OpenAI has developed a system for watermarking text generated by ChatGPT and a tool to detect the watermark. The company is considering releasing it, but some employees are concerned that it could hurt their bottom line. The watermarking adjusts how the model predicts words and phrases, creating a detectable pattern. A survey found that most people support having an AI detection tool.
OpenAI confirmed its work on watermarking in a blog post update, saying that its method is very accurate (99.9% effective) and resistant to tampering. However, it’s concerned about the potential stigmatization of AI tools for non-native speakers. Some employees feel that watermarking is effective, but others are worried about user sentiment.
To address these concerns, OpenAI is exploring embedding metadata into text, which would be cryptographically signed and have no false positives. The company is still in the early stages of this process and has yet to determine its effectiveness.
Source: https://www.theverge.com/2024/8/4/24213268/openai-chatgpt-text-watermark-cheat-detection-tool