Google’s Veo 3 AI Video Model Nails the Spaghetti—But Crunches the Sound

The “Will Smith eating spaghetti” benchmark returns, exposing AI’s lingering audio quirks

Google just upped the AI video game with Veo 3, the first major model to generate synchronized audio alongside visuals. The system produces 8-second HD clips complete with voices, dialogue, and sound effects—a leap beyond silent predecessors. But when put to the ultimate test—the infamous “Will Smith eating spaghetti” benchmark—it stumbled into a crunchy, surreal glitch.

“The audio sounded like Smith was devouring gravel, not pasta,” posted developer Javi Lopez after testing Veo 3 with the meme-worthy prompt.

The spaghetti benchmark began in March 2023 with a grotesque, low-res AI video from ModelScope, featuring a distorted Smith slurping noodles. It became shorthand for early generative video’s flaws, even prompting Smith himself to parody it in February 2024. Though Runway’s Gen-2 later surpassed ModelScope technically, the spaghetti clip endured as a cultural touchstone.

Lopez’s Veo 3 test initially seemed promising: the video rendered Smith’s face and movements convincingly. But the audio misfired, layering crunching sounds over the pasta—a likely training data artifact. “Models overindex on frequent associations,” explains an AI researcher familiar with Veo 3’s architecture. “If ‘eating’ in training data disproportionately links to crispy foods, you’ll get nacho sounds on spaghetti.”

Why AI still chews weird

Generative models live and die by their datasets. Skewed representation—like an overabundance of crunchy-eating audio samples—can warp outputs. Veo 3’s spaghetti snafu mirrors earlier text-to-image oddities, where “CEO” prompts defaulted to white men or “nurse” images skewed female. The fix? More balanced data—and lots of pasta-eating footage.

“Every AI benchmark reveals something new,” says Lopez. “This time? We need better chewing sounds.”

For now, Veo 3’s 8-second clips impress despite hiccups. But as the spaghetti test proves, even cutting-edge AI still bites off more than it can chew—literally.