It had been working on a text watermarking technique to identify essays generated using ChatGPT. Recently, a report came out saying that OpenAI hadn’t released this tool. OpenAI is said to have disclosed that the watermarking technique, while very accurate in some cases, was brittle to some sorts of tampering, including translation and rewording by other models, and text changes.
The company contemplates various methods for ascertaining the originality of the text. Among them are watermarking, studying classifiers, and metadata. As efficient as watermarking could prove to be, it might be associated with some drawbacks, like further stigmatizing the use of AI by non-native English speakers. These considerations are one important reason for keeping the tool under wraps.
Their selection of prioritizing authentication tools for audio-visual content before text watermarking illustrates the challenges and broad implications that text provenance can entail. The company envisions walking a fine line between effectiveness and ethics.
The release of the watermarking tool is still very indefinite as OpenAI continues to conduct its research. It assesses the risks and benefits associated with the different methods. For the present, the text watermarking tool remains in the working stage.
OpenAI is very careful when it comes to watermarking because of the many challenges involved. It seeks to find a resolution that supports both accuracy and ethical concerns.
