Also available in:

The debate around invisible markings in AI-generated texts has reached the online industry. Since early October 2026, OpenAI has been rolling out the “textGrain” system for users in the EU. It is designed to machine-readably mark ChatGPT and Codex outputs as AI-generated — invisible to the reader, but detectable by special software. For the adult industry, where thousands of description texts, stories and product descriptions are created every day, the question arises: What does this mean in concrete terms?

How the New Watermarks Work – and Whether Quality Suffers

textGrain does not change the visible text. It subtly influences the model’s word choice: when equally valid alternatives exist (for example “seductive” or “sensual”), a statistical pattern is applied. Over longer passages, an invisible signal is created. OpenAI itself emphasises that text quality does not suffer as a result. Benchmarks show no significant differences in performance, creativity or readability. Earlier tests in ChatGPT also revealed no changes in user feedback.

Similar methods are already used by Google (SynthID-Text in Gemini since 2024) and Anthropic (a variant of SynthID in Claude, worldwide since August 2026). All work on the same principle: statistical patterns in word choice rather than hidden characters or metadata. Heavy editing — rewriting, synonym replacement or translation — significantly weakens or destroys the signal. When 25 percent of the words are changed, OpenAI’s detection rate drops to around 17 percent.

Overview: Who Marks What? (as of October 2026)

The EU Rules: When Labelling Is Actually Mandatory

The EU AI Act has required transparency since August 2026. Providers such as OpenAI are obliged to apply machine-readable marking. For those who publish the texts (deployers), the labelling obligation only applies under narrow conditions: the text must be published, serve to inform the public, and concern matters of public interest (politics, health, economy, etc.).

The decisive exception is this: if the text has undergone human review or editorial control and a natural or legal person assumes editorial responsibility, the labelling obligation no longer applies. A simple spell-check is not enough — a substantive review is required.

Adult content generally does not fall under “public interest.” Neither detailed video descriptions nor stories or product texts on adult platforms meet the legal criteria. The deployer labelling obligation therefore practically does not apply here.

Impact on Operators of Adult Sites

Site operators who use AI for category descriptions, SEO texts, newsletters or internal content currently do not need to add visible “AI-generated” notices. The watermark itself is invisible and affects neither readability nor conversion rates.

However, operators should keep two points in mind: First, the origin of the texts. Anyone who exclusively takes unedited ChatGPT outputs from the EU risks detectors (currently still limited to researchers and authorities) finding the signal. Second, quality assurance. Google does not rank AI texts worse across the board — what counts is E-E-A-T and relevance for users. Thin, mass-generated texts without added value will be penalised, regardless of whether a watermark is present.

Anyone who thoroughly edits texts, adapts the style and enriches them with real user experiences is on the safe side — both legally and in terms of SEO.

What Creators Should Consider with Video Descriptions

Many creators use AI to produce long, detailed description texts for videos — often several hundred words with scene descriptions, keywords and calls to action. Little changes in day-to-day practice here: the watermark is invisible and does not disturb either viewers or platform algorithms.

Nevertheless, a conscious approach is worthwhile. Anyone who takes the generated text unchanged leaves a potentially detectable pattern with OpenAI texts from the EU. Anyone who heavily revises the draft, incorporates their own wording and preserves their personal style largely removes the signal and at the same time increases authenticity. It is precisely this authenticity that viewers and platforms reward in the long term.

In future, creators should pay attention to which models they use and whether they are working in the EU. Detectors are expected to become more accessible, and some platforms may introduce their own checks. A documented editorial process (own revision, final approval) provides additional protection.

Our Clear Recommendation: Grok for the Adult Sector

The editorial team expressly recommends the use of Grok (xAI) for the creation of adult content.

Unlike OpenAI, Anthropic and Google, xAI has not signed the voluntary EU Code of Practice and has so far not introduced any invisible text watermark. Grok texts therefore currently carry no comparable statistical signal. This gives operators and creators maximum flexibility — especially when they later heavily adapt texts or process them in several languages.

In addition there is the content suitability: Grok is less restrictive with adult topics and often delivers more direct, less “smoothed” formulations. For detailed scene descriptions, dialogues and atmospheric texts in the adult sector, this proves to be an advantage in practice. At the same time, text quality remains high and SEO suitability is given, provided the outputs are edited and given real added value.

Anyone who works professionally in the online adult market should therefore know the new marking systems, understand their limitations and make a conscious choice. With careful editorial post-processing and the targeted use of Grok, both legal risks and quality losses remain low — and the creative scope remains large.