Industry Odisha Bureau, Oct 06: In latest news, OpenAI is introducing an invisible watermarking system for the text generated by ChatGPT and Codex. This will allow detection tools to identify its AI-generated content. The company says that the system, called textGrain. Will initially be rolled out in the European Union (EU) to meet the requirements under the EU AI Act.
Unlike visible watermarks, textGrain works by changing patterns in the words generated by an AI model. These changes are not noticeable to readers but can be detected by specialised tools. OpenAI says the watermark can remain detectable when text is copied and pasted elsewhere.
Further, the company also says that the system does not identify the user or the prompt used to generate the text. When detected, the tool only indicates the presence of the watermark.
OpenAI says watermarked text performs similarly to non-watermarked text on benchmarks including the Artificial Intelligence Index and Automation Bench. The company is also making the watermarking system available globally as an opt-in feature for API customers on select models.
However, OpenAI says detection is not guaranteed in every situation. Longer passages are easier to identify, while shorter or highly constrained text can be more difficult. In its testing, the detector identified watermarks in around 80% of 200 token psychology passages and about 95% of 400 token passages. Detection was substantially lower for mathematical content.
Meanwhile, it was said that editing generated content can also weaken the watermark. OpenAI found that replacing 10% of words with synonyms reduced detection from around 92% to 66%, while replacing 25% reduced it to 17%. Similarly, translated text can also be difficult to identify.
OpenAI said the watermark does not establish whether content is accurate or not. It also does not determine how much a human contributed to the content. Similarly, the absence of a detected watermark also does not prove that the content was written by a human.
