OpenAI Launches Anti-AI Detection Tool: Two Years of Research Results Raise Questions

The subtle interplay of language.

ChatGPT, initially seen by students as a divine assistant for writing papers, may now become their nightmare. According to inside information, OpenAI has secretly developed an "anti-plagiarism tool" that can accurately detect whether ChatGPT was used in essays or reports. This technology has been debated within OpenAI for two years and was actually ready a year ago.

The principle of this anti-cheating tool is to subtly adjust the probability of word choices when ChatGPT generates text, embedding specific probability patterns as watermarks. Although undetectable by humans, OpenAI's detection system can identify them. It is claimed that as long as the text is long enough, the detection accuracy can reach 99.9%.

The specific implementation may be through changing the frequency of token occurrences in sentences. For example, in the sentence "My favorite fruit is __", certain fruits are given a higher probability of occurrence, forming a kind of code. These subtle probability changes do not affect reading but can be decoded by OpenAI.

Currently, OpenAI has not made this technology public, only acknowledging its research in a blog post on their official website. Various AI text detection tools are already available on the market, such as Google's SynthID and GPT Zero. However, these tools also have limitations, such as low accuracy in detecting complex texts.

The reasons why OpenAI has not released the watermarking function may include:

  1. The technical difficulty of watermarking text. There are various ways of expressing text, making statistical analysis challenging.

  2. The watermark might be cracked. Some have found that multi-language translation can bypass detection.

  3. There is controversy over whether watermarking is necessary. It may affect the development and application of AI technology.

  4. Concerns about misuse. It could be used for censorship or surveillance.

  5. Commercial considerations. It may affect user experience and product competitiveness.

Overall, AI text detection technology is still developing, and how to balance technological development, user experience, and ethical issues in the future is worth attention.