“The EU now mandates that AI-generated text be watermarked for identification, but tests on OpenAI's textGrain watermark — released in October 2024 — reveal it often fails when text is edited or falls under 150 words. With no defined legal benchmark for watermark reliability, OpenAI faces significant regulatory and financial risk in Europe. This gap between technical capability and legal expectation is a growing tension for the entire AI industry.”
Key Takeaways
- The EU now legally requires AI-generated text to carry detectable watermarks under its new transparency rules.
- OpenAI's textGrain watermark, released October 2024, fails to survive text editing and breaks down on content under 150 words.
- No legal benchmark currently defines how reliable a watermark must be to satisfy EU compliance, leaving OpenAI in uncertain legal territory.
OpenAI's new watermark tool may already be failing Europe's transparency rules.
Part of the 5 November 2026 AI news digest
trending_upWhy It Matters
This development exposes a fundamental mismatch between what current watermarking technology can deliver and what regulators expect — a gap that affects every major AI text provider operating in Europe, not just OpenAI. If textGrain is deemed non-compliant, OpenAI could face substantial fines under EU law, setting a precedent that could force costly redesigns across the industry. The lack of a clear technical benchmark also means companies have no reliable target to build toward, creating legal ambiguity that may chill AI product development in the EU market. Regulators and standardisation bodies will likely need to define measurable watermark thresholds soon, making this a space to watch closely in 2025.
FAQ
What is textGrain and why was it created?
textGrain is OpenAI's AI text watermarking system, released in October 2024, designed specifically to help comply with new EU transparency laws requiring AI-generated content to be identifiable. It embeds imperceptible signals into generated text to allow detection.
Why do the watermarks fail, and can this be fixed?
The watermarks degrade when text is edited or paraphrased after generation, and are unreliable on short texts under 150 words — common limitations of statistical watermarking methods. Fixing this is technically difficult because robust watermarks must survive natural text changes without distorting the original output.
What happens to OpenAI if textGrain doesn't meet EU rules?
Without a defined legal benchmark, it is currently unclear what penalty OpenAI would face, but non-compliance with EU AI transparency rules can carry significant fines. The uncertainty itself is damaging, as it puts OpenAI's ability to legally operate AI text services in Europe at risk.
