OpenAI has developed a system for watermarking ChatGPT-generated textual content and a software tool to detect the watermark over the past few years. Despite being divided internally, the company remains undecided on launching the initiative. On one hand, it appears that making a decision is necessary; on the other, it may harm its bottom line?
OpenAI’s watermarking technique involves subtly altering the language model’s predictions for likely phrases and sequences to create a discernible pattern. That’s a simplification; however, you may also try searching Google for additional information.
Developing methods to identify AI-generated content may prove beneficial in discouraging students from using artificial intelligence tools to complete academic assignments, thus maintaining the integrity of higher education. Despite implementing watermarking, the corporation found no significant impact on the quality of its chatbot’s written content. According to a recent survey commissioned by the corporation, a significant majority of people globally backed the idea of AI-powered detection software, with a substantial margin of 4:1 in its favour.
After presenting its story, OpenAI revealed that it has been working on watermarking text content in a way that allows it to identify original creations with 99.9% efficiency, as per documents obtained by the outlet. The company claims this methodology is resistant to tampering, including paraphrasing. However, it notes that methods like rewording using another model make it “trivial” for malicious actors to circumvent its detection. OpenAI also expressed concern about stigmatizing AI tools’ usefulness for non-native speakers.
OpenAI acknowledges a potential concern: if implemented, watermarking might alienate around 30% of its surveyed ChatGPT users who explicitly stated they would use the software less as a result?
Despite criticisms, some workers surprisingly assert that watermarking remains an effective tool. In response to lingering consumer concerns, some have called for alternative approaches that are “clearly less contentious among customers but untested.” Meanwhile, in a recent blog post update, the company revealed it’s “in the exploratory stages” of considering metadata integration. While still in its infancy, it’s premature to gauge the effectiveness of this system; nonetheless, due to its cryptographic signatures, the possibility of false positives is effectively eliminated.