On Tuesday, Google unveiled Veo 3, a new AI video synthesis model capable of achieving something no major AI video generator has managed before: producing a synchronized audio track along with video. Between 2022 and 2024, early AI-generated videos were silent and generally very brief. Now, Veo 3 delivers eight-second high-definition clips complete with voices, dialogue, and sound effects. Following the launch, people immediately posed the obvious benchmark question: How well can Veo 3 fake Oscar-winning actor Will Smith eating spaghetti? A quick recap: the "spaghetti benchmark" in AI video began in March 2023 with an early, rather unsettling AI-generated video created using an open-source synthesis model called ModelScope. That spaghetti example became so well known that Smith spoofed it nearly a year later, in February 2024. Here’s a reminder of what the original viral video looked like: What’s often forgotten is that at the time, the Smith parody wasn’t produced with the best AI video generator available—a model called Gen-2 from Runway had already delivered higher-quality results, though it wasn’t publicly accessible yet. Still, the ModelScope version was odd and memorable enough to become a reference point for early AI video limitations as the technology advanced. Earlier this week, AI app developer Javi Lopez responded to fans eager to revisit the spaghetti test using Veo 3, sharing his findings on X. However, when watching the results, the soundtrack seemed unusual: the fake Smith sounded like he was crunching the spaghetti. This glitch stems from Veo 3’s experimental capacity to add sound effects, likely because its training data included numerous examples of chewing accompanied by crunching noises. Generative AI models function as pattern-matching prediction systems, relying on sufficient training data across various media types to produce convincing outputs.

When certain concepts are over- or under-represented in that data, it leads to peculiar generation artifacts like this. We also ran the prompt ourselves on Veo 3, but “Will Smith” was blocked by Google’s content filters. Using the prompt “A black man eating spaghetti, ” however, produced a similar crunchy sound effect (Lopez may have had early filter-free access, or experimented with prompt variations that slipped through). Veo 3 impresses with its ability to generate coherent dialogue and music, already inspiring numerous striking examples on X. Not wanting to stop at just a video of a man eating very al dente noodles, we tested whether the figure could sing and eat simultaneously by prompting: “A man singing an English-language comedy opera about spaghetti at a kitchen table while eating it. ” We’ve made significant progress since 2023, and AI video generators will continue improving in realism and functionality. Were it not for Veo 3’s current celebrity filter, we might easily craft videos of Smith singing—or doing almost anything else—which highlights potential concerns around AI video technology. The cultural singularity is approaching fast. On that note, we recently conducted our own extensive series of video generation tests with Veo 3 and will share those results soon in a dedicated feature. For now, consider this a brief update on the Fresh Prince of noodletime. Bon appétit!