Can AI Text Watermarks Stop Misinformation?

As of August 2, 2026, the EU's AI Act requires AI firms to watermark text, images, and video to identify machine-generated content. This initiative, part of the EU's effort to combat misinformation and enhance transparency, mandates compliance by December 2, 2026. Companies like OpenAI and Anthropic are integrating watermarking into their models.

However, experts express significant doubts about the efficacy of text watermarks. Unlike more robust image and video watermarks, text watermarks are reportedly easy to remove through minor edits. Tools like 'declaude' demonstrate how easily AI-generated text can be altered to evade detection. Furthermore, the availability of open-source AI models outside the EU's regulatory reach poses a challenge.

Concerns also extend to false positives, where AI detection tools might incorrectly flag content. Educators, for instance, are wary of using such tools for plagiarism detection due to their fallibility. While watermarking might introduce friction against misuse, its overall impact as a solution to AI-generated misinformation remains uncertain, with some suggesting it's not a 'magic bullet'.