OpenAI has been facing an internal debate over the release of its text watermarking system for ChatGPT-created text and the tool to detect the watermark. The company has had the system ready for about a year but is unsure whether to make it available to the public due to potential implications on its bottom line.
OpenAI’s watermarking system works by adjusting how the model predicts the most likely words and phrases that will follow previous ones, creating a detectable pattern in the text. This system could be a game-changer for teachers looking to deter students from submitting AI-written assignments.
A survey commissioned by the company found that people worldwide supported the idea of an AI detection tool by a margin of four to one. This shows that there is a demand for tools that can help detect AI-generated content.
Despite concerns about potential negative effects on user sentiment, some employees at OpenAI believe that watermarking is an effective solution. The company claims that its method is 99.9% effective and resistant to tampering, such as paraphrasing.
In response to user concerns, OpenAI is reportedly exploring alternative methods, such as embedding metadata, that may be less controversial among users. While it is still early in the exploration process, the company is hopeful that this approach could address the issue without alienating users.
The internal debate at OpenAI highlights the delicate balance between protecting against misuse of AI-generated content and maintaining user trust and satisfaction. As the company continues to explore different strategies, it will be important to consider both the effectiveness of the solution and the impact it may have on users. Ultimately, finding a solution that strikes the right balance will be crucial for OpenAI’s future success.
Leave a Reply