October 7, 2026: OpenAI has begun introducing invisible text watermarking for ChatGPT and Codex outputs in the European Union, allowing automated systems to detect signals associated with AI-generated text.
The technology, called textGrain, subtly influences how the model chooses between possible words or word pieces while generating a response. These choices create a statistical pattern that is not visible to readers but can be detected by specialised systems.
OpenAI says the watermark does not add hidden characters, invisible spaces or unusual punctuation. Instead, the signal is embedded in the wording itself.
The company introduced the system as part of its compliance with transparency requirements under the European Union’s AI Act and commitments made under the EU Code of Practice on Transparency of AI-Generated Content.
For now, the watermarking rollout is focused on eligible ChatGPT and Codex text generated in the EU. OpenAI API customers worldwide can also choose to enable text watermarking, but it is not the global default.
OpenAI has also cautioned that text watermarking has limitations. Detection can become less reliable when passages are very short or when the original text has been substantially edited. A watermark can indicate that an OpenAI system may have generated or processed part of the text, but it cannot identify the user, prove authorship or determine whether the content is accurate.
Anthropic has announced a similar watermarking system for future Claude models, also citing EU AI Act requirements. Its approach likewise relies on subtle statistical patterns rather than visible markers.
The move reflects a broader push by major AI companies to make machine-generated content easier to identify as regulators increase transparency requirements around generative AI.