OpenAI is adding an invisible ChatGPT watermark to chatbot text and Codex coding output across the European Union. Yet its own tests show that swapping a quarter of the words for synonyms drops detection to just 17%.
The move answers the EU AI Act, the bloc’s AI rulebook, which demands machine-readable labels on AI-generated text. Still, OpenAI admits the technology is immature, so its detector stays out of public hands for now.
How Strong Is the ChatGPT Watermark Once Someone Edits It?
According to the announcement, the system, called textGrain, nudges the model’s word choices to leave a hidden statistical pattern. A separate detector scans passages for that pattern.
The watermark lives in the wording itself, OpenAI explains in its FAQ. It adds no hidden characters, invisible spaces, or odd punctuation. That means the common trick of scrubbing hidden characters has nothing to remove.
Length matters a lot. OpenAI measures text in tokens, the word fragments AI models process. One token equals about three-quarters of an English word.
In OpenAI’s tests, the detector caught about 66% of unedited answers with roughly 150 words. At around 300 words, the rate rose to about 92%.
Light editing weakens the ChatGPT watermark as well. In 300-word texts, swapping one word in 10 for a synonym cut detection from 92% to 66%. Replacing one word in four pushed it down to just 17%.
The watermark barely moved benchmark scores for OpenAI’s GPT-6 Astra, the company’s latest frontier model.
Why Won’t OpenAI Let the Public Run the Detector?
Instead of a public tool, OpenAI is taking applications from approved researchers and expert organizations. The company says a positive result reveals no user, prompt, or conversation.
A negative result settles little, too. Short, edited, or translated text can slip past the detector, and so can output from rival AI tools.
The absence of a detected watermark does not prove human authorship.
Source: OpenAI
Images and audio work differently. Anyone can upload those files to OpenAI’s public verify tool and check for its SynthID watermark.
Outside Europe, the ChatGPT watermark stays off by default. However, API customers worldwide can now opt in for select models through their project or organization settings.
The rollout follows the EU’s August start of AI Act transparency enforcement. For crypto, the stakes look concrete, as criminal AI adoption climbed 40% year-on-year, according to TRM Labs.
Therefore, the gap between what regulators demand and what detectors can prove will likely shape how every major AI lab handles compliance.
