.png)
April 22, 2025
OpenAI, known for its AI chatbot ChatGPT, is experimenting with implementing watermarks in AI-generated texts. After images created with the AI model GPT-4 were already provided with partially invisible watermarks, the company now seems to be extending this approach to texts as well. This apparently affects texts generated with the newer models GPT-3 and GPT-4-mini. However, an official announcement from OpenAI is still pending.
The watermarks consist of special Unicode characters, specifically the "Narrow No-Break Space" (U+202F). This character is indistinguishable from a regular space to the human eye and is transferred when the text is copied to other applications. By analyzing the distribution of these characters within a text, it could be determined whether it was generated by AI.
However, the method is not foolproof. Using common word processing programs that display Unicode characters, the markers can be easily identified and replaced with normal spaces. Thus, the effectiveness of the watermarks is limited and depends on how well known the method is.
Experts speculate that OpenAI might stop the test run as soon as the method becomes too widespread. On the other hand, OpenAI has already acknowledged similar challenges in watermarking images, such as the possibility of removal by cropping or re-uploading. Despite these limitations, the company emphasizes the value of such measures as a step towards responsible use of AI. The watermarks, according to OpenAI, are not perfect, but could still contribute to the transparency and traceability of AI-generated content.
The topic of AI-generated content and its labeling is becoming increasingly important. Mindverse, the German provider for AI-supported content creation, offers comprehensive solutions for text, images, research and more. As your AI partner, Mindverse also develops customized solutions such as chatbots, voicebots, AI search engines and knowledge systems that are optimally tailored to your needs. With Mindverse, you benefit from the latest AI technology and can optimally utilize the potential of Artificial Intelligence for your company.
The development of robust and at the same time practical methods for labeling AI-generated content presents a challenge. The easy manipulability of the watermarks used so far highlights the need for further research and development in this area. Future solutions must consider both the technical aspects of detection and removal as well as the ethical implications of AI content generation. An open dialogue between developers, users and the public is essential to ensure responsible handling of AI-generated content.
Bibliography: - https://t3n.de/news/openai-zeichen-chatgpt-texte-1683993/ - https://x.com/t3n/status/1914360576509108554 - https://www.threads.net/@t3n_magazin/post/DIt1qB9M-6g/openai-versteckt-offenbar-wasserzeichen-in-per-chatgpt-erstellten-texten-warum-d - https://t3n.de/tag/chat-gpt/ - https://www.facebook.com/100064654845221/posts/1105348191630332/ - https://t3n.de/tag/open-ai/ - https://t3n.de/ - https://t3n.de/tag/software-entwicklung/