Disclose, label and prove every AI output. One npm install.
Since 2 August 2026, AI apps serving the EU must tell users they're talking to AI and label AI-generated images, audio, video and text. Label50 implements the technical measures for any stack, and keeps the evidence.
Marking deadline for existing AI systems: 2 December 2026
import { withLabel50 } from "@label50/kit";
import OpenAI from "openai";
// Every image and text output is labelled and logged automatically
const openai = withLabel50(new OpenAI(), { app: "acme-studio" });
const img = await openai.images.generate({ model: "gpt-image-1", prompt: "a red bike" });
// → C2PA "AI-generated" manifest added, evidence logged, ready for your CDNPlanned API, shown for illustration. Join the waitlist for early access.
Why teams get stuck
Scattered building blocks
C2PA libraries, watermarks, signing certificates and detection APIs come from different vendors, each with gaps.
Labels get stripped
Screenshots, re-encoding and common CDN image optimisation remove metadata before users ever see the file.
Text has no standard
File-level standards don't attach to text responses, but generated text still counts as synthetic content.
No proof
When an enterprise customer asks "are you AI Act ready?", most teams have nothing to show.
How it works
1. Disclose
A React component and API that show "You're interacting with an AI system" at the first interaction, translated and accessible.
2. Label
C2PA labels on images, audio and video, an optional invisible watermark, and labels for text. Works with OpenAI, Anthropic, Vercel AI SDK and image models.
3. Prove
An evidence log of every labelled output, a CI check that labels survive your CDN, and a one-click PDF report for auditors and customers.
Built for small AI teams
| You ship… | Article 50 asks you to… | Label50 gives you… |
|---|---|---|
| A chatbot or AI agent | Disclose that it's AI at the first interaction | Disclosure component plus a server-side guard |
| AI images or video | Mark output as AI-generated in a machine-readable way | C2PA labels, watermark option, CDN survival check |
| AI voice or audio | Mark output in a machine-readable way | Audio labelling and evidence |
| AI-written text | Mark output in a machine-readable way | Text labels, response metadata, evidence log |
FAQ
Does this apply if we're not in the EU?
Yes. Article 50 applies when your AI system serves people in the EU, wherever your company is based.
What happens on 2 December 2026?
Generative AI systems already on the market before 2 August 2026 must add machine-readable marking to their outputs. Chatbot disclosure has applied since 2 August 2026.
Our model provider already adds labels. Isn't that enough?
Provider labels cover that provider's own output and can be stripped by your pipeline. Disclosure in your UI and evidence of what you labelled are still on you.
Does Label50 make us compliant?
Label50 implements the technical measures and keeps evidence. It isn't legal advice. Talk to a lawyer about your specific obligations.
Is it open source?
The core SDK will be open source. Paid plans add the hosted evidence log, reports and team features.
Get early access
The first 20 teams get 50% off for 12 months, plus hands-on help integrating before 2 December.
We'll only email you about Label50. Unsubscribe any time.