OpenAI Adds Hidden Watermarks to ChatGPT and Codex

Tired of generic AI-generated content? OpenAI is making it easier to identify text created by its tools through watermarking. The company plans to add an invisible watermark to text generated by ChatGPT and Codex. According to OpenAI, these marks will not be visible to humans but can be detected by its own analysis tools. For now, the watermark will only be implemented in the European Union (EU), with no plans for an immediate global rollout.

OpenAI calls its watermarking system "textGrain." This system works by slightly modifying the pattern of words used by the model. This pattern is imperceptible to humans, but a detector can identify it if specifically looking for it. Since the watermark relies on the words generated by the AI itself, it remains detectable even if the text is copied and pasted elsewhere.

In case you're wondering, the watermarking system does not significantly affect performance, according to OpenAI. The company noted that benchmark scores-such as those from the ‘Artificial Analysis Intelligence Index’ and ‘AutomationBench’-were similar for both watermarked and non-watermarked text.

OpenAI assures that the watermark does not identify the user or the original prompt. When the detector recognises the mark, it simply reports its presence. This system is similar to SynthID for text, developed by Google DeepMind. Anthropic, which launched its own watermark for Claude in August, likely uses a similar mechanism.

Why is OpenAI adding watermarks to ChatGPT? Under the EU AI Act, providers of generative AI are required to make AI-generated text identifiable through machine-readable formats. To meet this requirement, OpenAI will roll out textGrain in the EU over the coming weeks. The company noted that this limited launch would also allow it to learn from real-world usage and user feedback before deciding on a wider-scale implementation. Starting today, the watermarking system will also be available as an optional feature for OpenAI API customers worldwide, applicable to selected models. The company also plans to provide access to its detection tool to researchers and specialised organisations.

However, the company cautions that reliable detection cannot be guaranteed in everyday use scenarios. Shorter texts or those with greater constraints are more difficult to detect, as the system relies on word patterns. According to OpenAI, the tool identified watermarks in approximately 80 per cent of 200-token psychology excerpts, compared to 95 per cent for 400-token excerpts. The detection rate was significantly lower in subjects such as mathematics, where there is less flexibility in word choice.

OpenAI explains that editing can weaken the watermark. In an evaluation using 400-token excerpts, replacing 10 per cent of the words with synonyms lowered the detection rate from approximately 92 per cent to 66 per cent, while replacing 25 per cent reduced it to 17 per cent. The company also noted that translated texts are more difficult to detect.

Furthermore, the company clarifies that a text watermark does not verify accuracy or measure the extent of human contribution. The failure to detect a watermark does not prove that a passage was written by a human, due to various factors such as editing, translation, the use of older models, or the use of other tools.

Although OpenAI’s textGrain tool is currently limited to the EU, AI companies worldwide appear to be recognising the need to incorporate watermarks into text; however, it remains unclear whether these systems will be adopted globally in the near term.