AI companies stealing data from each other

Models start by training on data on the internet (which may or may not be properly copyrighted), other companies train models based on the outputs/distilled weights from the bigger models trained on stolen data

Additional comments:

The landscape of artificial intelligence development currently resembles a chaotic gold rush where the lines between innovation and replication are blurred. Large language models frequently ingest massive datasets from the open internet, often disregarding copyright protections or intellectual property rights. This creates a cycle where smaller or newer models refine their capabilities by consuming the outputs of established industry giants. The process often involves distilling weights from larger, pre-existing models, effectively turning one company’s expensive training efforts into another company’s shortcut. This systemic issue highlights a reliance on shared, potentially pilfered pools of information that sustain the entire sector. As these entities compete for dominance, the reliance on collective data practices creates a tangled web of origin stories for even the most advanced AI tools. Protecting original work remains a significant challenge as developers continue to harvest existing outputs to fuel their own technical progress.