OpenAI catching cheating students with text watermark tool

OpenAI has just announced that it would be building a tool that could potentially catch students who are asking ChatGPT to write their assignments. But the company is still deliberating whether they would want to release this tool due to the complexity of this issue

AI detection tool with 99.9% success rate

The detection tool that OpenAI is building will help users to analyse texts and grade the writing's similarity to an AI generated writeup. But this isn't the first time ChatGPT has released its new AI-detection tool. The previous model that OpenAI created - AI classifier - was taken down on July 20,2023 due to its low accuracy rate. Now, the whole new and improved version of plagiarism tool that OpenAI is developing is using text watermarking to detect AI generated contents. OpenAI's watermarking is described as adjusting how the model predicts the most likely words and phrases that will follow previous ones, creating a detectable pattern. OpenAI claims that the text watermarking method has a 99.0% accuracy rate. On top of its high accuracy rate, the tool is also resistant to tampering, such as paraphrasing. However, the tool is only capable of catching watermarks from its own AI model, so text generated by other language models will not be picked up by the tool.

User apprehension and alternative solutions?

OpenAI has said in their announcement that although they have built this tool, they are hesitant to release this tool. This stems from OpenAI's apprehension that watermarking could turn users off and stop them from using ChatGPT. This is supported by a survey conducted by the company that shows that 30% of users expressed that they would be less likely to use the software if watermarking was implemented. Taking these sentiments into consideration, the company is considering other potential solutions to these challenges.

In light of nagging user sentiments, the company is also trying other alternatives that are potentially less controversial among users. One of the other methods the company is exploring is embedding metadata into text. But the company says that it is still early to know how well it will work, but it holds promises due to its cryptographic signature, which would eliminate false positives.

Implementations remain a big question mark

The watermarking system has proven that it could be very beneficial for educators aiming to prevent students from submitting AI-generated writing assignments. But there is also complexity involved with the distribution of the tool, for example deciding whether to provide the detector directly to educators or partnering with thirty-party companies specialising in plagiarism detection. This would also highlight the complexities involving implementing the tool and ensuring it serves its intended purpose without unintended consequences.