OpenAI has admitted that watermarking AI-generated text remains a flawed technology with major shortcomings, even as the company begins deploying it across the European Union to satisfy incoming regulations.
The move comes as Brussels tightens oversight of ChatGPT under its most stringent platform rules. Starting in the coming weeks, invisible watermarks will appear on text produced by ChatGPT and Codex within the EU. Meanwhile, API customers worldwide can choose to enable text watermarking for certain models, and a limited group of researchers will gain access to test the detection tools.
Transparency mandates under the EU AI Act took effect this August. Those rules require AI systems to inform users when they are interacting with artificial intelligence and to clarify how content gets generated or modified.
OpenAI calls its watermarking method TextGrain. The system embeds an invisible statistical pattern into the model’s word selection, and a detector scans for that pattern to determine whether a passage carries the mark. But the technology struggles with short texts and human-edited content. Testing shows detection accuracy drops from nearly 100 percent on untouched output to under 20 percent when a quarter of the words change.
As a result, OpenAI itself frames the rollout cautiously. Company representatives described text watermarking as an early-stage tool with significant constraints, noting that the phased introduction reflects both legal requirements and the technology’s current limits.
The ChatGPT developer numbers among roughly 200 signatories to the EU AI Act’s voluntary Code of Practice on AI-generated content transparency. Anthropic, Mistral, Meta, and Microsoft have also signed on. OpenAI already labels supported images and audio with C2PA and SynthID watermarks, while Anthropic began embedding watermarks in EU-generated Claude text this August.
Regulators continue pushing detection technology forward despite its accuracy gaps, driven by mounting concerns over manipulated media’s societal impact. For now, the gap between regulatory ambition and technical reality remains wide.
