OpenAI Enables ChatGPT Watermarking by Default in the EU
OpenAI has announced that it will automatically watermark text generated by ChatGPT for users in the European Union, aligning its platform with regional regulatory requirements.
Radwan AbdulhadiEditor and technology writer
European Union Rollout and Regulatory Compliance
OpenAI has announced that it will begin automatically watermarking text generated with ChatGPT within the European Union. According to detailed coverage by Ars Technica, the feature will also be offered in other regions, but it will remain turned off by default outside of the EU, except in the ChatGPT and Codex apps where it is mandated by local rules.
The European rollout is directly driven by the need to comply with the EU AI Act, which officially took effect in August. The legislation requires artificial intelligence models to mark content they produce in a way that allows other tools to detect its origin.
How OpenAI's textGrain Watermarking Works
OpenAI's proprietary watermarking method is called textGrain. The company has published a technical paper explaining the mechanisms behind it, which can be reviewed by studying the textgrain entropy-calibrated watermarking for language model text documentation.
In general, the system operates similarly to other large language model watermarking tools introduced previously. It embeds subtle patterns within word choices that remain imperceptible to a human reader and do not meaningfully alter the overall quality of the generated output. However, individuals equipped with a key and a specialized detector can identify these patterns.
OpenAI stated that access to the detector will initially be granted to a limited number of researchers and organizations. Other parties will need to submit applications through a request-for-approval process to gain access over time.
Reliability and Limitations of Detection
Despite its technological design, experts and testing data show that textGrain is not entirely reliable and remains easy to circumvent. Similar solutions like SynthID and the C2PA project also face challenges, as they can be easily bypassed by anyone possessing basic technical know-how.
OpenAI's own tests of textGrain demonstrate a respectable 92 percent successful detection rate under ideal circumstances. However, those tests also reveal that modifying just 10 percent of the text in an output reduces the successful detection rate by nearly 30 percent. Furthermore, altering 20 percent of the text can lower the success rate by roughly 75 percent.
The company also noted that detection success rates tend to be lower for shorter or translated text passages compared to longer, original outputs. These performance metrics align closely with what the industry has observed in other similar watermarking solutions.
Comparison with Competitors and Deployment Timeline
The implementation by OpenAI follows similar moves by industry competitors. In August, rival AI firm Anthropic introduced watermarking for text generated by its own models. Unlike OpenAI, however, Anthropic chose to enable its watermarking globally rather than limiting default activation strictly to regions where regulators mandate it.
While OpenAI is making watermarking active by default within the EU for applications like ChatGPT and Codex, the feature will remain turned off but optionally available within the company's API. OpenAI reports that the new watermarking functionality will roll out to users in the European Union in the coming weeks.
Sources
- Ars TechnicaOpenAI will watermark ChatGPT outputs by default—but only in the EU
Continue chronologically
Related entity coverage
Related stories
OpenAI to Watermark ChatGPT Text in the European Union
OpenAI is rolling out invisible watermarks for ChatGPT and Codex text outputs in the European Union to comply with the new AI Act.
OpenAI Introduces Text Watermarking in ChatGPT and Codex
OpenAI is rolling out invisible machine-readable watermarks for ChatGPT and Codex in the EU to comply with transparency rules.
OpenAI Details Text Watermarking Approach for EU AI Act
OpenAI announces its phased text watermarking approach under the EU AI Act, featuring textGrain technology, API opt-ins, and researcher access.
OpenAI Adding Invisible Watermarks to EU Text and Code
OpenAI is rolling out invisible text watermarking for ChatGPT and Codex in the European Union to comply with regional AI transparency rules.
