OpenAI introduces invisible text watermarking for ChatGPT and Codex in EU

OpenAI is adding invisible watermarks to text generated by ChatGPT and Codex in the European Union, according to BleepingComputer, Digital Trends, and The Verge. Digital Trends, 9to5Mac, and TechCrunch report that the rollout is designed to comply with requirements under the EU AI Act, which mandates that providers of generative artificial intelligence systems mark AI-generated text so that it can be identified. As FTMQ Tech reported earlier, the transparency rules of the EU AI Act took effect on August 2. [1][2][3][4][6][16]

The new watermarking system is named textGrain, BleepingComputer, Digital Trends, and The Verge report. BleepingComputer explains that the watermark is not visible when reading or copying text. Instead, textGrain slightly alters the model's word choices to generate a statistical pattern that detector tools can subsequently recognize. [1][2][3]

According to Digital Trends and TechCrunch, OpenAI acknowledged that the technology faces limitations, as edits, short text, and other changes can make the invisible marks harder to detect. The Verge reports that OpenAI said textGrain matched or exceeded alternative approaches such as Google DeepMind's SynthID, but noted that the feature will not serve as the global default for now. [2][3][6]

BleepingComputer reports that developers using the application programming interface worldwide can opt in to watermarking starting today, though it remains turned off by default. BleepingComputer also noted that OpenAI is opening applications for its watermark detector tool. [1]