OpenAI and the Invisible Watermark: Transparency in the Age of AI
OpenAI has announced that it will begin adding an invisible watermark to texts generated by ChatGPT and Codex in the European Union. This measure aims to comply with the transparency rules of the EU AI Act, which came into effect on August 2. The legislation requires artificial intelligence companies to identify AI-generated content in a way that other systems can recognize.
The watermark, which is not a visible symbol, works by subtly shaping the model's word choices, creating a pattern that ordinary readers do not perceive, but that a detector can identify. This technique, called textGrain, was detailed in a technical report published by OpenAI in collaboration with researchers from the University of Pennsylvania and Yale. The method uses a secret key to order word predictions, allowing the detector to identify AI-generated content using only the text and the key.
How the Watermark Works
OpenAI's watermark does not alter the performance of the models and does not identify the user. It is incorporated directly into the words, meaning it remains in the text even when copied and pasted. However, the effectiveness of detection may be reduced. OpenAI's tests show that replacing 10% of the words with synonyms can decrease detection from 92% to 66%. Additionally, short passages, mathematical responses, and translated texts are more difficult to detect.
OpenAI has decided to provide initial access to the detector only to approved researchers and specialized organizations, which can help assess the reliability and responsible use of the technology. It is important to note that the absence of a watermark does not prove that the text was written by a human. The text may be very short, have been extensively edited, or have been generated by another AI.
Implications and Reactions
OpenAI's decision to implement the watermark comes two months after Anthropic announced a similar measure for Claude, which will be applied globally. This decision has generated criticism from some Claude users, who argued that they provided the instructions and context, while Claude was merely a tool.
OpenAI had previously developed a watermark for texts but hesitated to launch it due to concerns that users would migrate to competitors that did not use the marking. Now, with the implementation of EU rules, the company joins other tech giants like Google, Meta, and Microsoft, which have committed to following the EU code of practice for AI-generated content.
The introduction of the invisible watermark by OpenAI is a significant step towards transparency in the age of artificial intelligence. While there are still challenges to overcome, such as the ease of removing the watermark through edits, the initiative represents an effort to ensure that AI-generated content is clearly identified, promoting greater trust and accountability in the use of these technologies.
Comments (0)
Comments are moderated and if they violate our Terms and Conditions of use, the comment will be deleted. Persistence in violation will result in a ban of your account.