Could an Accurate ChatGPT Watermarking Tool Exist? OpenAI's Decision to Withhold It Explained

ChatGPT plagiarists, take note: OpenAI has created a tool that can detect GPT-4 writing output with an impressive 99.99% accuracy. However, the company has been deliberating for over a year on whether to make it publicly available. According to TechCrunch, OpenAI is taking a “deliberate approach” due to the complexities and potential consequences this tool could have on the broader ecosystem. An OpenAI spokesperson highlighted that “the text watermarking method we’re developing shows technical promise, yet we are weighing significant risks. These include the potential for circumvention by malicious actors and the disproportionate impact on groups such as non-English speakers.”

The text-watermarking system embeds a specific pattern into the model’s outputs, making it detectable by the OpenAI tool but invisible to users. While it can effectively identify text generated by its GPT-4 engine, it cannot differentiate outputs from other AI models like Gemini or Claude. Furthermore, the watermark can easily be erased by translating the text into another language and then back to the original.

This isn’t OpenAI's first venture into developing a text detection tool. Last year, it quietly discontinued a similar project due to its unsatisfactory detection rate and tendency for false positives. Released in January 2023, that earlier tool required users to input at least 1,000 characters of sample text for analysis. With only a 26% accuracy in correctly identifying AI-generated content, it mistakenly labeled 9% of human-written content as AI-generated. This even resulted in a Texas A&M professor incorrectly failing an entire class over alleged use of ChatGPT on final assignments. Additionally, OpenAI is wary of backlash from users; according to the Wall Street Journal, 69% of ChatGPT users consider such a tool unreliable, fearing it may lead to false accusations of cheating. Another 30% indicated they would switch to alternative models if OpenAI released the feature. Concerns also linger that developers might reverse-engineer the watermark to create counter-tools. As OpenAI continues to weigh the pros and cons of launching its watermarking system, other AI startups are racing to introduce text detection tools, including GPTZero, ZeroGPT, Scribbr, and Writer AI Content Detector. However, the overall lack of accuracy in these tools means that the human eye remains our most dependable method for spotting AI-generated content, which is not a comforting thought.