OpenAI, an artificial intelligence lab, has developed a system that can effectively identify text produced by AI chatbots such as ChatGPT, but it’s currently unsure about disclosing this tool publicly. This development comes as difficulties in differentiating between AI-written text and human-written text continue to rise and become more complex. Particularly educators are finding it hard to tell if students’ assignments are their original work or if they are leveraging AI tools like ChatGPT to generate them.
OpenAI’s watermarking method has demonstrated substantial precision and has shown resistance to common altering practices such as paraphrasing. However, the tool isn’t infallible and can be tricked. For example, translating the text or using another LLM for rewording can circumvent the watermarking system. Similarly, using an AI tool to add and then remove a special character in between words can disable the watermark’s effectiveness.
But these aren’t the only reasons OpenAI is hesitant to release the watermarking method. ChatGPT has proven to be advantageous to non-English speakers or those trying to improve their language skills. OpenAI is concerned that the availability of watermarking may unduly affect these users and may stigmatize the use of AI-assisted writing.
Alternatively, OpenAI’s engineers are exploring the use of metadata as an alternative text provenance tool. Images produced by AI models like DALL-E 3 already utilize “C2PA metadata”. There are potential benefits to metadata, such as its immunity to false positives due to cryptographic signing. However, one of the main concerns around using metadata is how easily it can be removed. While the specifics of how metadata would be applied to text are still under wraps, it has been shown that removing C2PA metadata from AI-generated images is surprisingly straightforward.
OpenAI’s reluctance to release its tool could also be business-related. The detector only identifies text generated by their AI, ChatGPT. Users may shift to another platform if they realize their AI-made content can be easily detected. Moreover, according to a Wall Street Journal report, OpenAI’s tool is almost 99% efficient and has been ready for launch for over a year now.
A global survey conducted by OpenAI reveals that people by a 4:1 margin favor an AI detection tool. However, almost one-third of ChatGPT users don’t approve of an AI-text detector, fearing that their AI-produced content could be easily identified. Therefore, despite the push for transparency in AI-generated content, the end-users want it to be easily distinguishable only when it’s somebody else’s content but not when it’s theirs.
Leave a Reply