Content · 14 August 2026

Since 2 August 2026, synthetic audio, image, video and text content must carry a machine-readable marking. Deepfakes used in a professional setting must in addition be labelled visibly. Anthropic has announced that Claude models, from that date, embed an invisible text watermark. OpenAI relies on SynthID (DeepMind) for image and audio. The C2PA standard sits in the same logic. For an SME that feeds its website, social networks and brochures, the question is no longer “should we say it”. It is “how to trace it without breaking the sales journey”.

What Europe actually requires

Article 50 of the AI Act imposes two distinct layers. The first is machine-readable marking: a signal that platforms, detection tools and, in time, browsers will be able to read. The second, for deepfakes in professional use, is the visible label: the human eye must understand that the image, voice or video is not a raw recording of reality.

Systems already placed on the market before 2 August 2026 have until 2 December 2026 for the machine-readable marking under Article 50(2). That timetable does not suspend user information on chatbots, nor the visible labelling of professional deepfakes. An agency, a hotel or a manufacturer that publishes a generated visual for a campaign today does not have four months of silence.

The AI Office may investigate providers of general-purpose models since 2 August 2026. About 190 organisations have signed a voluntary transparency code (list of 31 July), including Anthropic, Google, Meta, Microsoft, Mistral and OpenAI on the provider side. A code is not a law. It does not exempt a French SME from labelling what it publishes under its own name.

What this changes for an SME’s marketing and website

The risk is not theoretical. A product visual, a presentation voice-over, a room photo “staged”, a director’s portrait retouched by a model: as soon as the content can pass for real when it is not, the visible label is required in professional use. AI-generated text of public interest with little human review must also indicate the use of AI.

Conversely, product-sheet copy reviewed and signed off by a salesperson is not automatically “text of public interest”. Distinguish content that simulates reality (photo, voice, video, clone) from clearly editorial content, reviewed by your teams.

Anthropic’s invisible watermark, like SynthID, is a provider tool. Anthropic’s detector is not yet public. You cannot, today, “scan” your Claude archives and obtain an enforceable certificate. C2PA, for its part, attaches to origin metadata (who created it, with which tool, on which date). The two approaches complement each other. Neither replaces your deployer responsibility for what you put online.

Five concrete actions for an SME site and communications

• Inventory generated or heavily assisted content. Photos, banners, voice, short videos, avatars, texts published without substantial review. Classify them: simulation of reality / mere drafting aid.
• Decide a visible-label rule. For any image, voice or video that could be taken for a real recording, a clear mention next to the media (caption, credits, overlay). Not a hidden footer link.
• Require metadata from your suppliers. Agency, freelancer, internal tool: ask whether the chain exports C2PA or equivalent machine-readable marking. Write down what the supplier warrants, and what it does not (Anthropic’s detector not being public, nobody can promise you complete proof).
• Separate the “company truth” corpus from the generated corpus. Spec sheets, prices, photos of sites or real rooms: these items must remain identifiable. This is also the right moment for data cleaning with a view to RAG: an agent that cites an AI photo as proof of a real product is a trust incident, not a marketing detail.
• Train marketing and front desk. The person who posts on LinkedIn or sends an email with a generated attachment must know the rule. Article 4 (AI literacy) has applied to deployers since 2 February 2025.

Hospitality, construction, brochure sites: the cases that keep coming back

An AI agent for a hotel or B&B that sends an “idealised” room photo without saying so exposes the establishment to a classic claim (misleading advertising) on top of the AI Act strand. A tradesperson who generates a before/after of a façade for a quote is in the same case: professional use, simulation of reality, visible label.

Fines on the transparency strand go up to €15 million or 3% of worldwide turnover. For a VSE, the theoretical cap is not the daily issue. The daily issue is a customer who feels misled, a marketplace that takes the visual down, or an insurer asking questions after a dispute.

What Cybernecs can frame with you

We help VSEs and SMEs set a simple rule: which content is marked, where the label appears, how site agents stay honest. That sits in our AI solutions for VSEs and SMEs. For a journey review (site, chat, publications), go through the contact page.

The watermark of the large providers is a building block. It does not do your compliance for you. It is your name the customer reads.

Sources: Artificial Intelligence Regulation (AI Act), Touteleurope, L’Express, Anthropic and OpenAI / DeepMind communications (SynthID), C2PA documentation.