Watermarked AI text: what ChatGPT's marking changes for a Swiss SME

Since 5 October, OpenAI has been adding an invisible watermark to text produced by ChatGPT and Codex in the European Union. For a Swiss SME that drafts proposals, reports and emails with AI, the question is no longer whether it will show, but how to own it.

What OpenAI announced on 5 October

On 5 October 2026, OpenAI published a twenty-page technical report on textGrain, its text watermarking method, designed with researchers from the University of Pennsylvania and Yale, as reported by TechCrunch.

The principle is subtle: the model slightly steers its word choices according to a secret key. Readers see nothing, but a detector that knows the key finds the statistical pattern, even after a copy and paste.

The rollout comes in three parts:

• ChatGPT and Codex: eligible text is marked for users located in the EU, "over the coming weeks", on every plan.

• API: available worldwide since 5 October, but off by default, through the "Text provenance" setting in organisation settings.

• Detector: access is restricted, for now, to approved researchers and expert organisations, including ETH Zurich and Cornell.

The driver is regulatory. Since 2 August 2026, Article 50 of the AI Act has required AI-generated content to be marked in a machine-readable format.

The digital omnibus regulation gives systems already on the market until 2 December 2026 to comply. OpenAI is not alone: in August, Anthropic announced it would watermark Claude's text worldwide.

Why a Swiss SME is affected, even outside the EU

Switzerland is not named in the scope announced for ChatGPT. One could conclude that the topic does not concern us. That would be a misreading, for three reasons.

• Your cross-border or travelling staff use ChatGPT from France or Germany, and their text will be marked.

• Tools from other providers, such as Claude, already watermark regardless of country.

• Your European clients apply the AI Act and will expect compatible practices from their suppliers.

On the Swiss side, the Federal Council decided in February 2025 to follow the Council of Europe Convention on AI, with targeted adjustments to existing law. The framework remains lighter, but your European clients' expectations apply today.

In other words, watermarking becomes a commercial reality before it becomes a legal obligation for you.

What the watermark proves, and what it does not

OpenAI is remarkably cautious about the limits of its own tool. This is the part of the announcement we recommend reading first, before drawing any hasty conclusion.

According to OpenAI's technical report, on an English passage of about 400 tokens, roughly 300 words, the detector recognises marked text in about 95% of cases, at a 1% false positive rate. At 200 tokens, it drops to 80%.

Robustness falls quickly once the text is edited. Replacing 10% of words with synonyms cuts detection from 92% to 66%. With a quarter of the words replaced, it falls to 17%.

Other EU languages are less well covered: between 42% for Romanian and 69% for Spanish in the first tests. OpenAI has since strengthened the signal for the weakest languages.

Above all, OpenAI states it plainly: the absence of a watermark does not prove that a human wrote the text. Its presence says nothing either about the share of human work, responsibility or the accuracy of the content.

For an SME, this has two direct consequences:

• Do not rely on rewording to "erase" AI: it is a poor strategy, and often an admission in itself.

• Never use a detector as evidence against a candidate, an employee or a supplier: false positives exist and the signal remains probabilistic.

An example: the fiduciary firm that drafts with AI

Take a fifteen-person fiduciary firm in Lausanne. Its staff use ChatGPT to prepare year-end closing comments and recommendation letters for their clients, several of which are French subsidiaries.

Two employees live in Haute-Savoie and work from home two days a week. Their text will be watermarked. Their colleagues in Vaud will not be, using the exact same tool.

The risk is not the watermark itself. The risk is that a client discovers the use of AI without having been told, and starts wondering what they are really paying for. Trust, in a fiduciary relationship, is the product itself.

We recommend three simple measures for this kind of firm, all fitting on a single page:

• An AI usage policy listing authorised tools, prohibited data and the level of review required.

• A clear statement in the engagement terms: AI assists the drafting, a professional validates and signs.

• A single company account rather than scattered personal accounts, to keep control over settings and data.

Cost of the exercise: about two half-days of work. Benefit: a potentially awkward topic becomes an argument for transparency with clients.

What we take away

Text watermarking does not change what AI can do. It changes what can be hidden. For an SME, that is rather good news: it pushes firms to formalise practices that should have been in place long ago.

Our reading comes down to three points:

• Own your use of AI and explain how you govern it, rather than trying to hide it.

• Place your value in review, expertise and human sign-off, not in the volume of text produced.

• Treat detectors as clues, never as proof.

If you want to know where your company stands on these topics, our AI Barometer gives you an initial diagnosis in a few minutes.

To build your usage policy or review your engagement terms, let's talk.