🛡 OpenAI launches invisible watermarks for AI text
OpenAI has published its approach to EU rules on the provenance of AI-generated text. The invisible watermark is the textGrain technology: a statistical shift in word selection that is then detected by a separate detector. Starting today, OpenAI API clients worldwide can opt in to enable labeling for specific models — it is disabled by default, cloud partners will have it enabled in the coming weeks, and the watermark will soon appear in the output of ChatGPT and Codex, for now only in the EU and on all plans.
🌍 For the industry, this is the first practical implementation of the EU AI Act's requirements — the EU Code of Practice on AI-generated content — for machine-readable labeling of generative text by a major provider. Along with the launch, OpenAI publishes the textGrain technical report: Entropy-Calibrated Watermarking for Language Model Text and promises to open-source the code. According to benchmarks for the flagship model Astra (GPQA Diamond, BrowseComp, Terminal-Bench 4.0), output quality with watermarking changes only slightly, but at launch only watermark generation is functional: the detector is closed, and its quality is confirmed by the vendor itself.
👤 The watermark is not visible to ordinary users: access to the detector at launch will be granted only to approved researchers and expert organizations under the rules of the EU Code of Practice. Accuracy depends on text length — about 80% for 200 tokens and about 95% for 400 tokens at a 1% false positive rate, and replacing 25% of words with synonyms reduces detection to 17%. Image and audio verification is already open at openai.com/verify, along with the Content Provenance API for developers; the absence of a watermark is not a sign of human text, and a 1% false positive rate carries the risk of accusing a living author of machine generation.
Source 1: https://openai.com/index/eu-text-provenance/ Source 2: https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf