OpenAI is introducing an invisible, machine-readable watermark in the text generated by ChatGPT and Codex, with the rollout initially limited to users within the European Union.
The new feature is designed to embed a signal into written output that software can detect, helping to identify content produced by artificial intelligence. At launch, the capability is being offered only to those in the EU, reflecting the growing regulatory focus on labelling machine-generated material across the bloc.
How textGrain Compares to Rival Systems
According to OpenAI, its textGrain watermarking “matched or exceeded” alternative approaches, including Google DeepMind’s SynthID for text. That same SynthID technology also underpins the watermarking that Anthropic announced in August. Both OpenAI and Anthropic have moved to introduce these measures in order to meet the requirements set out in the EU’s AI Act.
The arrival of such tools has not been universally welcomed, however, with concerns raised following the announcement of the additions. OpenAI has nonetheless pointed to results from AI benchmarks indicating that watermarked and unwatermarked text deliver similar performance, suggesting the feature does not meaningfully degrade output quality.
Limits of the Watermarking Approach
Despite the positive comparisons, OpenAI has been clear that the technology is not infallible. The company notes that textGrain “does not guarantee” the detection of AI-generated text in every case, acknowledging that the system has its boundaries.
The rollout forms part of a wider industry trend in which major developers of generative systems are adopting watermarking in response to legislative pressure. The EU’s AI Act has become a central driver of these changes, prompting companies to build in mechanisms that allow AI-produced content to be flagged and traced.
For now, the watermarking feature in ChatGPT and Codex remains an EU-only measure, with no broader availability confirmed beyond that region.
Source
Image: theverge.com
