EU AI Act mandates watermarking for all AI text, audio, images from August 2nd

The EU AI Act comes into full force on August 2nd, imposing strict requirements on all AI systems that generate or modify text, audio, or images. Providers must implement two layers of detection: cryptographic metadata tagging (e.g., C2PA) for downloadable files like PDFs, and statistical watermarking embedded in the text itself. Simply adding an "AI Generated" label is insufficient—the output must be machine-readable and robust enough to resist tampering. Fines for non-compliance start at 35 million Euros (about $40 million) or a percentage of annual revenue, whichever is higher. These rules apply to any service accessible to EU citizens, even tourists using VPNs, making them global in scope.

Open-source tools and models are deeply affected. Platforms like Ollama, llama.cpp, vLLM, ChatGPT, Claude, Copilot, Cursor, and Stable Diffusion Web UI must comply if they serve EU users. High-profile models such as Qwen 3.6, Deepseek Flash, GLM, and Kimi are classified as "GPAI with systemic risk" and cannot claim exemption. The Act also introduces a code of conduct that, while voluntary, increases fine risk if ignored. Providers must maintain risk assessment documentation, compliance records, and detection services. The technical challenges—especially statistical watermarking of text without degrading quality—are substantial, and the law’s broad reach means even hobbyist projects could face legal exposure if they reach EU users.