OpenAI will begin automatically watermarking text generated by ChatGPT in the European Union to comply with the EU AI Act. The feature will be optional in other regions. The company's proprietary watermarking method, textGrain, has shown a 92% detection rate in tests, but its effectiveness is significantly reduced by even minor text alterations.
The implementation of AI watermarking, particularly in the EU due to regulatory requirements, raises questions about the reliability and effectiveness of detecting AI-generated content, impacting trust and authenticity in digital communications.
OpenAI will begin automatically watermarking text generated with ChatGPT in the European Union, the company announced. This move is primarily driven by the need to comply with the EU AI Act, which took effect in August and mandates that content produced by AI models be marked in a way that another tool can detect.
While OpenAI is making this feature default in the EU, it will be available but off by default in other regions. The company's proprietary watermarking method, called textGrain, works by embedding imperceptible patterns in word choices that can be detected by a specialized tool. OpenAI plans to grant access to the detector to a limited number of researchers and organizations, with a process for others to be added over time.
However, the effectiveness of textGrain, like other similar solutions, is not absolute. OpenAI's own tests indicate a respectable but not entirely reliable 92% successful detection rate. The company's tests also revealed that altering just 10% of the text in an output can reduce the detection rate by nearly 30%, and changing 20% of the text can lower it by approximately 75%. These figures are comparable to other LLM watermarking tools, which are generally easier to circumvent with basic technical knowledge. The success rate is also noted to be lower for shorter or translated text.
In August, competitor Anthropic also introduced watermarking for its models, but it has enabled the feature globally, unlike OpenAI's current approach of defaulting it only where regulators require it. The new watermarking is expected to roll out to EU users in the coming weeks.
Pick the topics you care about. Get only what matters, on your cadence.
