OpenAI has developed a powerful tool capable of detecting GPT-4’s writing output with a reported accuracy of 99.99%, causing concern among ChatGPT plagiarists. The company, however, has been indecisive about releasing it to the public for over a year.
According to TechCrunch, OpenAI is adopting a cautious approach due to the potential impact on the broader ecosystem and complexities involved. An OpenAI spokesperson mentioned that while the text-watermarking tool shows promise, there are risks to consider such as circumvention by bad actors and impact on non-English speakers.
The text-watermarking system embeds a detectable pattern in the written output, invisible to the user but recognizable to the OpenAI tool. Although effective in spotting GPT-4 generated content, it cannot detect outputs from models like Gemini or Claude. Interestingly, the watermark can be removed by running the text through translation tools.
OpenAI’s previous attempt at developing a text-detection tool had a low accuracy rate, leading to its discontinuation. This tool incorrectly detected human-generated content as AI, causing issues like a professor failing a class based on incorrect accusations.
While other AI startups rush to release text detectors, concerns about their accuracy persist. The human eye remains the most reliable method for spotting AI-generated content, stating the need for caution when relying solely on AI tools.
