With concerns growing about students using ChatGPT to cheat on their homework, there’s been a lot of talk about how to spot AI-generated work. Although there are some tools designed to detect AI text, their reliability is often shaky. Recently, OpenAI has hinted at a solution: a new "text watermarking method." However, this method isn’t ready for widespread use just yet.
Watermarking And Metadata: How They Work
OpenAI has been developing two main methods to detect AI-generated text:
Watermarking: This technique involves adding a hidden text watermark to the AI-generated content. It’s proven to be quite effective, even against attempts to alter the text, like paraphrasing.
Metadata: Instead of using watermarks, OpenAI is also exploring the use of metadata. This method involves adding cryptographic signatures to the text, which makes it hard to tamper with or forge.
Another method they’re considering is classifiers, similar to those used in email systems to sort messages into spam or important categories. These classifiers could help identify AI-generated text in essays. Despite these promising methods, there are challenges to overcome:
Tampering Issues: While watermarking is good at detecting some types of tampering, it’s less effective against more sophisticated methods. For example, translating text through different languages or rewording it with another AI model can bypass the watermark.
Impact on Users: There are also concerns about how these tools might affect certain groups. For non-native English speakers, AI can be a valuable tool for writing. Using these detection methods might unfairly penalize those who rely on AI for help with their language skills.
OpenAI’s new methods for detecting AI-generated text are promising but not without flaws. The company is still working on refining these tools to ensure they are both effective and fair.
