-
OpenAI has developed a tool to detect ChatGPT-written text with 99.9% accuracy, but hasn’t released it due to internal debates and concerns.
-
The tool uses a watermarking method that subtly changes token selection in text generation, creating an invisible pattern detectable by OpenAI’s technology.
-
OpenAI is cautious about the tool’s impact on user experience and fairness, and is exploring less controversial alternatives.
OpenAI has developed a tool to detect if ChatGPT is used to write essays or research papers, but the company is hesitant to release it due to internal debates and concerns about user reactions. The tool uses a watermarking method that subtly changes token selection in text generation, creating an invisible pattern detectable by OpenAI’s technology.
The watermarking method is reportedly 99.9% effective when enough new text is generated. However, there are concerns that the watermark could be easily removed through techniques like translation or adding and removing emojis. OpenAI is also weighing the risks and benefits of releasing the tool, including potential false accusations and impact on ChatGPT’s performance.
Teachers and professors are eager for tools to combat AI misuse in education, with 59% of middle- and high-school teachers believing students have used AI for schoolwork. OpenAI is cautious about the anti-cheating tool’s impact on user experience and fairness, and is exploring less controversial alternatives.
