OpenAI plans to introduce invisible watermarks in eligible ChatGPT and Codex text outputs across the European Union. The ChatGPT text watermarking rollout will use a technology called textGrain, which creates detectable statistical patterns through subtle changes in word selection.
According to the company’s announcement, the rollout will take place over the coming weeks. Readers will not see a label or symbol within the generated text.
However, the technology has limitations. Editing, translation and short passages can make the watermark harder to detect.
Watermarks Rely on Word Choices
Rather than inserting a visible marker, textGrain slightly adjusts the words a model chooses while generating a response. Those choices create a statistical pattern that a detector can subsequently examine.
The watermark therefore remains invisible when someone reads or copies the text. Its purpose is to help identify eligible watermarked output, rather than change how that output looks.
For now, OpenAI is limiting the planned ChatGPT and Codex rollout to the EU. It has not announced watermarking as a worldwide default.
Meanwhile, API developers worldwide can opt in for supported models. Watermarking remains disabled by default for those developers.
Detector Access Will Initially Be Restricted
OpenAI is accepting applications for access to its watermark detector. However, the initial program will serve approved researchers and expert organizations.
According to the company, detecting a watermark does not reveal the person who generated the text. It also does not expose their account, prompt or conversation.
Furthermore, a positive result cannot establish how much of a finished piece a human wrote or edited. The detector identifies a signal in the text, rather than reconstructing its complete editing history.
Ordinary Editing Can Weaken Detection
OpenAI’s evaluations show that relatively modest wording changes can substantially reduce detection rates.
In a test involving 400-token passages, replacing 10% of words with synonyms lowered detection from approximately 92% to 66%. When replacements reached 25%, detection fell to 17%.
These results illustrate an important limitation of ChatGPT text watermarking: the statistical signal can weaken as the wording changes.
Translation can also interfere with detection. Consequently, the absence of a watermark does not establish that a human wrote the text.
Passage Length and Subject Matter Affect Results
The amount of available text also influences detection performance.
At a target false-positive rate of 1%, OpenAI detected watermarks in approximately 80% of 200-token responses to psychology questions. For 400-token responses, that figure rose to roughly 95%.
However, those results reflect specific evaluation conditions. They do not establish a universal detection rate across every subject or type of writing.
Mathematics responses produced weaker results because the model had less flexibility in its word choices. That constraint leaves fewer opportunities to create a detectable pattern.
Short passages present a similar challenge by providing less material for analysis.
OpenAI Reports Little Change in Model Quality
OpenAI says enabling textGrain does not meaningfully affect GPT-6 Astra’s quality. According to the company, benchmark results remained broadly similar with watermarking switched on.
Nevertheless, output quality and detection reliability are separate questions. Text can remain readable while its watermark becomes difficult to identify after revision.
The planned rollout therefore adds a way to recognize some AI-generated output, but it does not provide a definitive test of human or AI authorship.
0 responses to “ChatGPT Text Watermarking Set for EU Rollout Alongside Codex”