OpenAI is rolling out invisible AI text watermarking for ChatGPT in the European Union, introducing a new way to identify whether text was generated or processed by its models.
The move comes as the EU AI Act’s transparency requirements take effect, putting greater emphasis on machine-readable signals for AI-generated content. The European Commission says the relevant transparency obligations began applying on August 2, 2026.
OpenAI says eligible ChatGPT and Codex text outputs in the EU will receive an invisible watermark over the coming weeks. API customers worldwide can also choose to enable watermarking for selected models, although it remains off by default.
How OpenAI’s AI text watermark works
The technology, called textGrain, does not add visible symbols, hidden characters or strange formatting.
Instead, it subtly changes the statistical pattern of the model’s word choices. A specialized detector can then look for that pattern and estimate whether an OpenAI watermark is present. OpenAI says it plans to eventually release the technology as open source.
The technology is designed to work differently from conventional AI detectors, which attempt to identify AI writing by analyzing linguistic patterns after the text has already been produced.
But AI watermarking has a major weakness
A watermark is not an AI lie detector.
OpenAI’s own testing shows that detection becomes harder with short or highly constrained text. In one evaluation, detection for 200-token psychology passages was about 80%, compared with roughly 95% for 400-token passages at a 1% false-positive rate.
Editing can weaken the signal even more. Replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. Mathematics was also harder to detect because there is less freedom in word choice.
That is why OpenAI initially plans to restrict access to its text detector to approved researchers and expert organizations.
What the watermark cannot prove
Finding an OpenAI watermark does not reveal who used ChatGPT, how much AI contributed, who owns the text or whether the information is accurate.
Conversely, failing to detect a watermark does not prove that a human wrote the content. Rewriting, translation, short passages and unsupported models can all make detection unreliable.
The bigger AI transparency battle
OpenAI is pursuing a broader provenance strategy that combines watermarking, C2PA Content Credentials and verification tools for different types of AI-generated media.
The challenge is increasingly clear: as AI-generated text becomes indistinguishable from ordinary writing, proving where content came from may become as important as generating it.
But OpenAI’s own research highlights the central problem—an AI watermark can provide a signal, not absolute proof of authorship.
