Source : INDIA TODAY NEWS
Tired of AI slop? OpenAI is making it easier to identify what was written by its AI tools with watermarks. The company is set to add an invisible watermark to text generated by ChatGPT and Codex. These watermarks, OpenAI says, will not be visible to humans, but can be seen by its detection tools. For now, the watermark will only be rolled out in the European Union (EU) with no global launch planned.
advertisement
OpenAI calls its watermarking system textGrain. This system works by tweaking the pattern of words that a model uses. This pattern cannot be seen by humans, but a detector can pick it up if it looks for it. Since the watermark works by the words generated by the AI itself, it can still be detected even after it is copied and pasted elsewhere.
In case you are wondering, the watermark system does not bring any meaningful change in performance as per OpenAI. The company cited similar scores across benchmarks, including Artificial Analysis Intelligence Index and AutomationBench, for both non-watermarked and watermarked texts.
OpenAI says that the watermark does not identify the user or the prompt given. When the detector recognises the watermark, it will only report it as is. This system is similar to Google DeepMind’s SynthID for text. Anthropic, which released its own watermark for Claude in August, likely uses a similar mechanism as well.
Why is OpenAI adding watermarks to ChatGPT?
Under the EU AI Act, generative AI providers are required to make AI-generated text identifiable in a machine-readable way. To meet this requirement, OpenAI is rolling out textGrain in the EU in the coming weeks. The company said this limited launch also gives it room to learn from real-world use and feedback before deciding on wider deployment.
The watermark system will also be available from today as an opt-in feature for OpenAI’s API customers worldwide on select models. The company is also planning to give researchers and expert organisations access to its detector tool.
At the same time, the company says reliable detection cannot be guaranteed in everyday use. Shorter or more constrained text is harder to detect as the detector relies on word patterns. According to OpenAI, the detector identified watermarks in about 80 percent of 200-token psychology passages, compared with about 95 percent of 400-token passages. Detection was substantially lower for subjects such as mathematics, where there is less flexibility in word choice.
OpenAI explains that editing can weaken the watermark. In one evaluation of 400-token passages, replacing 10 per cent of the words with synonyms reduced detection from about 92 per cent to 66 per cent, while replacing 25 per cent reduced it to 17 per cent. The company also said translated text is also harder to detect.
The company also clarifies that a text watermark does not verify accuracy or measure how much a human contributed. The absence of a detected watermark does not prove a passage was written by a human for a number of reasons – be it editing, translations, older models or using other tools.
While OpenAI’s textGrain is limited to the EU by default as of now, it appears that AI companies globally are recognising the need for watermarks in text. Though it is unclear whether we will see such systems going mainstream at a global level just yet.
– Ends
SOURCE :- TIMES OF INDIA
