Introduction

The European Union's new transparency rules for AI-generated content are now in force, and one of the most prominent technical attempts to comply is already under scrutiny. OpenAI introduced textGrain in October as a watermarking system for its language models, designed to flag machine-written text. However, early testing indicates the method can fail on short passages and after light editing, leaving open whether it satisfies the EU AI Act's requirements.

What Happened

Under Article 50 of the EU AI Act, which took effect on August 2, providers of systems that produce synthetic text, audio, images, or video must mark their outputs in a machine-readable format. The law asks that these technical solutions be effective, interoperable, robust, and reliable to the extent that is technically feasible. OpenAI's textGrain was released with compliance in mind, but internal tests described in a technical report show detection rates of roughly 80 percent for passages around 150 words and about 95 percent for texts near 300 words. The system also carries a 1 percent false-positive rate. When researchers replaced roughly 10 percent of words with synonyms in 400-token passages, detection dropped from 92 percent to 66 percent; at 25 percent replacement, it fell to 17 percent. OpenAI has stated that watermarking alone cannot determine whether content is AI-generated and should work alongside other provenance tools, such as metadata credentials and Google DeepMind's SynthID for images and audio.

Why This Matters

The EU's AI Act does not yet set measurable thresholds for watermark performance, and the bloc's voluntary Code of Practice exempts outputs shorter than 200 tokens. That ambiguity creates uncertainty for companies like OpenAI. If regulators ultimately determine that textGrain does not meet the legal standard, OpenAI could face penalties of up to 7 percent of its annual income. Independent verification is still pending; researchers at Cornell University and the Kempelen Institute of Intelligent Technologies in Slovakia have only recently received access and have not completed their evaluations. One AI ethics researcher has also urged OpenAI to release the system as open-source so the broader community can stress-test it.

Key Takeaways

  • The EU AI Act's transparency obligations for synthetic text began on August 2.
  • OpenAI's textGrain watermark is invisible to readers but can be difficult to detect in short texts or after editing.
  • Detection rates fall significantly when even a small portion of a passage is reworded.
  • The law requires machine-readable marking but lacks specific performance benchmarks.
  • Non-compliance could expose OpenAI to fines reaching 7 percent of annual revenue.
  • Independent academic review of textGrain has not yet been published.

Conclusion

textGrain represents a notable step toward labeling AI-generated text, but its current limitations highlight a broader challenge: turning regulatory principles into reliable technical practice. Until independent researchers finish their assessments and the EU clarifies what effective and robust means in measurable terms, it remains unclear whether OpenAI's approach will satisfy European regulators. For now, the watermark is best understood as one piece of a larger provenance strategy rather than a standalone solution.