Sora AI video goes viral for being creepy — here’s why these anomalies happen

OpenAI’s Sora AI video generator produced a surreal and technically flawed video of a gymnast performing impossible movements, including sprouting extra limbs and temporarily losing her head during what was meant to be an Olympic-style floor routine.

Technical breakdown of the issue: The video synthesis errors stem from Sora’s fundamental approach to generating content through statistical associations rather than true understanding of physics or human anatomy.

• Sora creates videos by analyzing training data that pairs video footage with text descriptions
• The system makes continuous next-frame predictions based on the previous frame
• While Sora attempts to maintain coherency by looking ahead at multiple frames, it struggles with rapid movements and complex physics

Training data limitations: The model’s performance is directly tied to the quantity and quality of its training examples.

• The system relies on AI-generated labels to describe training videos
• Limited gymnastics footage in the training data leads to inconsistent predictions
• Complex movements prove especially challenging due to insufficient examples of precise limb-level movements

Industry perspective: This phenomenon extends beyond just Sora to other AI video generators.

• Runway’s Gen-3 and Hunyuan Video face similar challenges with complex movements
• These issues emerge when prompts exceed the boundaries of the model’s training data
• The problems highlight the imitative nature of transformer-based AI models

Technical terminology: Industry experts have begun referring to these AI-generated anomalies as “jabberwockies” – nonsensical outputs that fail to produce plausible results, distinct from mere hallucinations or confabulations.

• The term derives from Lewis Carroll’s nonsense poetry
• These errors represent complete failures in generating coherent output
• Similar issues have appeared in various AI video applications, from commercials to celebrity deepfakes

Future developments: Improvement in AI video generation will require significant advances in several areas.

• Massive amounts of well-labeled training data
• Increased computational power for processing
• Better understanding and implementation of physics rules
• More sophisticated approaches to maintaining consistency across frames

Looking ahead: While current limitations in AI video generation are evident, the technology’s rapid evolution suggests significant improvements are likely, though true physics-based “world simulation” remains a distant goal. The progression may mirror the development of AI image generation, which advanced from abstract shapes to highly realistic imagery in a relatively short time.