OpenAI has developed a tool that could identify students using ChatGPT to complete their assignments, but they’re unsure whether to release it. The Wall Street Journal reported that the company is experimenting with a method called text watermarking. This technique involves making tiny changes to how ChatGPT picks words, creating a hidden “watermark” in the text that can be detected later.
An OpenAI spokesperson told TechCrunch that the company is being careful about releasing this tool due to its complexities and potential impact. They’re worried about issues like people finding ways to bypass the watermark and the possibility of unfairly affecting non-English speakers.
It’s “highly accurate”
This approach is different from earlier attempts to spot AI-generated text, which have mostly failed. For example, GPTZero, another tool designed to detect AI-written text, faced challenges in accuracy and effectiveness. Even OpenAI stopped using its own AI text detector last year because it wasn’t accurate enough. The new watermarking tool would focus only on catching text written by ChatGPT, not other AI models.
In an updated blog post, OpenAI noted that text watermarking has been “highly accurate” against simple tampering, like changing a few words. However, it’s less effective against more complex methods, like translating the text into another language or using another AI model to reword it. The company also mentioned concerns that this tool could unfairly stigmatize AI use, especially for non-native English speakers who might use AI for help with writing.
As OpenAI considers these factors, they’re taking time to decide whether or not to release this tool, weighing the pros and cons carefully.
All images are generated by Eray Eliaçık/Bing