Our approach to EU text provenance rules

• Our approach to EU text provenance rules

  OpenAI NewsOct 5, 2026First report

  UpdateUpdate to: Supporting Europe’s work in ensuring a trustworthy AI ecosystem, 11 Jun

  Our summaryRead the original

• OpenAI will start watermarking ChatGPT’s text in the EU

  TechCrunch AIOct 5, 2026

  Update

  Our summaryRead the original

• OpenAI to watermark ChatGPT, Codex text in EU under AI Act

  DigiTimesOct 6, 2026

  Update

  Our summaryRead the original

1. Revision 3Assessed 6 Oct

   Update

   Update from DigiTimes

   • Impact: medium → low
   • Indicators to watch: Whether the EU rollout reaches eligible ChatGPT and Codex users as described, Independent evaluation of detector accuracy across text types and editing levels, Whether approved researchers gain access to the detector and publish reliability findings, Whether API watermarking adoption extends beyond opt-in use → EU rollout timing and geographic coverage, Detection reliability after text editing or paraphrasing, Whether watermarking is applied consistently across ChatGPT and Codex
   • Evidence limitations: The article reports OpenAI’s announcement and test results; it does not provide independent validation of the method or its detection performance. → The supplied report is brief and does not state a rollout date, technical method, coverage limits, or independent validation.
2. Revision 2Assessed 5 Oct

   Update

   Update from TechCrunch AI

   • Status: assessed → updated
   • Indicators to watch: Whether watermarks are deployed to eligible ChatGPT and Codex text output in the EU as planned, Detector access criteria and availability to approved researchers, Independent replication of the reported detection rates, Whether the planned open-source release occurs and under what terms → Whether the EU rollout reaches eligible ChatGPT and Codex users as described, Independent evaluation of detector accuracy across text types and editing levels, Whether approved researchers gain access to the detector and publish reliability findings, Whether API watermarking adoption extends beyond opt-in use
   • Evidence limitations: Performance results and planned access arrangements are reported by OpenAI; the source notes that real-world detection may be less reliable and does not provide independent validation. → The article reports OpenAI’s announcement and test results; it does not provide independent validation of the method or its detection performance.
3. First assessmentAssessed 5 Oct

   UpdateUpdate to: Supporting Europe’s work in ensuring a trustworthy AI ecosystem, 11 Jun

   First assessment, from OpenAI News

   • Impact: not set → medium
   • Status: developing → assessed
   • Confidence: not set → medium
   • Event type: model_release → product_launch
   • Assessment: not set → OpenAI’s phased text-watermarking rollout adds a practical provenance measure, but its own results show substantial limits that make independent validation and careful interpretation important.
   • Indicators to watch: (none) → Whether watermarks are deployed to eligible ChatGPT and Codex text output in the EU as planned, Detector access criteria and availability to approved researchers, Independent replication of the reported detection rates, Whether the planned open-source release occurs and under what terms
   • Evidence limitations: not set → Performance results and planned access arrangements are reported by OpenAI; the source notes that real-world detection may be less reliable and does not provide independent validation.
   • Representative source: not set → Our approach to EU text provenance rules (OpenAI News)

No maturity ladder applies to this desk.