EU watermark rules test AI text detection

The European Union's push to require watermarks on AI-generated text is running into a practical problem: detection marks can be fragile. IEEE Spectrum reports that current watermarking technology may not satisfy the rules, based on tests of OpenAI's textGrain watermark, which was released in October to help comply with the EU law. The tests found that the marks often disappear or become unreliable after text is edited, paraphrased, or copied. That is a significant weakness for regulators who want watermarking to serve as an enforcement tool. A watermark is designed to leave a detectable signal in generated content, but it must survive normal use to be useful. If a short round of editing removes the signal, bad actors can evade detection easily, and platforms cannot confidently label content as AI-generated. The finding does not mean watermarking is useless. It may still help identify some unmodified outputs and support transparency efforts. But it suggests that watermark-based detection should not be treated as a silver bullet. Enforcement may need to combine watermarking with provenance metadata, account-level signals, content authentication standards, and human review. The EU rules are part of a broader effort to make AI-generated material more transparent as synthetic text becomes cheaper and more convincing. Yet the tests highlight a gap between policy goals and technical reality. Regulators want reliable labels; researchers can only provide tools with known limits. For platforms, the challenge is to avoid overclaiming what detection can do while still giving users useful context. The debate is likely to intensify as watermarking requirements expand and as AI systems learn to produce text that mimics human writing more closely. The core question is whether current methods can be made robust enough for law and platform policy, or whether expectations need to change.