An OpenAI text watermark can help investigate a text's origin. Your organisation still needs to assess the transparency requirements that apply to its publication. A technical signal cannot make the decision about visible disclosure or human editorial review for you.
On 5 October 2026, OpenAI announced a phased EU rollout for eligible ChatGPT and Codex text, worldwide opt-in for selected API models, and restricted detector access. API watermarking is off by default. See OpenAI's EU text provenance announcement. This article examines the implications for publication workflows, as of 5 October 2026.
How does textGrain work?
During generation, textGrain connects token selection to randomness derived from a key. A detector looks for the resulting statistical signal. The method balances detectability against the variety of responses a model can produce. The technical report by OpenAI and co-authors describes a configurable budget for the loss of sampling variation.
OpenAI's product guidance says this does not add hidden characters or extra watermark-only tokens. Rewriting and translation can weaken detection. A generic AI detector classifying writing patterns after generation examines something different from an embedded watermark signal.
For procurement, ask: which detector examines which signal, in which outputs? A supplier displaying an 'AI-written' percentage has not, by that fact alone, demonstrated machine-readable marking.
Who carries the Article 50 obligations?
Article 50(2) assigns machine-readable marking to providers of generative AI systems. Technical solutions must be effective, interoperable, robust and reliable insofar as technically feasible. Exceptions include assistive standard editing or not substantially changing input or its meaning.
For publishing organisations, paragraph 4 addresses AI text informing the public on matters of public interest. The editorial exception requires human review or editorial control and an identified natural or legal person holding editorial responsibility. A watermark does not establish that exception.
The Commission's Article 50 guidelines help distinguish roles and applications. An application built on an API needs its own role assessment. Procurement should not assume that every provider obligation automatically remains with the model supplier.
Use the Article 50 text and decision tree to structure that assessment. Our article on AI text informing the public examines the publication question in more detail.
Which deadline applies to text watermarking?
Article 50 has applied since 2 August 2026, as confirmed in the Commission's transparency guidance.
Systems placed on the market before that date have a transition until 2 December 2026 for Article 50(2) only. The legal reference is Article 1, point 39(b), of Regulation (EU) 2026/1744, adding the transitional provision to the AI Act. This does not postpone the other transparency obligations.
Assess the particular system, its role and market introduction. One supplier's feature rollout date is not a universal legal deadline for your organisation.
A decision matrix for publication workflows
These are Praxikon's illustrative scenarios, not regulatory findings. The final column contains recommended working practices; the AI Act does not prescribe this dossier format.
| Situation | Decision to investigate | Recommended record |
|---|---|---|
| A communications adviser drafts internal text with AI | Who checks the text before external use? | Purpose, sources and responsible editor |
| A municipality automatically publishes information about a public scheme | Is this public-interest text, and how will disclosure work? | Publication example, audience and disclosure decision |
| An editorial team publishes after human review | Is reliance on the editorial exception supported? | Review steps, corrections and editorial responsibility |
| A software company builds a text generator on an API | Which provider role and marking solution apply to its system? | Role assessment, model version and supplier explanation |
| An employer investigates suspected AI use in a document | Which additional facts support a conclusion? | Provenance investigation, the person's response and uncertainties |
A useful process starts with substance. For municipal information, for example, check the competent authority, applicable scheme and exceptions. Detection status does not assess the quality of that legal review.
Why a detector score cannot make the final decision
A positive result needs context; a negative result does not exclude AI use. OpenAI's announcement acknowledges false positives and false negatives and explains that a watermark does not measure human contribution or factual accuracy.
Our recommendation: do not base an employment or publication decision solely on a detector score. Investigate document versions, applicable agreements and the person's account. Record what the result supports and what remains uncertain. This keeps the tool within an evidence-gathering role.
Distinguish your own testing from a supplier's claims. If you lack access to the relevant detector, do not report its performance on your texts as independently tested. Request relevant test documentation and state the verification gap explicitly.
Five questions for suppliers and editorial teams
- Which application are we using? Record the product, model, version, settings and publication channels. Reassess after a material change.
- Which obligation follows from our role? Have procurement, product and editorial teams assess the same application. Assign one decision owner.
- What has actually been verified? Separate supplier explanations, your own observations and untested assumptions. Ask about test scope and limitations.
- Who checks the substance? Establish who verifies facts, citations and conclusions, and who can correct an inaccurate publication.
- What does the reader see? Inspect a real published page. A recorded decision is insufficient if the intended disclosure disappears during publishing.
The Code of Practice on Transparency of AI-generated Content offers an implementation framework. Adherence is voluntary; the legal obligations remain. The Commission and AI Board have assessed the code as an adequate voluntary tool. That is not an individual approval of textGrain or your organisation.
The practical result should be an explainable publication decision: which application was assessed, why disclosure is or is not required, and which evidence supports the conclusion. Keep an example of the final publication alongside it. This makes the next change in tools, editorial practice or regulation easier to assess.
Preparation: written with AI assistance. The primary sources listed below were consulted on 5 October 2026. Scenarios and process recommendations are Praxikon analysis; no independent technical test of textGrain was performed.
Want to determine which obligation applies to your own use case? Walk through the Article 50 decision tree: per role and use case the paragraph, the steps from the guidelines and examples of good notices.
Frequently asked questions about text watermarking
Primary sources, consulted on 5 October 2026
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe