China Warns of AI Security Risks from Data Contamination

The Ministry of State Security warned of AI security risks from contaminated training data, a core challenge to AI safety. Data pollution, including false information and bias, undermines AI reliability, especially as AI integrates into China's economy.

The ministry highlighted "recursive pollution" where AI-generated errors contaminate future training data. Risks include financial manipulation, public panic, and medical misjudgments. They proposed stricter data supervision, risk assessments, and data-cleansing.

Professor Zhang Xi noted China's vulnerability due to a shortage of high-quality Chinese-language data, forcing reliance on lower-quality sources. The scale of training data makes manual inspection impossible, increasing contamination risks in fields like medicine and autonomous driving. He also warned of malicious data poisoning.

Zhang advocated for greater investment in domestic data collection and legal mechanisms to criminalize data poisoning. Professor Shen Yang defined data pollution as erroneous or manipulated content, weakening AI's reliability. He noted that malicious actors could manipulate AI, eroding public trust. The public must critically analyze AI-generated content.