Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- October 5 distinguishes global API opt-in for selected models that day from coming-weeks rollout for eligible EU ChatGPT/Codex text.
- API watermarking defaults off. Check organization defaults and project settings, select currently supported models and save; no fixed list or per-request parameter is invented.
- Text detection requires an eligible organization's approved application. Public image/audio verification is not a public text detector.
- textGrain changes statistical word-choice patterns without hidden characters. Short text, code, paraphrase and translation affect detection; a negative result does not prove human authorship.
API opt-in is available; EU rollout is still a plan
OpenAI has a documented API activation path, but not all ChatGPT/Codex output can be called watermarked. The October 5, 2026 EU text-provenance announcement enables global API opt-in for selected models that day and describes introduction to eligible EU ChatGPT/Codex text over coming weeks. This October 6 article separates announcement and publication dates.
This affects API provenance configuration and EU understanding of generated text. OpenAI frames it as responding to EU AI Act provenance-marking requirements. API defaults off; initial EU scope is not a global default. Check product, region and settings without claiming all accounts completed rollout.
Read help-page present tense alongside rollout timing
The announcement covers eligible EU users across plans and ChatGPT/Codex text, without a universal completion date. Coming weeks does not establish reliable detection of all text, code or old content from announcement day. Cloud partners are also progressing, with availability dependent on partner and output type.
The help summary uses present tense for watermarked EU ChatGPT text while the October 5 announcement still says coming weeks. Read the help as mechanism and announcement as rollout timing. They do not establish completion; absence of a visible mark does not determine your account's coverage.
Check organization defaults and project settings
The organization path is Organization settings > Data controls > Text provenance. Enable Allow text watermarking, select models and Save. Verify the current organization first; another organization's setting is not your project's result.
For a project, open Project Settings > Text provenance, confirm project, enable, select models and save. The settings page shows currently supported models. The announcement says selected models, and help describes expansion over coming weeks; names, age or benchmark charts cannot establish the list.
Once enabled for a supported model, watermarking occurs during generation; customers do not tag every response separately. These are documented steps, not performed account changes or API tests. If absent, check help and actual settings rather than adding undocumented request parameters.
What textGrain changes
textGrain gently adjusts random choices of words or word pieces during generation, creating a statistical pattern the detector can seek. The signal is wording, not invisible characters, extra whitespace, unusual punctuation or watermark-only tokens. There is no icon to remove manually.
Copying unchanged wording may retain the signal, while substantial rewriting, paraphrase or translation weakens detection. OpenAI says assessed quality differences stayed within normal evaluation variation; that guarantees no identical outcome across applications, languages or outputs. Watermarking is neither fact-checking nor plagiarism determination.
Generating a watermark is not detector authorization
The detector is not public at initial announcement. Help/developer docs describe approved research, academic and other eligible organizations with case-by-case applications. API watermarking, an API account or submitting an application does not establish approval.
The application form covers higher image/audio limits and text-detection access. Text uses include provenance research, reliability and interpretation of results. Explain organization and actual research and await review; ordinary commercial approval, immediate access or fixed quotas are not promised.
Keep public verification within its content scope
Public verify and Content Provenance API primarily inspect supported images/audio for OpenAI signals. Developer docs separately require approval for text. Do not treat public image/audio entry points as arbitrary text detectors or transfer upload formats and limits to text.
Images may have Content Credentials metadata and images/audio SynthID watermarks; text uses textGrain. These signal types and preservation after editing differ. Select docs by content type rather than assuming provenance gives shared eligibility and workflows.
Short text, code and results have limits
Short text often lacks enough content for reliable detection. Code offers fewer reasonable alternative phrasings, as do precise mathematics and verbatim user quotations. Help says the relevant EU code of practice does not require watermarking outputs under 200 tokens or code snippets; that does not classify every code block as human.
False negatives and false positives are possible, with language and edits affecting results. Detection supports a corresponding OpenAI signal, not author identity, human contribution, truth or copyright ownership. No detection does not establish fully human authorship. Combine provenance, time, content type and edit history rather than one result.
Apply this to your workflow
API developers should decide whether to opt in, confirm organization/project, inspect supported models and record saved settings. EU ChatGPT/Codex users should watch official updates and retain creation/edit records, without declaring all historical output watermarked from unknown account rollout.
Organizations needing detection should check purpose and approval. Third-party classification is not OpenAI watermark detection. Machine-readable marks replace neither required visible labels nor other disclosures; assess applicable uses separately. Track supported models, EU/partner rollout and detector access. Text-detection prices, quotas and EU completion date remain unpublished.
Sources
Frequently Asked Questions
Are all EU ChatGPT and Codex accounts covered?
Not established. October 5 says rollout over coming weeks. Help's present tense does not prove completion; account/output scope needs further official evidence.
Is API watermarking on by default?
No. Opt in through organization or project Text provenance and select supported models shown there, not inferred from evaluations.
Does copied text contain hidden characters?
OpenAI says no. textGrain adjusts statistical wording without hidden characters or dedicated tokens; substantial edits and translation affect detection.
Does enabling it grant a public detector?
No. Eligible organizations need approved text-detection access. Public verify and image/audio API scope do not transfer to text.
Does no watermark detection prove human authorship?
No. Length, code, model/time, rewriting and translation affect detection. Positive signals also establish neither truth, copyright nor author identity.