AI companies face a junk data crisis that threatens physical AI and world model development

The AI industry has prioritized feeding models ever-larger quantities of data, but this approach is hitting a wall: low-quality 'junk data' that doesn't advance model development is now degrading performance, slowing time to market, and risking unpredictable outcomes in safety-critical applications like autonomous vehicles.

Physical AI and world models (systems that learn to operate in the physical world) require rich, multifaceted data that cannot simply be downloaded from the internet. Training these systems demands significant time and effort—often involving hours of simulated scenarios to teach robots and self-driving cars to distinguish typical situations from edge cases like wrong-way driving or obstructed views of pedestrians.

OpenAI's shutdown of its Sora video app and reassignment of that team was driven by a junk data problem: the world model lacked sufficient understanding of physics to generate realistic predictions, illustrating the real-world consequences of insufficient data quality.