OpenAI said, while acknowledging that text watermarking and detection remain early technologies with significant limitations, that its phased approach reflects both the EU AI Act requirements and the technologies’ limitations.  The company also said that it was opening applications to access its text watermark detector. However, this access will initially be limited to approved researchers and expert organisations that assist the company with evaluation and improvement of the technology.

Story continues below.
Subscribe to see fewer ads.

How does the watermark work?

In its blog, OpenAI said that its technology named textGrain adds something it describes as an ‘invisible statistical signal’ to the model’s word choices. Later, the detector scans for that signal to determine whether a passage contains an OpenAI watermark. OpenAI has published a paper, ‘textGrain: Entropy-Calibrated Watermarking for Language Model Text’, shedding light on the inner workings of the technology.

Further, OpenAI said that in its evaluations, textGrain matched or exceeded the performance of other approaches that it tested, including SythID (Google DeepMind’s digital watermarking tech) for text. OpenAI also revealed that, along with strong performance, under ideal conditions, this technology could not guarantee reliable detection in everyday use. The company said that the detectors can make two kinds of errors – they can detect a watermark where none is present or miss a watermark that is present. Some of the key hurdles to watermark detection are that shorter or more constrained text is harder to detect, while editing text generated by AI can weaken its watermark.

The limitations with watermarks

OpenAI also pointed out factors that may deem watermarks inconclusive.  It said that a watermark may not necessarily mean information about the extent of human judgement, editing or creativity. It does not establish any form of ownership or responsibility; most importantly, it does not identify the user. It is also not a barometer to verify accuracy. The company claimed that the absence of a detected watermark does not translate to human authorship.

Article 50 of the AI Act, which applies from August 2, 2026, sets out transparency obligations for providers and deployers of certain AI systems, including generative and interactive AI systems and deepfakes. While providers are required to design AI systems that ensure individuals are explicitly informed whenever they interact with an AI system directly, they must add machine-readable marks to enable the detection of AI-generated or manipulated content. On the other hand, deployers of AI systems must inform users when they are exposed to emotion recognition and biometric categorisation tools, deepfakes, and text publications on matters of public interest without human review or editorial control. According to the Act, these obligations are aimed at building trust and integrity in the information ecosystem.

Story continues below this ad

OpenAI’s latest announcement follows a similar measure taken by Anthropic in early August when it declared that it is adding invisible watermarks to text and files generated by its Claude models. Anthropic also made the announcement in its bid to comply with the EU’s AI Act on Transparency Code. The company said that watermarks will be applied globally to all new models released after August 2, with plans to roll out support for older legacy models.