As OpenAI continues to push the boundaries of artificial intelligence, it’s introducing a new feature that aims to combat AI-generated text scams in the European Union. Starting this month, ChatGPT and Codex will invisibly watermark text generated by these models, making it easier to detect when users copy or share AI-created content without permission.
The invisible watermarks won’t be visible to the naked eye, but they’ll leave a statistical pattern that can later be detected using OpenAI’s proprietary technology. This means that if you’re copying or sharing text from ChatGPT or Codex, it will be possible for the owner of the original text to identify whether it was generated by AI or written by a human.
OpenAI is introducing this feature in response to growing concerns about AI-generated content being misused online. The company says its watermarking technology won’t be enabled globally at first but can be opted into by developers using ChatGPT and Codex APIs worldwide. However, OpenAI warns that the watermarks can disappear if the text is edited or modified in any significant way.
While this new feature may seem like a straightforward solution to combating AI-generated text scams, it’s not without its limitations. According to OpenAI’s own tests, editing just 10% of words with synonyms reduced the watermark’s detectability from 92% to 66%. This raises questions about how effective these watermarks will be in real-world scenarios where texts are frequently edited or translated.
The lack of reliability is further complicated by the fact that a detected watermark doesn’t reveal who generated the text, their account information, prompt, or conversation history. It also can’t determine how much of the final work was written or edited by a human. This means that while the watermarks may provide some level of assurance, they won’t be able to pinpoint the source of AI-generated content with certainty.
Despite these limitations, OpenAI claims that watermarking doesn’t significantly affect the quality of GPT-6 Astra’s performance. With benchmark results remaining broadly similar when textGrain is enabled, it seems that this feature may not compromise the AI’s capabilities in any meaningful way.
As we navigate the increasingly complex landscape of AI-generated content, it’s essential to be aware of these new developments and their implications for online security. For users, the takeaway from OpenAI’s watermarking initiative should be to exercise caution when sharing or copying text generated by AI models. Be aware that even with watermarks in place, there are still risks associated with relying on AI-generated content.
To stay ahead of emerging threats, consider implementing robust authentication and verification processes for your organization’s online communications. This may involve investing in AI-powered detection tools that can identify potential security risks related to AI-generated content. By being proactive and informed about these developments, you’ll be better equipped to protect yourself and your business from the evolving landscape of AI-powered attacks.
Source: Bleeping Computer — 2026-10-05