According to The Wall Street Journal, OpenAI has been working on a watermarking system for ChatGPT text for approximately a year, along with a tool to identify the watermark. However, there are disagreements within the corporation about whether to release it. While doing so might appear responsible on the one hand, it might be detrimental to the company’s profit line on the other.
To create a discernible pattern, OpenAI’s watermarking is defined as modifying the model’s prediction of the most likely words and phrases that will come after earlier ones.
Also Read: Microsoft Lists OpenAI as Competitor Despite $13 Billion Investment
Teachers attempting to dissuade students from turning in writing projects to artificial intelligence may find it helpful to provide any means of identifying content written by AI. According to The Journal, the company discovered that watermarking had no impact on the text output quality of its chatbot. “People worldwide supported the idea of an AI detection tool by a margin of four to one,” according to a survey the business commissioned, according to the Journal.
In a blog post update that TechCrunch saw today, OpenAI verified that it has worked on watermarking text after the Journal published its article. According to materials the Journal examined, the company claims that its approach is highly accurate (99.9% effective) and impervious to “tampering, such as paraphrasing.”
However, it claims that methods such as rewording using a different model make it “trivial to circumvent by bad actors.” The business expresses concern regarding the stigmatization of AI technologies’ applicability to non-native speakers as well.
Also Read: OpenAI ChatGPT Voice Rolled Out Plus Users, Check How to use it on Mobile
However, it appears that OpenAI is also concerned that watermarking could alienate ChatGPT users since over 30% of them reportedly informed the business that they would use the program less if watermarking was added.
It is alleged that some staff still believe watermarking works well despite this. However, the Journal notes that some have proposed using approaches that are “potentially less controversial among users but unproven” in response to persistent user attitudes.
The corporation stated that it is “in the early stages” of investigating embedding metadata in an update to its blog post. Although it’s “too early” to tell how effectively it will function, it states that there won’t be any false positives because it’s cryptographically signed.
Also Read: OpenAI’s SearchGPT: AI-Powered Search Engine with Advanced Summarization Features
