A mark is coming before a public checker

OpenAI announced text watermarking plans on October 5. Eligible ChatGPT and Codex output in the European Union will gain the mark over the coming weeks, the company says. API customers worldwide can opt in for selected models, with watermarking off by default.

The detector has a different front door. OpenAI is accepting applications from researchers and expert organisations, with initial access granted case by case. It is not releasing a public text checker at launch. That leaves ordinary readers without a tool they can use to check this particular signal themselves.

The signal sits in choices, not hidden characters

The technique is called textGrain. Its technical report describes controlled changes to the probabilities used when a model selects its next token, the small unit of text it generates. A detector with the relevant secret key looks for a pattern across a passage. The method has to balance a detectable pattern against the freedom needed to produce useful text.

OpenAI's help page says the mark does not depend on invisible characters, extra spaces or odd punctuation. Coverage also varies with the product, model and when the output was created. It is not a universal test for anything written by AI.

In one company evaluation of 400-token English passages, replacing a tenth of the words with synonyms reduced detection from roughly 92% to 66%. That is a reported test result, not a promise about arbitrary documents. How well the system holds up across everyday editing remains an important question.

Origin and authorship are different questions

OpenAI's documentation says a positive result cannot establish ownership, identify the user or measure how much a person contributed. Nor does it verify that the text is accurate. A negative result cannot prove a human wrote it: unsupported models and altered or unsuitable samples can escape detection.

The European Commission's transparency code treats machine-readable marking and visible labelling as separate parts of its framework. Finding a statistical signal is not the same as explaining the editorial responsibility behind a published piece.

For anyone assessing a disputed passage, the sensible role of this tool is narrower than a verdict. It may add evidence about a supported model's involvement. It cannot replace checking the sources, the editing process and the person accountable for the final text.

Sources

  1. OpenAI: Our approach to EU text provenance rules, October 5Primary dated announcement of staged EU rollout, global API opt-in, restricted detector access and company-reported editing evaluation.
  2. OpenAI Help: Provenance signals in generated contentPrimary documentation on coverage, absence of hidden characters and limits of authorship, identity and accuracy conclusions.
  3. OpenAI and co-authors: textGrain technical reportPrimary technical report describing entropy-calibrated token sampling and keyed detection; not presented as independent validation.
  4. European Commission: AI-generated content transparency codePrimary context distinguishing providers' machine-readable marking from deployers' labelling. No compliance deadline or legal conclusion asserted.