OpenAI adds text watermarking to ChatGPT and Codex, starting with EU users

Why This Matters

• AI-generated text is difficult to distinguish from human writing, and watermarks give platforms and regulators a tool to trace content back to its source.
• The EU-first rollout shows how the EU AI Act is directly shaping product decisions at major AI labs.
• If the watermark proves robust, similar requirements could spread to other jurisdictions.

Background

Text watermarking works by embedding a subtle statistical pattern in generated text that can be detected later. The EU AI Act requires AI-generated content to be identifiable, which has pushed companies including OpenAI and Anthropic to ship practical solutions. The approach has tradeoffs: watermarks aid attribution but can be evaded, and some techniques have faced criticism over their effect on non-English text.

Key Perspectives

OpenAI: Says its watermarking matches or exceeds rival approaches and shows minimal impact on performance in benchmarks. Anthropic and Google DeepMind: Their SynthID system is the reference point OpenAI is measured against, and Anthropic adopted the same technology for its own watermarking in August. Critics: Not everyone welcomed the feature, and the company has not guaranteed the method is foolproof, leaving questions about evasion and unintended effects.

What to Watch

• Whether OpenAI expands watermarking beyond the EU to other regions.
• Independent testing of how well the watermark survives editing, rewriting or translation.
• Any regulatory reaction to claims that the watermark can be evaded or produces false positives.