Data Poisoning
-
AI Data Poisoning Alert: 0.01% Fake Training Text Can Increase Harmful Content by 11.2%
China’s Ministry of State Security warns of “data poisoning” as a critical threat to AI. Inaccurate, fabricated, and biased data corrupt AI training datasets, leading to flawed models and security risks. Even minimal data contamination (0.01% fabricated text) can significantly increase harmful content generation (11.2%). The proliferation of AI-generated content further amplifies the issue, creating a “post-contamination legacy”. Authorities highlight dangers in finance, public safety, and healthcare, where data manipulation can trigger market volatility, social panic, and incorrect medical advice.
