OpenAI has built a text watermarking method to detect ChatGPT-written content — company has mulled its release over the past year

OpenAI has already built and tested a tool to detect whether any written content has been created using ChatGPT. However, the Wall Street Journal reports that the company is holding back the tool from public release because of several concerns.

The tool adds a pattern to how the large language model (LLM) writes its output, allowing OpenAI to detect if ChatGPT created it. However, the pattern remains unnoticeable to humans, thereby not impacting the LLM’s quality. Internal documentation says that the tool is 99.9% effective in detecting ChatGPT’s output, but OpenAI has yet to release it.

While text watermarking is highly effective for detecting content written by ChatGPT, it cannot work with output from other LLMs like Gemini AI or Llama 3. Furthermore, this technique can be easily circumvented. For example, you can insert ChatGPT’s output in Google Translate, convert it to another language, and then back to English, effectively removing the watermarking.

It could even be as simple as asking the AI to insert a unique character, emoji, or even a short phrase between words and then deleting them later (via Microsoft Word’s Find and Replace function, for example) or asking another LLM to rephrase the entire output entirely is enough to break the detection tool.

Another point of contention OpenAI has on releasing the tool is that it says it could be biased against non-native English writers. The company previously released a generative AI text detection tool last year. Still, it only took seven months before OpenAI pulled the plug on it because of its low detection rate and propensity to deliver false positives. It even has led to the point that a professor once failed an entire class because all the submitted papers were inaccurately marked as AI-generated.

Furthermore, OpenAI also considers its customer base, with customer surveys showing that 69% of ChatGPT users believe that the tool would result in false accusations of AI cheating. 30% even said they would likely switch to a rival LLM if OpenAI deployed the tool. The company is facing another issue with the broad deployment of the ChatGPT detector, as other users could reverse engineer OpenAI’s watermarking technique and release a plug-in or app to neutralize it.

Nevertheless, OpenAI understands the risk that AI-generated content poses to society and has been looking at alternatives to the text watermarking technique. There’s also a demand for an AI detector, with internal data showing that 80% of people asked globally support its existence.

It’s still a question of whether OpenAI will release its text watermarking tool and detector. However, it also understands that as one of the leading organizations in the development of AI, it must also take steps to ensure the responsible use of its tools. Sources say that OpenAI has, until this fall, to sway public opinion around AI transparency. We don’t know what this means, but there’s one thing we’re sure of: we must be critical of what we read and ensure we’re getting the truth.