Regulators want people to know when text came from an AI system. The hard part is marking it without wrecking readability—and without pretending a detector can prove everything.
On 5 October 2026, OpenAI described how it will implement text provenance under EU rules using textGrain: an invisible watermark that biases token sampling so a detector can recognise OpenAI-generated text. [1]
How the rollout is supposed to work
In plain terms, a “watermark” here is not a visible stamp. As the model picks each next word (token), the sampler is nudged in a patterned way. A matching detector can later look for that statistical fingerprint.
According to OpenAI’s announcement and a community/staff summary: API customers worldwide can opt in for select models; over coming weeks, eligible ChatGPT and Codex text generated in the EU will carry the watermark; the API stays off by default and is not a global default at launch. [1][2]
Detector access will initially be limited to approved researchers and expert organisations, because short passages and edits weaken the signal. [1][2]
What a watermark does—and does not—mean
OpenAI says the mark indicates OpenAI generation or processing of part of a text. It does not identify the user, ownership or accuracy. Absence of a watermark does not prove human authorship. [1]
TechCrunch covered the same EU-focused text-watermarking plan on 5 October 2026. [4]
A technical report PDF on textGrain, also dated 5 October 2026, accompanies the policy post. [3]
Why it matters
This is a concrete compliance move for EU transparency rules and a rare production text-watermark rollout from a major provider. It also comes with clear limits: detectors will miss short or heavily edited text, and a clean human-looking passage without a mark is not proof it was written by a person.
What we don’t know
- How reliable detection remains after paraphrase or translation is limited by design; OpenAI itself flags short text and edits as weak spots.
- Broader Help Center provenance material has historically emphasised images and audio (SynthID/C2PA) and only a goal of expanding to text; the dedicated 5 October EU provenance post and community summary are the clearer launch sources for this text rollout.
- How widely approved researchers will share detector results publicly is not settled in the materials cited here.
The Bottom Line
OpenAI is starting to watermark eligible ChatGPT and Codex text generated in the EU with textGrain, while leaving the API opt-in and off by default. Treat a positive detection as a clue that OpenAI touched the text—not as a full audit of who wrote what, or whether it is true.
Sources
- OpenAI, “Our approach to EU text provenance rules,” 5 October 2026 — https://openai.com/index/eu-text-provenance
- OpenAI Developer Community summary, 5 October 2026 — https://community.openai.com/t/openais-approach-to-eu-text-provenance-rules/1403521
- OpenAI, textGrain technical report PDF, 5 October 2026 — https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf
- TechCrunch, 5 October 2026 — https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/
