OpenAI has developed a new tool designed to identify when students solicit ChatGPT’s help for completing assignments. However, as reported by The Wall Street Journal, there’s ongoing debate within the company about whether to make this tool publicly available.
A representative for OpenAI has confirmed to TechCrunch that the organization is exploring the text watermarking technique highlighted in the Journal’s article. Yet, the spokesperson emphasized the cautious stance OpenAI is taking towards a public release, citing “the complexities involved and its likely impact on the broader ecosystem beyond OpenAI.”
The representative elaborated that the text watermarking initiative shows technical promise but is fraught with challenges, including potential evasion by malicious users and the likelihood of adversely affecting particular groups, like those who are not fluent in English.
This initiative marks a departure from earlier attempts to identify AI-generated text, which have not been very successful. In fact, OpenAI discontinued its former AI text detection tool last year due to its ineffectiveness.
Text watermarking would allow OpenAI to pinpoint submissions crafted by ChatGPT by implementing subtle changes in word selection during text generation. This method aims to embed an undetectable mark that could later be recognized by specific detection tools.
In light of the Journal’s report, OpenAI revisited a May blog post, updating it with insights on their research into AI-generated content detection. The update discusses how text watermarking has been “highly accurate and even effective against localized tampering,” yet less effective when up against extensive manipulation, such as through translation services, rewording with another AI model, or manipulating the text with special characters and then removing them.
Consequently, OpenAI acknowledges the limitations of this method, noting it could be easily bypassed by determined individuals. The update also reiterated concerns about the potential for text watermarking to discriminate against individuals who rely on AI for English language assistance, highlighting the essential balance between technology use and ethical considerations.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
