Design
The AI label tells users, at first exposure, that they are interacting with an AI system or looking at content an AI generated or modified. It is based on the European Commission’s icons for labelling AI-generated content, adapted for localised copy, interface sizes, wrapping and accessibility. It is not a quality indicator, a feature badge or the place for accuracy copy, and it does not replace the machine-readable marking that sits with the model provider.
Anatomy
Section titled “Anatomy”- Label
- AI mark
- Text
When to use it
Section titled “When to use it”| Copy | What it discloses |
|---|---|
| You’re chatting with an AI | The user is interacting with an AI system |
| Generated | The content was entirely created by AI |
| Modified | Existing content was partially altered by AI |
| AI (icon only) | Same as Generated or Modified, when the full label does not fit |
Media in a component. Use Generated when the whole image or video was created by AI. Use Modified when existing content was partially altered by AI, for example a generated person or environment added to a real photograph. An accurate CAD digital twin of a vehicle is not AI content, and standard retouching or colour grading does not trigger the label.
Generated text. Use Generated when all three are true:
- The surface returns synthesised, model-generated text to the user.
- No person reviews that output before it is shown.
- The output is presented as an answer, not as a verbatim quotation of a source page.
The search field also carries “You’re chatting with an AI”, at or before the user’s first interaction, and keeps it for as long as users can keep asking. When an answer is on screen, both labels appear: one on the field discloses the interaction, one heading the answer discloses the content.
Don’t use it when:
- The surface returns retrieved content unchanged, such as a link list or a snippet from a published page. Retrieval is not generation.
- Content was drafted with AI but reviewed and signed off by a named person. The human review exception applies and the label would be misleading.
- Only part of the page is generated. The label marks the generated element, not the page.
Properties
Section titled “Properties”Colour
Black (default) or white, chosen by contrast and never as a stylistic choice: a black label on light areas, a white label on dark areas. The label must reach at least 3:1 against the area right behind it, measured against that area and not against an average of the image. On media, check it per image. On generated text, decide once per surface and document it.
Size
Large (default) and small. The text is never smaller than 14px, on every breakpoint, and scales with the user’s text size settings. The label does not scale with the image it sits on: a small card and a full-bleed hero carry the same size.
Icon only
The “AI” mark alone, for media where the full label does not fit. It stands in for Generated or Modified only. “You’re chatting with an AI” never goes icon-only, because the mark alone does not tell users they are interacting with an AI system.
Wrapping
Long strings wrap to a second line rather than stretching or truncating the label. The two-line state is designed, not an exception. In right-to-left languages the label mirrors, with “AI” on the right.
Best practices
Section titled “Best practices”Do
Use the same “AI” mark everywhere, pick black or white by contrast, and attach each label to the element it discloses.
Don't
Don’t hide the label, shorten its text, change its colour or size, animate it, or make it the smallest legal text on the page.
Content guidelines
Section titled “Content guidelines”- “AI” stays untranslated in every market and language. The text next to it is localised.
- Use the approved strings. They are translated and versioned centrally; markets cannot edit them. The current strings are in the Figma translation table.
- Sentence case in every variant. It reads more easily than capitals, takes less room in longer languages and makes “AI” stand out.
- Full words over acronyms, no brand voice, no first person, no exclamation.
- Never cut text. A shorter approved string is better than truncation. No new translation may be longer than the longest approved one, so the label always fits in two lines.