\n\n
OpenAI announced that it will embed an invisible text watermark—using its textGrain technique—into ChatGPT and Codex outputs for users in the European Union, a move required by the EU AI Act’s transparency rules. The watermark works by subtly biasing word choices to create a statistical signal that can be detected by a specialized tool.
Key developments include OpenAI’s rollout plan, which will cover all ChatGPT and Codex plans in the EU over the next few weeks, while offering an opt‑in toggle for API users worldwide. However, internal tests show the watermark can be weakened: substituting just 10% of words with synonyms drops detection from roughly 92% to 66%, and a 25% substitution cuts detection to only 17%.
Meanwhile, Google has broadened access to its SynthID Detector, launching a public portal on October 7 2026 that lets anyone upload images, video, or audio to check for watermarks. SynthID, which has already marked over 180 billion images and videos and 240 000 years of audio since 2023, now supports checks for watermarks from Google and partners such as OpenAI and NVIDIA.
Why it matters: Both moves aim to satisfy the EU AI Act, which mandates machine‑readable markings on AI‑generated content. OpenAI’s text watermark seeks compliance for text, while Google’s SynthID expands verification to visual and audio media, offering a broader toolbox for detecting synthetic content.
What to watch next: Analysts will monitor how effective OpenAI’s textGrain proves in real‑world use, especially as users experiment with paraphrasing to evade detection. In parallel, the adoption rate of Google’s public SynthID Detector and any extensions to cover additional partners or media types will indicate how quickly the industry can build a reliable, cross‑platform watermarking ecosystem.