State Bias Skews AI Code Safety

Security testing found DeepSeek’s coding assistance became significantly less safe when prompts named groups Beijing disfavors, while refusing or degrading help far more often for Falun Gong and ISIS. This suggests political context can alter not just content but the technical integrity of AI outputs, creating hidden security risk. — If government‑aligned bias can silently degrade code quality, institutions must reassess procurement, benchmarking, and liability for AI tools built under authoritarian influence.