OpenAI is rolling out invisible watermarks to ChatGPT and Codex text output in the European Union — starting this week. The feature, called textGrain, does not alter formatting or add visible markers. Instead, it nudges word selection at scale to embed a statistical fingerprint detectable only by OpenAI’s dedicated detector tool.
It’s not a global rollout: the EU is the first and only region where watermarking will be enabled by default in the coming weeks. Elsewhere, API developers can opt in now, but it remains off by default worldwide. The detector itself is opening to applications today, though access is tightly restricted: only approved researchers and expert organisations will get early access.
OpenAI’s own testing shows the system breaks down fast under real-world editing. Swap just 10% of words with synonyms in a 400-token passage, and detection drops from 92% to 66%. At 25% synonym replacement? Detection falls to 17%.
Shorter texts fare worse: for 200-token psychology responses, detection hovers around 80% (at a 1% false-positive threshold). For maths-heavy outputs, it’s lower still. The model has less lexical flexibility to encode the pattern. OpenAI stresses two things clearly: a missing watermark doesn’t prove human authorship, and a detected one doesn’t identify the user, prompt, or conversation.
It also doesn’t measure how much of a final document was AI-written versus edited. And yes, GPT-6 Astra’s benchmark performance holds steady with textGrain active. No measurable hit to quality. What happens next?
The EU rollout begins imminently. The detector application window is open. And OpenAI says it’s watching how reliably the watermark survives translation, summarisation, and paraphrasing. All common, low-effort ways to strip it.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.