OpenAI looking into ways to ensure students don't cheat in literature class

A report from The Wall Street Journal says that OpenAI, ChatGPT's creator company, has a tool that can spot essays written by its generative AI pretty accurately. Now, the company has shared a bit more about its research into test watermarking in an update to a blog post, and why it hasn't made its detection method available just yet.

OpenAI says that its teams have developed a text watermarking method that they are considering, but are also looking into alternatives on ways to prevent text written by ChatGPT from being used in cases where it shouldn't be... (like if your high schooler suddenly "came up" with an epic essay).

According to OpenAI, watermarking is one of the solutions. Watermarking may include classifiers and metadata. Watermarking is accurate but can be less useful if you use a translation system or if you reword the text using another generative model. It can also be bypassed if you ask the AI to insert a specific character after every word and then delete it later (boy, some of these high-schoolers are creative, it's that been done!)

Text watermarking could also impact other users, like for example non-native English speakers who'd like to refine their text and use AI to help them with writing.

According to the blog post, OpenAI has been considering all aspects of the issue with text provenance.

I agree with the fact that addressing this issue is a complicated thing. Of course, we don't want kids to be incapable of writing an essay themselves (or any text longer than 200 characters, for that matter). But also there are cases where using Generative AI to write things is totally acceptable and shouldn't be that big of a deal. Regulating AI is one of the challenges we face as the technology evolves, and in general, we should be looking more into that moving forward.