OpenAI Develops Secret Watermark to Identify AI-Generated Text
OpenAI is building a statistical watermarking tool to secretly mark text generated by systems like ChatGPT. This hidden signal aims to prevent academic plagiarism, mass propaganda, and impersonation by making it easier to distinguish human writing from AI output.
OpenAI is developing a statistical watermarking tool to secretly identify text generated by AI systems like ChatGPT. Currently serving as a guest researcher at OpenAI, computer science professor Scott Aaronson reveals that this technology embeds an unnoticeable signal into the generated text to indicate its artificial origin. OpenAI engineer Hendrik Kirchner already builds a working prototype, and the company hopes to integrate this feature into future systems to prevent AI outputs from being passed off as human writing.
The need for this watermarking stems from the rapid rise and ethical concerns surrounding ChatGPT, a highly capable chatbot that excels at answering questions, writing poetry, and solving programming puzzles. However, bad actors easily exploit this powerful technology to write malicious phishing emails, generate harmful malware, or help students cheat on assignments. Furthermore, the system struggles with factual consistency, a major shortcoming that prompts platforms like Stack Overflow to ban AI-generated answers entirely.
This watermarking approach exploits the fundamental mechanics of how language models actually work by manipulating the mathematical probability distributions they use to select the next token. By subtly altering the random selection process of words and punctuation, the tool creates a hidden pattern that humans cannot detect but algorithms can easily verify. Ultimately, OpenAI hopes this method stops malicious applications like mass propaganda campaigns and malicious impersonation without degrading the quality of the AI's output.