OpenAI's text generators create hidden watermarks unintentionally

Hidden watermarks have been discovered in text generated by OpenAI's GPT-o3 and GPT-o4 mini models, raising concerns about AI-generated content. These invisible characters are detectable in specialized editors. The "watermarks," specifically non-breaking spaces, appear in longer texts and could potentially be used to identify AI-generated content. OpenAI claims the symbols are a byproduct of their reinforcement learning process, not an attempt to deceive. This discovery comes as AI detection tools are becoming more prevalent, particularly in academic settings. The presence of these watermarks could provide a more reliable method for identifying AI-generated text.