Initially an outlandish internet prank, an artificial intelligence-generated image of Will Smith trying to grip spaghetti with his hands, in January 2023, model- Scope opened up the internet to the concept of a ‘spaghetti test’ with a video that was dismally bad and distorted. This video showed Smith’s face in a warped fashion, had his hands moving around chaotically, and he was not able to eat the spaghetti. This video highlighted the challenges of the first-generation text-to-video technologies, with creating an accurate and lifelike portrayal of a person’s face and movement continuing to be an enormous technical challenge similar to climbing Mount Everest.
Two and a half years from the last time, when the line between what is virtual and what’s real changed very little, we have arrived at a place where everything is made far more realistic with Google’s Veo 3.1 and OpenAi’s Sora 2. With Veo 3.1, we will be able to emulate levels of detail and fidelity for all real-life objects we were once bound to our imaginations. Examples include micro-expressions, the delicate motions associated with eating, and the chaotic physics of spaghetti. This has resulted in a virtual environment that is nearly indistinguishable from the physical world in photorealistic texture fidelity. Similarly, Sora 2 has utilized temporal attention layers in keeping the objects consistent across frames of a series and has incorporated a generative audio pipeline to keep all audio relative to all video in sync. Both will greatly enhance your enjoyment of your spatial experiences in virtual reality and will continue to get better with time as technology evolves.
This leap in quality has intensified the arms race between OpenAI, Google, and Elon Musk’s xAI. Sora and Veo 3 can synthesize full-motion video directly from text prompts, while Grok Imagine synthesizes still images to be animated and realistic audio is a challenge. But xAI bills speed as its differentiator, generating clips in seconds-an advantage for meme culture but a limitation for cinematic realism.
The technology’s new precision has collided head-on with legal and ethical boundaries. Sora lets users upload facial scans to insert themselves-or others-into the app’s generated scenes. That capability spurred a trademark suit from Cameo, the celebrity video platform, which argued that OpenAI’s branding in “blatant disregard for the obvious confusion it would create” put it in direct competition with its marketplace for authentic celebrity greetings. More troubling for rights holders is Sora’s opt-out policy for copyrighted characters and likenesses, which studios argue inverts the default protections of copyright law.
Warner Bros. responded flatly: “Content owners do not need to ‘opt out’ to prevent infringing uses of their protected IP.”
The legal stakes go well beyond trademarks. The entertainment industry’s challenge to Sora parallels the open question of whether training on copyrighted material without permission is infringement. In ongoing litigation against OpenAI, the company relies heavily on a “fair use” defence, arguing that model training is transformative and outputs are different in character and purpose from source works. But the Supreme Court’s cabining of “transformative” in *Warhol v. Goldsmith* makes that argument more difficult, at least where outputs serve the same expressive purpose as the originals.
Regulators are starting to respond. With the implementation of the newly passed TAKE IT DOWN Act, the criminalization of distributing non- consensual intimate images (including artificial intelligence-generated deepfake content) now carries up to three years imprisonment. Additionally, elected officials are considering an increase in the regulatory framework for Artificial Intelligence, which ranges from implementing a possible 10-year moratorium on the regulation of Artificial Intelligence at the state level, to requiring hosting companies to exercise reasonable care when allowing the posting/distribution of synthetic media. The aim here is to tackle not only overt abuse but also the more insidious, looming threat of political misinformation-an area where hyper-realistic video could be weaponized.
On the technical side of things, the same advances that make the spaghetti test a passable feat in itself further make deepfake detection harder. Early systems of detection relied on catching frame-level inconsistencies or compression artifacts. With diffusion-transformer hybrids creating temporally coherent frames in concert with naturalistic motion blur, the detection shifts to physiological signal analysis-tracking micro-pulse patterns in skin tone, for example-or blink dynamics that may not be visible to the naked eye.
Even brands are leaning into the tech. Coca-Cola’s newest holiday campaign blended assets from Sora, Veo 3, and Luma AI, showing corporate adoption that is outpacing regulatory clarity. Today’s AI labs must not only render realism but also incorporate compliance layers that can authenticate provenance, filter prompts, and watermark outputs without limiting creative potential.
The spaghetti test, which started out as a meme, is now considered a turning point in the development of synthetic media from a glitchy curiosity to a tool that can change political discourse, advertising, and entertainment. The question is no longer whether AI can make Will Smith eat spaghetti convincingly. It’s whether society can digest the implications of that capability.
