Skip to content
How to · 11 min read

How to label AI-generated content correctly in 2026

Article 50 of the EU AI Act says any AI-generated text, image, audio, or video must be labelled. Here is what counts, what does not, and how to do it on a real website.

The rule in one paragraph

Anyone who publishes AI-generated content to the public in the EU must mark that content so that users can recognise it as AI. The label has to be clear and timely. For images, a machine-readable provenance marker (C2PA metadata) is the de facto recommended addition, because most platforms now read C2PA tags automatically.

Two layers of labelling

Article 50 splits the responsibility in two.

  • Provider layer (machine-readable). The maker of the AI generator (OpenAI, Anthropic, Adobe, Midjourney) must embed a machine-readable marker in the output, in a way that platforms can detect.
  • Deployer layer (visible). You, the publisher, must show a visible label to readers when the content is displayed.

You are responsible for the visible label. The provider is responsible for the machine-readable marker. In practice, well-known generators already meet the provider obligation: DALL-E, Adobe Firefly, and Midjourney v6 embed C2PA metadata automatically. Some text generators experiment with watermarking but no production standard exists yet.

What needs a visible label

  • Blog posts or social posts that were written by an AI tool, even if a human edited lightly afterwards.
  • Product descriptions, listing copy, marketing emails authored by AI.
  • AI-generated photos, illustrations, mockups, infographics, banners.
  • AI-generated audio (podcasts, voice-overs, audio ads).
  • AI-generated video (explainers, social cuts, ads).
  • AI-translated content from one language to another (where the meaning was rebuilt, not just glossed).
  • AI-generated reviews or testimonials. These are also banned as misleading under consumer protection law.

What probably does not need a label

  • Content where AI was used as a research helper and the final wording is fully human.
  • Content where AI fixed spelling or grammar (Grammarly-style edit) on otherwise human writing.
  • Content where AI selected which human-written assets to show (recommendation engine output, like a Netflix carousel).
  • Internal documents not published to the public.

The line between "AI helped" and "AI wrote" is blurry. To stay safe, the rule of thumb most companies adopt is: if the first draft was AI, label it.

What the label should say

Short, neutral, factual. Three patterns that work.

  1. Above the title. A small badge or pill near the post title that says "AI-assisted" or "AI-generated".
  2. Author byline. The byline reads "Written by AI, reviewed by [Editor Name] on [Date]".
  3. Footnote. A line at the bottom of the post that says "Parts of this article were drafted by AI and reviewed by a human editor."

You can mix. A small badge near the title plus a longer note at the bottom is a common, audit-friendly choice.

How to add the label on your CMS

WordPress

Two options. First, use the ActHub inline label snippet inside the post content (paste in the visual editor). Second, register a custom field "ai_label" on your post type and render it in your template just below the post title. The site-wide ActHub script then auto-styles every label consistently.

Shopify

Add a metafield called "AI Label" on your product, populate it for AI-described items, and render the metafield in your product template. Use the ActHub label class to keep visual consistency.

Webflow / Wix / Squarespace

Drop the ActHub embed script in the site footer. Then add the inline label tag inside each AI-assisted post. The script styles all tags with one set of CSS.

The machine-readable layer (C2PA)

C2PA (Coalition for Content Provenance and Authenticity) is the format that big platforms have agreed to read. Meta, TikTok, LinkedIn, and Google's Search Generative Experience all check for C2PA metadata in uploaded images. If the metadata says the image was AI-generated, the platform automatically adds its own "AI info" label.

You do not have to handle C2PA yourself if you use a modern generator. DALL-E (since June 2024), Adobe Firefly (since launch), and Midjourney (since v6) embed it by default. If you use an older or self-hosted generator, you can add C2PA metadata as a separate step using free tools like c2patool or the Adobe Content Credentials browser extension.

Common mistakes that fail the requirement

  1. A watermark painted into the image pixels. Not valid as machine-readable disclosure under most regulator and platform guidance. It is removable. Use metadata instead.
  2. A site-wide footnote that says "Some content on this site is AI". Too generic. The Act expects the label to attach to the specific piece of content.
  3. Labelling only the obvious cases. If you label one AI image but skip another, the inconsistency itself is suspicious. Decide a rule and apply it everywhere.

Worked examples

Example 1. A bakery blog post. The bakery uses ChatGPT to draft a weekly post about a new seasonal cake. Their workflow: prompt ChatGPT, the marketing lead rewrites about 30 percent of the text, the post is published. The byline reads "Drafted with AI, edited by Maria on 28 June 2026". A small pill above the title says "AI-assisted".

Example 2. An e-commerce product photo. The shop uses Adobe Firefly to generate hero images of accessories on a clean background. C2PA metadata is embedded automatically by Firefly. The product page shows an "AI-generated image" label below the photo. The label is the only visible disclosure on the photo itself.

Example 3. An AI voice over for a YouTube ad. ElevenLabs generates the voice over. The video description on YouTube contains the line "Voice over generated with AI". YouTube's own "altered or synthetic content" tag is set to "yes" in the upload form.

Going further

Sources