OpenAI to Add Invisible Watermarks to ChatGPT Text in the EU

Introduction

OpenAI says it will begin adding an invisible watermark to text generated by ChatGPT and Codex in the European Union. The move is intended to satisfy the transparency provisions of the EU AI Act, which took effect on August 2 and requires AI-generated content to carry a mark that other systems can identify. The rollout is expected over the coming weeks for eligible ChatGPT and Codex users in the EU, across all plans.

Key points

• The mark is embedded in word choices. It is not a visible symbol. Instead, the model makes small adjustments to its next-word selections, creating a statistical pattern that readers cannot see but a detector can analyze. Because the pattern is part of the text, it can remain after copying and pasting.
• The initial default is regional. OpenAI does not plan to make text watermarking a global default at launch. Developers using the OpenAI API anywhere in the world can enable it for selected models, although the setting is off by default.
• Detection has clear limitations. In one OpenAI test, replacing about 10% of the words with synonyms reduced detection from roughly 92% to 66%. Short passages, mathematical answers, and translated text are also more difficult to assess. A missing watermark does not prove human authorship: the passage may be too short, heavily edited, or generated by another AI system.
• It is not an identity tag. OpenAI says the watermark does not identify the user and that enabling it caused no meaningful change in model performance. The company also released a technical report on textGrain, a method that uses a secret key to organize next-token predictions and accumulates many small nudges into a detectable signal.

Why it matters

The rollout shifts AI transparency toward a machine-checkable provenance signal. That could be useful for platforms, educators, publishers, and regulators that need an additional clue about how text was produced. It does not, however, establish how much of a passage came from a model or how much human judgment, editing, and creativity shaped the final result. Since modest rewriting can weaken the signal, the watermark is better understood as evidence of possible OpenAI involvement than as conclusive proof for enforcement or authorship decisions.

OpenAI had reportedly developed a text watermark earlier but delayed release partly because users might move to competitors without similar controls. Anthropic later announced worldwide watermarking for Claude text, prompting debate among some users about whether AI-assisted work should be labeled when people provide the instructions, context, and decisions. With OpenAI, Anthropic, Google, Meta, and Microsoft among the companies committed to the EU’s code of practice on AI-generated content, watermarking may become part of compliance competition. Reliability, detector access, false positives, and cross-provider compatibility will remain central questions.

Source: TechCrunch AI