OpenAI is rolling out an invisible, machine-readable watermark called textGrain to ChatGPT and Codex outputs. The feature launches first in the European Union.
The watermark embeds information into text that AI systems generate, making it possible to detect machine-authored content programmatically. OpenAI claims textGrain "matched or exceeded" competing approaches, specifically naming Google DeepMind's SynthID for text and a watermarking method Anthropic announced in August as comparison points.
According to The Verge, Anthropic's watermarking is also based on the same underlying approach as SynthID.
The rollout begins with EU users, suggesting compliance with European regulations may drive adoption. The EU has pursued strict rules around AI transparency and content provenance. Watermarking AI-generated text addresses one element of that regulatory framework: the ability for downstream users, platforms and regulators to identify when content originated from a generative system rather than a human author.
Text watermarking operates differently from image watermarking. Rather than embedding visible or semi-visible markers, textGrain inserts patterns into word choice and phrasing that remain imperceptible to readers but carry machine-detectable information. This preserves readability while enabling verification. The approach protects against one category of AI misuse: passing off synthetic text as human-authored without disclosure.
The expansion beyond the EU timing and scope remain unspecified in the available details.