In 2023, the internet couldn’t stop laughing-or cringing-at a warped, jittery AI-generated clip of Will Smith awkwardly “eating” spaghetti. The so‑called spaghetti test became an unofficial benchmark for text‑to‑video AI: could a model render a human convincingly performing a simple, physical act without breaking anatomy or physics? Back then, the answer was a resounding no. Faces melted, limbs multiplied, noodles levitated, and chewing looked like a bad claymation loop.
Fast-forward to late 2025, and that benchmark is all but obsolete. Models like OpenAI’s Sora 2, Google’s Veo 3.1, and xAI’s Grok Imagine now crank out scenes with photorealistic motion, coherent physics, and synchronized audio capable of fooling the casual viewer into thinking they’re watching real footage. Sora 2’s architecture extends the transformer‑based diffusion pipelines but integrates advanced physics simulation layers that enable objects to interact with believable weight, inertia, and collision dynamics. Veo’s latest iteration adds frame‑to‑frame motion consistency and naturalistic camera movement, which cuts down on the “floaty” artifacts that plagued earlier generations.
The leap in realism is most obvious in facial rendering: Where 2023’s ModelScope spaghetti clip produced caricatured features, Sora’s “cameo” system can ingest high‑resolution facial scans and replicate micro‑expressions, skin pore detail, and subtle eye movements. This personalization pipeline encrypts facial and audio data at rest and in transit, but it also raises deepfake concerns. As OpenAI confirms, “Uploaded cameos are kept by OpenAI to use for new Sora videos… Uploads are automatically deleted within 30 days after users delete them.” Even with opt‑out controls, the ability to drop a public figure’s likeness into any scene has already set off backlash from talent agencies and lawmakers.
That backlash is amplified by the sheer volume of copyrighted material embedded in these models’ training data. Public feeds for Sora, for example, were initially overflowing with clips featuring Nintendo characters, SpongeBob, and Star Wars icons-clear signs of training on massive datasets of protected IP. Rights holders from Disney to Warner Bros. say being forced to “opt out” flips copyright law on its head. As Warner Bros. put it, Decades of enforceable copyright law establishes that content owners do not need to ‘opt out’ to prevent infringing uses of their protected IP.
Technically, removing such content from a model is daunting. As analysts note, unlearning copyrighted data requires retraining from scratch-a process that’s both expensive and potentially destructive to the model’s capabilities. That’s why bypassing guardrails has become a sport among users-misspelling names or using descriptive proxies can slip past keyword filters, yielding uncanny recreations of banned characters or real people.
The competitive race is another driver of this rapid, messy rollout. Google’s Veo 3 set the bar high, with coherent dialogue and environmental audio baked into its generation stack, while xAI’s Grok Imagine swaps out quality for speed, with a two‑step text‑to‑image‑then‑animation workflow optimized for “fun and shareable content.” Sora’s direct text‑to‑video approach, capable of up to one‑minute clips, positions it as the most cinematic of the lot, but also the most legally exposed.
Under the hood, these systems share core generative principles-large multimodal transformers, trained on billions of paired text-video samples and fine-tuned with reinforcement learning from human feedback to improve adherence to prompts. The leap in realism is due to innovations in temporal coherence modules that ensure objects maintain consistent shape and lighting across frames, and differentiable physics engines that let the model “reason” about gravity, friction, and fluid dynamics. In their latest iterations, noodles drape themselves naturally over a fork, sauce clings with believable viscosity, and chewing follows the anatomically correct motion of the jaw-all characteristics driven by these physics-sensitive architectures.
Yet the more convincing the output, the higher the stakes. Legal experts warn that if courts side against “fair use” in ongoing cases like New York Times v. OpenAI, companies may need licenses for every copyrighted element their models can reproduce. That could force a fundamental redesign of training pipelines or push the industry toward synthetic‑only datasets. Meanwhile, studios are exploring monetization schemes everything from granular opt‑in controls to revenue‑sharing for AI-generated content as a way to channel this technology without eroding IP value.
To the tech-savvy observer, the Will Smith spaghetti saga now reads like some kind of primitive artifact from AI’s adolescence. The new clips pass the spaghetti test with ease-but they also expose the unresolved engineering, ethical, and legal challenges of a medium hurtling toward photorealism.
