Google's Veo 3 nails the spaghetti test, but makes it crunchy

Two years ago, a nightmarish clip of a warped, AI-generated Will Smith frantically shoving fistfuls of spaghetti into his face became the internet's favorite symbol of how broken AI video was. This week, Google tried to recreate that legendary benchmark with its latest model. The visuals look drastically better, but there is a bizarre catch: the fake Will Smith is crunching on his pasta like it's a bowl of potato chips.

That odd audio artifact comes courtesy of Veo 3, Google's new video generation model unveiled at I/O. For the past two years, AI video tools have produced mute, dreamlike snippets. Veo 3 changes the playing field by generating eight-second high-definition clips with synchronized dialogue, voice acting, and ambient sound effects baked right into the output.

Adding lines you never wrote

Veo 3 does not just generate ambient sound; it often decides what characters should say completely on its own.

In hands-on testing by The Verge, a prompt describing a police scene produced an officer making a firm gesture and declaring they needed to clear the street, even though the original text prompt mentioned zero dialogue. Another test featuring two cartoon cats on a pier resulted in the characters spontaneously complaining that the fish were not biting, accompanied by jaunty background music.

That unprompted realism makes for convincing outputs. Alejandra Caraballo, a clinical instructor at Harvard Law School's Cyberlaw Clinic, produced a clip of a television anchor announcing the death of US Secretary of Defense Pete Hegseth. The official is alive, but the broadcast looked and sounded disturbingly authentic. Meanwhile, a Reddit thread featuring AI-generated characters protesting the prompts used to create them gained over 50,000 upvotes, with models delivering spoken lines alongside hospital breathing tubes and simulated gunfire.

Google has built in guardrails to block the most obvious hazards. Prompts trying to depict Joe Biden tripping, a news anchor reporting a presidential assassination, or a tech CEO laughing in a shower of dollar bills are rejected immediately. Supplying photos of yourself to manufacture bespoke talking-head clips also gets blocked by the safety filters.

Crunchy pasta and slop machines

Why did the spaghetti crunch? Generative AI is essentially a pattern-matching engine. The training data behind Veo 3 likely paired chewing mouths with crunchy sound effects far more often than soft noodles, causing the model to default to the wrong audio texture when presented with eating motions.

Beyond strange audio glitches, Veo 3 makes churning out low-quality digital filler easier than ever. The internet is already overflowing with toddler-focused YouTube slop, like endless loops of poorly animated monster trucks driving into vats of colorful paint. Veo 3 can generate that exact formula, sound design included, in just a few minutes with simple text instructions.

For now, the tool has clear boundaries. You need an expensive subscription to Google's AI Ultra plan to access it. Furthermore, if you attempt to extend an eight-second clip into a longer scene, the system falls back to Veo 2, instantly stripping out all sound.

Why it matters

We think Veo 3 shifts the competitive benchmark for AI video from pure pixel fidelity to integrated audiovisual storytelling. Until now, creators had to jump through multiple software hoops, stitching silent Runway or Pika clips together with third-party voice generators and sound libraries. Condensing motion, lip-syncing, voice acting, and Foley work into a single text prompt changes the economics of short-form production overnight.

However, Google's decision to let the model invent its own dialogue is a double-edged sword. It creates spontaneous, lively clips, but it also means the model can inject misleading or alarming spoken narratives into seemingly innocuous scenes. Even with strict filters around public figures, giving everyday users the power to generate broadcast-ready disaster coverage and convincing bystander audio within seconds creates an undeniable headache for trust and safety teams.

Right now, high subscription fees and the rigid eight-second ceiling will keep it out of full cinematic workflows. Instead, we expect Veo 3's earliest real-world footprint to show up exactly where creators prioritize speed over substance: in the relentless, automated churn of social media feeds and low-effort children's entertainment.