MikhbarMIKHBAR
Artificial Intelligence

OpenAI Adds Invisible Watermark to EU Text and Code

OpenAI announced plans to incorporate an invisible digital watermark into text and code generated within the European Union, following a similar move by Anthropic earlier this year.

OpenAI Adds Invisible Watermark to EU Text and Code

Compliance With European Union Transparency Rules

OpenAI recently announced its plans to comply with part of the European Union's new AI transparency rules. The artificial intelligence developer will soon begin adding an invisible watermark to text and code generated in the EU, following a similar move by Anthropic earlier this year. This implementation is designed to meet strict regulatory frameworks established across the region.

Article 50 of the EU's AI Act mandates that providers of generative AI systems make text outputs identifiable in a machine-readable way. While this regulation already pertains to new companies entering the market, pre-existing entities have until December 2 to comply. This mandated timeline impacts major organizations including OpenAI, Anthropic, Microsoft, Google, and Meta, among others.

The Technology Behind textGrain

To achieve this watermarking capability, OpenAI is utilizing a proprietary system called textGrain. According to the company, the software adds an invisible statistical signal to the model's word choices. Once this signal is embedded, a specialized detector can spot the modifications and ascertain whether a watermark is present in the generated output.

OpenAI reports that its system has matched or exceeded the performance of other approaches, such as SynthID for text. Despite these encouraging performance benchmarks, the company offers important cautions regarding everyday reliability.

An OpenAI logo on a smartphone.
Image related to the report from Engadget · Source: Engadget

Performance Limitations and Edge Cases

Although the underlying technology shows strong initial performance, OpenAI cautions that robust results under ideal conditions do not automatically guarantee reliable detection in everyday use. Shorter text outputs present distinct challenges, yielding roughly an 80 percent success rate during testing. Content categorized as mathematics faces similar identification hurdles.

Furthermore, basic editing of the generated text tends to bring down the overall success rate of the detection mechanism. To foster broader improvement and transparency, OpenAI plans to make the underlying technology available in open source so that other developers and researchers can study and tweak it.

Rollout and Access Controls for Detection

Regarding implementation details, the watermark will not be enabled by default across all platforms. Instead, it will apply directly to eligible ChatGPT and Codex text outputs located within the European Union. Customers will retain the ability to opt-in for select models, though the company has not yet specified which models will eventually become available to the general global public.

OpenAI is adopting a cautious approach toward the distribution of its detection software. Access to the detector will initially be limited to approved researchers and expert organizations. Additionally, the developer notes that the detection tool does not identify individual users or reveal sensitive underlying prompts and conversations.

Sources

  • EngadgetOpenAI will add a digital watermark to text and code generated in the EU

Continue chronologically

Related entity coverage