AI watermarks in texts: what changes for Italian companies

Your marketing office uses ChatGPT every day to write copy. Starting in August 2026, those texts must be labeled by law — and the fine for non-compliance can reach up to 15 million euros.

Is it worth adapting immediately or waiting for the rule to become operational?

The rule is already operational as of August 2, 2026. Those who started using AI tools before that date have until December 2, 2026 to get in compliance, but waiting is a concrete risk.

Article 50 of EU Regulation 2024/1689 — the so-called AI Act — entered into force on August 2, 2026. From that day on, anyone who produces content with generative artificial intelligence tools and distributes it in the European market must mark it in a machine-readable format. The European Commission confirmed that systems already in use before August have a transition period until December 2, 2026.

For Italian companies, the message is simple: the time for waiting is over. Every week of delay is a week of exposure to risk.

What is the invisible watermark and how does it really work?

An invisible watermark is a statistical signal hidden in the text, not visible to the naked eye, which an automated detector can identify to figure out if the content was generated by a machine.

OpenAI has developed a system called textGrain : instead of adding a visible signature, the model chooses words a bit differently than usual, creating a recognizable statistical pattern. It is like a hidden code in the word sequence.

The detector looks for this pattern. With texts of at least 400 tokens (about 300 words in Italian), the system finds the watermark in 95% of cases, with a false positive rate — meaning human texts mistaken for AI — of 1%. On shorter texts, around 150 words, accuracy drops to 80%.

The critical point: if someone replaces 25% of the words with synonyms, the detection rate plummets from 92% to 17%. The watermark is robust for normal use, but it is not unbreakable.

The detector is not public: at the moment it is only accessible to researchers and expert organizations approved by OpenAI. For companies, this means they won't be able to independently verify whether their texts are watermarked — at least for now.

Who has already introduced the watermark and who hasn't?

OpenAI and Anthropic have made their moves, but with different approaches. The difference matters for those using multiple AI tools in the company.

Anthropic activated watermarking on all Claude models starting August 2, 2026, across all platforms and globally — including instances hosted on AWS, Google Cloud, and Microsoft Foundry. It did this because, at launch, it lacked a reliable way to restrict it geographically, as reported by pasqualepillitteri.it .

OpenAI has chosen a more gradual approach: textGrain will reach ChatGPT users in the EU in the coming weeks. Developers using the APIs (the programming interfaces to integrate ChatGPT into their own systems) can already enable it voluntarily, but it is disabled by default.

Around 190 organizations had signed the Code of Practice — the European voluntary code of conduct on AI content transparency — by the end of July 2026, including Anthropic, Google, Meta, Microsoft, and OpenAI, as reported by dev.to . Signing the code is not mandatory, but those who do get a presumption of conformity before supervisory authorities.

What changes in practice for those who produce content within the company?

The marking obligation falls on those who publish the content, not on the software that generates it. This means that Italian agencies, marketing departments, and creative studios are directly responsible.

The AI Act defines deployer — meaning anyone who deploys an AI system — anyone who produces content with generative AI tools and distributes it. Communication agencies, marketing departments, creative studios: everyone falls into this category, as clarified by aipia.it .

The point that surprises many: if an agency produces AI texts for a client and the client publishes them without the correct markings, the risk of non-compliance falls on the publisher. This changes how contracts with content suppliers must be written.

Three concrete scenarios: how does all this translate for an Italian company?

The impact varies a lot depending on how the company uses AI today. Here are three typical situations with concrete actions to take.

Illustrative examples: the numbers and situations are meant to show the reasoning, they are not real customer data.

What are the real sanctions and risks for those who don't comply?

Fines for violating Article 50 of the AI Act can reach up to 15 million euros or 3% of global annual turnover, whichever is higher.

The figure is the one published by CityNext based on the text of the regulation. For an SME with 10 million in revenue, 3% means 300,000 euros. For a medium-sized company with 100 million, it reaches 3 million.

The risk is not only financial. A company that publishes unmarked AI content can be reported by competitors, journalists, or consumer associations to national supervisory authorities. Reputation is at stake just as much as the balance sheet.

There is also a contractual risk: if a client discovers that the content received from the agency was non-compliant, they can dispute the payment or claim damages. This applies in both directions of the supplier-client relationship.

Why are we talking about it now and not for months?

Because until August 2026 it was theory. Now it is applicable law, with concrete deadlines and the first technical tools available.

The AI Act was approved in 2024, but its obligations have come into force gradually. The transparency obligations of Article 50 became applicable on August 2, 2026. Before that date, companies could wait. Now they no longer can.

The topic blew up in October 2026 because OpenAI announced the rollout of textGrain for ChatGPT users in the EU — the first major provider to implement the watermark directly in the consumer interface, not just via API. This made the issue real for millions of users who use ChatGPT every day without messing with code or technical integrations.

