OpenAI Adds Opt-In ChatGPT Text Watermarks for EU Users Ahead of AI Act Deadline

OpenAI will introduce a text watermark for ChatGPT outputs in the European Union to comply with Article 50 of the EU AI Act. The documentation treats text watermarking as an active switch within API project settings rather than a universal default applied to every output. Images and audio will carry two distinct signals, combining C2PA metadata with Google DeepMind's imperceptible SynthID watermarks. The company previously reported its internal text watermark prototype was 99.9% effective in 2024, but held it back due to robustness concerns and potential bias against non-native English speakers. Regulators will need to clarify if an opt-in approach satisfies the machine-readable transparency requirements before the December 2, 2026 transition deadline they are bracing for.