OpenAI's Sora video generator stumbles with gymnastics

A notable weak point of Sora, which has been causing fits of laughter on social media, is its comically bad attempts at generating videos of gymnasts. Instead of smooth movements, the results show distorted and chaotic limbs, resembling scenes from a horror film rather than an Olympic performance.

Venture capitalist Deedy Das highlighted the issue by sharing clips of Sora's bizarre gymnastics attempts. Das jokingly wrote, "As cool as the new Sora is, gymnastics is still very much the Turing test for AI video." He posted several examples of the AI's uncanny creations, which have left viewers amused and unsettled.

The AI for Humans podcast also joined in, describing gymnastics as "the funniest way to break Sora." One video showed a gymnast flipping rapidly before landing in a bucket of mustard. The clip included a stream of yellow paint-like material trailing from the gymnast's feet, adding to the absurdity. Another video featured a Raygun-style breakdancer from the Paris Olympics, also plagued by limb distortions.

Even in less demanding scenarios, Sora has its flaws. It often struggles to spell words correctly in on-screen text, a common issue with AI-generated images. For instance, in one video, a mask worn by a character misspelled "obey" as "oeybey."

Popular YouTuber Marques Brownlee also pointed out the issue during his early access to Sora. While the AI could produce photorealistic human figures, it consistently generated garbled text. This discrepancy led one user to question how Sora could create lifelike visuals but fail at basic spelling.

Despite some impressive outputs, Sora's performance so far suggests it is less advanced than many had expected. The launch has added to the ongoing conversation about overhyped AI tools and their real-world limitations.