OpenAI Follows Anthropic: ChatGPT Text Watermark in the European Union, Detection May Weaken

OpenAI has officially applied an invisible watermark to text generated by ChatGPT and Codex through a tool called textGrain. The move follows Anthropic's earlier decision to embed a hidden watermark in Claude's output as a form of compliance with European Union regulations. Although the goal is similar, the two companies' technical approaches differ greatly.

According to OpenAI's explanation, textGrain inserts a special code into the text so that AI detectors can identify whether content was produced or edited by their system. The tool prepares several versions of word probabilities that are slightly altered, keeping their average close to the original so quality does not decline, then uses a key to select which version is used at each step. Meanwhile, Claude's method works like a keyed lottery for each word: every candidate word is given a secret score based on the key and the preceding text, then the highest-scoring word is chosen so that a pattern emerges over long passages.

The effectiveness of textGrain watermark detection depends heavily on context. Reports cite a success rate of about 80% for 200-word texts and 95% for 400-word texts in the psychology domain, but "far lower" in mathematics. When 10% of words are replaced with synonyms in a 400-token passage, detection drops from about 92% to 66%; replacing 25% of words suppresses that figure to around 17%.

OpenAI says it will open-source textGrain as standards mature. Currently, API customers worldwide can opt to enable text watermarking on certain models, but the feature is off by default, except in the European Union. That policy raises questions about uniformity of implementation across jurisdictions.

"We want transparency, but we also want to maintain output quality. Watermarking is one way to balance the two," an OpenAI spokesperson said in a statement cited from a source.

For Indonesia, this development is worth watching closely. Data protection and digital content regulations in the country do not yet specifically govern AI text watermarking. However, with the growing adoption of generative AI in education, media, and public services, the need for content verification mechanisms is becoming more urgent. If the watermark is active only in Europe, users in Indonesia may receive text without markers, which could complicate efforts to enforce academic or journalistic integrity.

On the other hand, the limitations of detection—especially when words are replaced with synonyms—show that watermarking is not a single solution. Cybercriminals or disinformation spreaders can easily modify text to evade detection. Therefore, a layered approach involving source verification, digital literacy, and platform oversight remains necessary.

Going forward, the question is whether watermark standards will be adopted globally or remain fragmented along regional regulations. If OpenAI and Anthropic compete fiercely on compliance, there may be pressure on other AI providers to follow a similar path. However, without international agreement, the effectiveness of watermarking as a verification tool will continue to be tested by a variety of text manipulation techniques.