From Will Smith Eating Spaghetti to Now: A Short History of AI Video

There is a video that every AI researcher remembers: Will Smith eating spaghetti, March 2023. A melting, many-fingered horror devouring pasta that flowed like lava — generated by an early open-source model, and so gloriously bad that it became the unofficial benchmark of the field. "Wake me up when it can do the spaghetti."

Nobody laughs anymore. Here is how three years turned a punchline into a verification crisis.

2023: The uncanny year

Early text-to-video models produce two-second clips at postage-stamp resolution: warping faces, limbs that multiply, physics from a dream. Runway's Gen-2 and Pika make "AI video" a consumer phrase, but everything they make is recognisably synthetic within a frame or two. Detection, at this point, needs no software — a pulse is enough.

2024: The Sora shock

In February 2024, OpenAI shows Sora: minute-long, coherent, photoreal-adjacent clips that break the collective assumption of "we have years before this gets scary". The demo videos — a woman walking through Tokyo, historical footage that never happened — circulate with the energy of a moon landing.

The rest of the year is an arms race. Kling arrives from Kuaishou and floods social platforms with cheap image-to-video. Runway ships Gen-3, Google previews Veo. By December, Sora is publicly available. The first viral incidents follow immediately: fake disaster footage, fake celebrities, fake "phone videos" of events that never occurred.

2025: Good enough to fool everyone

Three shifts turn a novelty into an epidemic:

• Audio arrives. Google's Veo 3 generates video with synchronized sound — speech, ambience, effects. The silent-movie era of AI video ends, and with it another easy tell.
• Fakes go social-native. Sora relaunches as a feed app with a cameo feature: your face, your friends' faces, one tap. AI video stops looking like stock footage and starts looking like someone you know.
• Open source catches up. Alibaba releases Wan openly; anyone can generate video on their own hardware — no watermarks, no content rules, no logs. The compliance story that big labs tell regulators quietly stops covering the actual threat.

Somewhere in this year, the default question flips: from "is this AI?" (asked about weird clips) to "is this real?" (asked, increasingly, about everything).

2026: The verification era

Today's frontier — Veo, Kling 3.x, Seedance — produces clips that pass every casual eyeball check: correct hands, plausible physics, imitated camera shake, matching audio.

The visual tells still exist (we catalogued nine of them), but they are subtle, and they shrink with every model release.

Sora is not on that list any more. OpenAI retired the app in April 2026 and is closing the API in September, folding the compute back into other work. It is a strange kind of ending: the model that started the panic is gone, and its clips are still everywhere — millions of them, mostly with the watermark cropped off, outliving the thing that made them. Generators come and go; what they produced does not.

What does not shrink is the statistical gap. A generator paints pixels; a camera captures photons on a physical sensor. The manufacturing marks of generation are baked into every frame — invisible to eyes, readable by detector models trained on thousands of examples. Detection has moved from eyesight to statistics, which is why it now takes software.

The spaghetti test, revisited

The distance from spaghetti-horror to "indistinguishable at feed speed" took roughly three years. Whatever the next three bring — longer clips, real-time generation, perfect UGC mimicry — the direction is not in question. The tools to make fake video will keep getting better. The habit of checking video is the part you control.

It takes thirty seconds. And if a clip matters enough to share, it matters enough to check first.

Both spaghetti clips are public domain — output of a computer algorithm carries no copyright. The 2023 original is by chaindrop on Reddit, the 2025 one by Javi Lopez; both via Wikimedia Commons.