The topic of EU digital regulation is also broader: as we saw with the fine for Meta for the DSA , European authorities are becoming increasingly active in enforcing new digital rules. Waiting is no longer a strategy.

What tools and paths are available to you to comply?

There are three levels of intervention: technical (the AI provider's watermark), visual (official EU icons), and contractual (updating agreements with suppliers and clients).

The European Commission has adopted three official icons that deployers can use where visual disclosure is possible: one for entirely AI-generated content, one for AI-modified content, and a base one with an interactive second level. These are practical tools, not just symbolic ones.

On the technical side, the provider's invisible watermark (OpenAI's textGrain, Anthropic's SynthID-Text) does not replace visual disclosure: it complements it. A text can have both, or just one. The regulation requires that content be machine-readable — the technical watermark fulfills this requirement.

For those who use ChatGPT via API in their own systems — for example a CMS that automatically generates product descriptions — activating the watermark is a technical operation in the API settings. It is disabled by default: it must be explicitly enabled. This is a point that many technical teams have not yet checked.

How to get started: concrete steps to comply?

Five sequential actions are enough to cover the main risks by the December 2026 deadline.

1. Map AI content in use. Make a list of all texts, images, and audio produced with AI tools that are published or distributed. Include those produced by external agencies.
2. Check the settings of your AI tool. If you use ChatGPT via API, check if the watermark is active. If you use Claude, it is already active by default. If you use other tools, check their stance on AI Act compliance.
3. Add the official EU icons. Wherever you publish AI content — website, newsletter, socials, PDF — add the appropriate icon. The European Commission provides the official icons.
4. Update contracts with suppliers and customers. Include a clause specifying who is responsible for labeling when AI content changes hands. Do not leave this point ambiguous.
5. Consider signing the Code of Practice. If you systematically produce AI content, signing the European voluntary code gives you a presumption of conformity. Check the official page of the Commission for details.

What are the most frequent mistakes companies are making?

The three most common mistakes are waiting for the AI supplier to fix everything, failing to update contracts with agencies, and ignoring AI content produced by third parties but published by the company.

The first mistake is thinking that OpenAI's or Anthropic's technical watermark is enough. The invisible watermark satisfies the machine-readable labeling requirement, but it does not replace visual disclosure to the end user, which is required separately by the rule.

The second mistake is not updating contracts with agencies. If an agency produces AI content for you and delivers it without watermarking, you — as the publisher — are the one exposed to the penalty. This also applies to the content of the European retail using AI for product descriptions .

The third mistake is treating the watermark as proof of authorship. OpenAI is explicit: a missing watermark does not prove the text was written by a human, and a present watermark doesn't tell who wrote what. Using it as a criterion in employment contracts or editorial policies is wrong.

A less obvious mistake involves short texts. The rule doesn't require watermarking for texts under 200 tokens — about 150 words. But many companies produce longer texts that get cut down for publication. The watermark gets weaker with edits: a 400-token text with 25% of the words replaced only has a 17% chance of being detected.

How to understand if the adaptation process is working?

There are concrete indicators you can track without advanced technical tools, at least in the initial phase.

So, is it worth adapting immediately or waiting until December?

It is worth it if you already have an active AI content production workflow, publish texts on websites, newsletters, or social media, or receive AI content from external agencies. Every week of delay is direct exposure to risk.

A middle ground you use AI sporadically and the content never goes public. You have until December, but it's best to use this time to set up a process rather than put it off.

Not yet if you use AI only for internal use — drafts, notes, analysis — and no output is ever distributed. The labeling obligation does not apply, but keep an eye on how the rule evolves.

How we see it at SHM

For us, the AI watermark is not a technical problem: it is a process problem. The technology — textGrain, SynthID, EU icons — is already available. What is missing in most Italian companies is a clear workflow that knows where the machine's contribution ends and the human's begins.

We believe AI should be integrated to free companies from routine tasks, not to create new legal risks. A well-built process — featuring watermarking, disclosure, and updated contracts — turns a regulatory obligation into a competitive advantage: it shows customers that the company uses AI responsibly.

We work with Italian companies to build exactly this: AI integration strategies that comply with the regulations and produce real value. If you are evaluating how to structure your AI content workflows, or how to train your team on the obligations of the AI Act, our service of AI training for businesses is the fastest starting point. For those managing content campaigns at scale, the topic also ties into the SEO visibility in the AI era .

You can find more insights on EU digital regulations and technological process management in our cluster Project management and technology .

Frequently Asked Questions

Watermarking is not the end of creative freedom with AI: it is the beginning of a mature and responsible use of the tools you already use every day.