OpenAI introduces invisible watermarking for ChatGPT text in the EU
OpenAI is adding invisible watermarks to text generated by ChatGPT and Codex for users in the European Union to comply with the EU AI Act. The company uses a technique called textGrain that embeds a statistical signal in word choices, rather than adding visible labels or metadata. OpenAI says the watermark can be weakened by editing and is planned to be rolled out gradually in the coming weeks, while API customers worldwide can opt in.
OpenAI’s system, called textGrain, alters word selection slightly so a statistical pattern emerges across many choices. The mark travels with copied text rather than relying on removable metadata. The company says it may later release the technology as open source.
Rollout is gradual in the EU for eligible ChatGPT and Codex outputs across subscription tiers. API users worldwide may opt in, though it remains off by default. Detection weakens after edits, translations, or shorter outputs, and works less reliably for constrained text such as math.
The EU rollout could give regulators, platforms, and readers a new tool for spotting some AI-written text, while API opt-in may extend the signal beyond Europe. Yet because editing, translation, or brevity can weaken detection, people may overtrust results or misjudge authorship. Researchers and institutions gaining limited access may study its reliability, but the mark cannot show who prompted the text or whether it is accurate, so it may work best as one clue among many.