OpenAI is actively researching a text watermarking method aimed at identifying AI-generated text, alongside other detection solutions like classifiers and metadata.

The company has developed a tool that can detect text generated by ChatGPT with 99.9% accuracy, but it remains unavailable to the public.

Internal debates at OpenAI have delayed the release of this detection tool for two years, despite its technical readiness for about a year.

OpenAI is taking a cautious approach to the release of the detection tool due to the complexities and potential broader impacts it may have.

Concerns about misuse and the potential stigmatization of non-English speakers have led OpenAI to withhold the tool from public access.

The proposed text watermarking method could disproportionately affect non-native English speakers, raising concerns about its impact on their use of AI writing tools.

OpenAI has identified potential methods that bad actors could use to bypass the detection system, contributing to the decision to withhold the tool.

While the watermarking method shows high accuracy against localized tampering, it is less effective against broader text alteration techniques.

The watermarking technology aims to address concerns about the authenticity of AI-generated content, embedding an imperceptible watermark for later identification.

In a similar effort, Google is beta testing its own watermarking tool, SynthID, to detect text generated by its Gemini AI.

The project has been a topic of internal debate at OpenAI for the past two years, reflecting the challenges in balancing transparency and user attraction.