On the Value of Human Ideas: What data poisoning research reveals about "autonomous" AI breakthroughs

Research into data poisoning reveals that even a small amount of poisoned data can significantly impact large AI models. For instance, just 250 poisoned documents, making up 0.00016% of training tokens, reliably implanted a backdoor in a 13B-parameter model. Similar success was seen in fine-tuning experiments, where 50–90 poisoned examples achieved over 80% attack success. This raises questions about intellectual property and credit in a future where AI synthesizes human ideas into breakthroughs.

This report uniquely highlights how a minuscule 0.00016% of poisoned data can reliably implant backdoors in large AI models, unlike previous discussions focusing on general data integrity.