Web Desk
Aug 05, 2024
In the era of immensely developed artificial intelligence (AI), the boom in tools to detect and label AI-generated content is too notable.
Similarly, OpenAI too had in place a system of watermarking ChatGPT-generated text and a tool to detect that watermark.
Citing a WSJ report, The Verge reported that both of the watermarking tools have been in the OpenAI books for almost a year now, but an internal discord in the company has been causing a delay in its release. This means that OpenAI as of now is not up to roll out its watermarking tool for ChatGPT-written content.
Read more: M4 MacBook Pro, M4 Mac mini, M4 iMac to launch this year — Rumour
The watermarking opted for by OpenAI is about adjusting how its LLM predicts the most likely words and phrases to follow previous ones, paving the way for a repetitive and detectable pattern.
The provision of any tool or medium to detect and label AI-written content could be of great aid for teachers to prevent students from resorting to AI assistants to get their writing assignments done.
The report stated that OpenAI inspected and found its watermarking system did not alter ChatGPT's text output quality.
The company also conducted a survey to assess people's views on its AI watermarking, finding that people across the globe in favour of an AI detection tool outnumbered the opposers by a margin of four to one.
In a blog post published today, OpenAI claims that its method of watermarking ChatGPT-written text is “99.9% effective,” and resistant to “tampering, such as paraphrasing.”
However, OpenAI says techniques like rephrasing with another model make it “trivial to circumvention by bad actors.”
