With the rising popularity of ChatGPT and other large language models, internet users have started flooding the web with AI-generated text. They are seen by many as the game changer that is redefining the way people work, as they help automate tedious tasks and save time.
Nevertheless, the AI language tools also come with risks, as they create an illusion of correctness by generating sentences that seem factually accurate even when they are not. Moreover, they pave way for the spread of misinformation and the rise in plagiarism and cybercrime.
The question: is it possible to detect AI-generated text?
It is tricky, but not impossible.
Detection Tools
At the end of January, OpenAI, the American research laboratory behind ChatGPT, released a classifier trained to distinguish between AI-written and human-written text. OpenAI is upfront that the classifier has limitations and is not 100% reliable, especially on short texts containing less than 1,000 characters. Hence, it says, the tool “should not be used as a primary decision-making tool, but instead as a complement to other methods of determining the source of a piece of text”.
Edward Tian, a computer science student at Princeton University, has also come up with an experimental tool called GPTZero to detect whether a text is written by ChatGPT. In the process, the tool measures the complexity of text and compares the variations of sentences before determining the result. But still, it has its limitations, including the inability to detect a mix of AI-written and human-written text.
Watermarks
Researchers at the University of Maryland have looked into ways to use watermarks to spot text created by large language models. Simply put, the watermarks are hidden patterns put in AI-written text, allowing people to detect machine-generated sentences. This could be useful, though it would only work if it is embedded in the chatbot system by the creators right from the start. Moreover, the watermark algorithm can potentially make false accusations of plagiarism.
Scott Aaronson, a guest researcher at OpenAI, said that the team working on the watermarking scheme at the company exploited the fact that OpenAI controlled its own servers.
“So, it can do the watermarking using a secret key, and it can check for the watermark using the same key,” Aaronson said. “In a world where anyone could build their own text model that was just as good as GPT, what would you do there?”