The Extended Brief

AI Text Watermarking May Not Be Able to Meet the EU’s New Rules

Brief by The AI News AI newsroom · Nov 5, 2026, 10:14 AM EST edition

Original reporting by IEEE Spectrum — AI — Anna Desmarais · published Nov 5, 2026, 10:00 AM EST

OpenAI's textGrain watermark, built to satisfy the EU's AI-text transparency mandate, often fails after editing or on short passages, leaving compliance and detection in doubt.

Key points

• Tests show OpenAI's textGrain watermark often fails to survive editing or text under 150 words. source ↗
• The EU now requires AI-generated text to be watermarked so it can be identified as machine-made. source ↗
• OpenAI released textGrain in October, developed to comply with the new EU law. source ↗
• No clear legal benchmark exists for how well watermarks must work, leaving textGrain's compliance uncertain. source ↗
• textGrain treats entropy as a budgetable resource, letting developers trade watermark detectability against output variation, co-author Xiang Li says. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Test textGrain or any watermark you rely on against your own editing and paraphrasing pipelines before counting it toward EU compliance.
• Benchmark watermark detection rates on your short outputs specifically, since texts under 150 words often lose the signal.
• Plan a fallback disclosure or provenance mechanism, because with no legal robustness benchmark, watermarking alone may not satisfy EU transparency rules.

Context

Text watermarks embed a hidden statistical pattern into AI-generated text by using a secret key to nudge which tokens the model picks; a detector holding the same key can then check for the pattern. The EU now mandates such marking so AI text can be identified, but editing or short passages can erase the signal. OpenAI's textGrain, released in October, lets developers set an entropy budget controlling how strongly the watermark is embedded.

What to watch

• EU guidance or a ruling defining a concrete robustness benchmark would settle whether textGrain complies.
• Independent replication of the editing and short-text tests would confirm or soften the reported failures.

Related briefs

• In Its Race to Field AI, The Pentagon’s Testing Can’t Keep Up
• AI on the Ballot
• DeepSeek V5 is planned to be pre-trained on Huawei Ascend chips.
• China’s Largest Startup Deal Ever Lifts Global Venture Funding To $59B In October

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.