AI content icons are a workflow problem for SMEs

The EU has now given companies a visible way to say that content was made or changed with AI. That is useful. It is also easy to misunderstand. An icon is not a policy, and a label pasted at the end of a publishing process will not fix a workflow that never decided when disclosure is required.

On 10 June 2026, the Commission published the final Code of Practice on transparency of AI-generated content. The page ties the code to Art. 50(2), Art. 50(4) and Art. 50(5) of the AI Act, with the transparency obligations applicable from 2 August 2026. It also points to a new set of EU icons that deployers may use when labelling AI-generated content.

The icon solves only the visible layer

The icon set is the least abstract part of the transparency package. It gives a recognisable label for content generated or manipulated by AI. For an SME, that is better than inventing a disclosure mark in a hurry.

But Article 50 is not just a design problem. It separates provider duties from deployer duties, machine-readable marking from visible labelling, and public-interest publishing from ordinary internal drafting. The icon sits near the end of that chain. The decision to use it has to sit earlier.

Question	Why it belongs before publication
Is the content audio, image, video or text?	Article 50 treats content types differently.
Is it generated, substantially altered or merely edited?	Minor editing is not the same as synthetic content.
Is the publication informing the public on a matter of public interest?	That is the text trigger in Art. 50(4).
Has a human reviewed it under editorial responsibility?	The text exception depends on real review and accountability.
Is the mark visible, accessible and timely?	Art. 50(5) requires disclosure at first exposure or interaction.

Source: Regulation (EU) 2024/1689, Article 50; European Commission Code of Practice and draft Article 50 guidelines. Last verified 2026-06-30.

Machine-readable marking is a vendor file

The Code of Practice makes one distinction worth preserving inside the company. Providers of generative AI systems are expected to mark outputs in a machine-readable format so they are detectable as artificially generated or manipulated. That is not the same job as a deployer deciding whether to show a visible label on a web page, report, advert or customer message.

Most SMEs should not try to build their own provenance system. They should keep a supplier file. For each tool used to generate public content, record what the vendor says about metadata, watermarks, export settings and detection. If the answer is vague, that is procurement information, not a reason to invent a local technical standard.

The visible disclosure rule then stays short: when the business publishes AI-generated or materially manipulated content that a normal reader, viewer or listener may rely on, the label appears where that person sees it before acting on the content.

Human review needs a record

The tempting shortcut is to say that everything passes through a person, so nothing needs labelling. That is too loose. Article 50(4) refers to text that has undergone human review or editorial control and where a natural or legal person holds editorial responsibility.

For a small firm, the record can be boring:

• the channel, such as website news, product page, customer email or social post;
• the owner who approves publication;
• whether AI generated a first draft, translated, summarised or materially rewrote the text;
• whether the piece informs the public on a matter of public interest;
• the label decision, including “no label” when editorial responsibility is the reason.

That five-line record is more useful than a full policy document nobody opens. It lets the company show that it made the disclosure decision before publication, not after a complaint.

The August deadline is operational

The Article 50 date matters because the work is visible. A high-risk classification debate may stay inside legal and procurement files for months. A missing AI label is on the page, in the video, in the chatbot or in the PDF.

The smallest useful SME rule is therefore a publishing gate: no public AI-assisted content goes live until the owner has answered the label question. If the content is an image, audio or video, check the vendor marking file and decide on the visible label. If it is text on a public-interest matter, record whether human editorial responsibility applies. If it is a chatbot or voice interface, make the AI interaction notice visible at first use.

The EU icons are helpful because they lower the friction of disclosure. They do not lower the need to decide. Treat the icon as the final mark on a workflow, not the workflow itself.