-generated ads were easy to spot, but the new generation of video generators have outclassed these experiments.
What does this mean for the flesh-and-blood users of the internet?
The Dead Internet Theory Is Coming
The Dead Internet Theory started life as a meme, inspired by a paranoid 4chan post proposing that the majority of online activity had been replaced by bots.
This wasn’t true at the time, but today, it risks manifesting into reality.
The internet’s most famous viral tweets and TikToks did not emerge from bots, and still don’t. AI-generated content doesn’t have the personality and insight to inspire viral trends.
However, there is a danger of the web slowly filling up with AI-generated slop of all kinds—photos, videos and comments, crowding out all the human activity, like plastic particles contaminating the ocean.
It’s already happening on Facebook, and AI-generated videos are starting to creep into TikTok and YouTube, even if they don’t define the culture of these platforms.
The realism of the latest models inspired fear from many internet commentators, who instantly saw the terrible potential of the technology—AI-generated doppelgangers can be used to make incriminating videos that are difficult to distinguish from real footage.
This is an exceedingly powerful technology, with the potential for terrible abuses—at best, it will unleash a tsunami of mass-produced spam on the internet that risks overwhelming the human-crafted work that we really want to see.
However, AI-generated video is still in the “novelty” phase, and many of the most impressive videos are only impressive because they are artificial. Would we care to watch them if they were real?
Arguably, the most interesting pattern to emerge from the AI-generated era of the internet is the rise of absurd, surreal imagery, like Shrimp Jesus and Italian Brainrot.
Novelty is impossible to sustain, and internet users are likely to get tired of the wave of AI-generated content.
Soon, many users could migrate to platforms that manage to block the “sloppification” of the web, while others stay behind, surrounded by artificial content.
AI-generated Will Smith successfully eating spaghetti is a sign of incredible technological progress, but it could also be viewed as the canary in the coal mine.
MORE FROM FORBES
