OpenAI to watermark ChatGPT and Codex text output in the EU under AI Act compliance
2026-10-05
OpenAI announced it will begin watermarking ChatGPT- and Codex-generated text in the European Union to comply with the EU AI Act's provenance rules, with detection access starting with researchers. TechCrunch notes editing text can make the invisible watermarks harder to detect.
Significance 3: A concrete, named regulatory-compliance commitment tied to a specific law is a meaningful policy/governance development, though its real-world detection efficacy is already undercut by OpenAI's own admission that editing weakens it.
Implications · machine-drafted, not owner judgment
This sets a compliance template other labs serving EU users will likely need to match, and researcher-first detection access is a modest transparency step; but OpenAI's own caveat that editing defeats the watermark means the practical content-provenance problem remains largely unsolved even where compliant.
- Whether Anthropic, Google, or Meta announce comparable EU watermarking commitments
- Independent testing of watermark robustness against common edits
Sources
- primary Our approach to EU text provenance rules retrieved 2026-10-06
- press OpenAI will start watermarking ChatGPT's text in the EU retrieved 2026-10-06