Fingerprint & verify any document — free, no upload Try it now →

Home / Answers / Synthetic-content marking evidence (Art. 50(2))

EU AI Act · Article 50(2) · Synthetic content

How do I mark AI-generated content under Article 50(2) — and prove the marking held?

Under Article 50(2), providers of AI systems that generate synthetic audio, image, video, or text must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Marking is the obligation; the evidence problem is proving the marking was applied at generation and that a specific output carried it — because marks can be stripped, and claims can be disputed. The durable answer is to bind the output’s content hash, the marking mechanism, and the generation context into a tamper-evident receipt anchored to the public chain. LedgerProof produces this — it complements watermarking and C2PA (the mark is the label; the receipt is the independent proof it existed at that time). It produces evidence, not a compliance verdict.

What Article 50(2) requires

Article 50(2) places the obligation on providers of generative AI systems: outputs (synthetic audio, image, video, or text) must be marked in a machine-readable format and be detectable as artificially generated or manipulated. The technical solutions must be effective, interoperable, robust and reliable as far as is technically feasible. These obligations apply from 2 August 2026.

Watermarks, metadata, and C2PA manifests are ways to apply the mark. None of them, on their own, prove to a third party that a given output was marked at the moment it was generated — which is what a dispute or an audit turns on.

The evidence gap: a mark can be stripped or contested

Metadata gets removed on re-encode. Watermarks get cropped. When a specific piece of content is challenged — “was this actually marked, or added later?” — you need an independent record that the marking existed at generation. A tamper-evident receipt closes that gap:

PropertyScreenshot / self-written logTamper-evident receipt
Independently checkableNo — requires trusting youYes — anyone can verify
Tamper-evidentNo — editable in placeYes — any change is detectable
Timestamp you can't backdateNo — system clockYes — cryptographically anchored

How to produce the evidence

  1. Capture the event. At the moment the disclosure, marking, or label is applied, record the exact text or marking, the output's content hash, and the context.
  2. Fingerprint it. The SDK computes the SHA-256 fingerprint locally — the content never leaves your systems (hash-only, GDPR- and confidentiality-clean).
  3. Anchor it. Fingerprints are folded into daily RFC-9162 Merkle roots and anchored to the public chain via a public-ledger transaction.
  4. Verify it — anytime, by anyone. Re-check the receipt against the public chain with the open-source verifier, no account required.

Because the receipt format is an open IETF Internet-Draft implementing the SCITT architecture, the evidence you hold today keeps verifying even if any single vendor disappears.

Anchoring doesn’t replace your marking mechanism — it makes it provable. The watermark or C2PA manifest is the label; the anchored receipt is the independent proof that this exact output carried that label at that time. See the general Article 50 evidence guide for how the obligations fit together.

What this does and doesn't do. LedgerProof produces independently verifiable evidence that a disclosure, marking, or labeling event existed, unaltered, at a point in time. It does not make you “compliant,” does not authenticate authorship, and is not a guarantee of court-admissibility — your auditor, regulator, or counsel determines compliance and admissibility. Proofs are tamper-evident, not tamper-proof. This is not legal advice.

Frequently asked questions

Does anchoring a receipt replace watermarking or C2PA?

No — it complements them. Watermarking and C2PA are how you apply the machine-readable mark Article 50(2) asks for. An anchored receipt is the independent proof that a specific output carried that mark at generation time, which survives even if the watermark is later stripped or the claim is disputed.

What counts as machine-readable marking under 50(2)?

The Act calls for outputs to be marked in a machine-readable format and detectable as artificially generated or manipulated, with solutions that are effective, interoperable, robust and reliable as far as technically feasible — for example embedded watermarks, cryptographic provenance (C2PA), or metadata. LedgerProof does not judge which mechanism satisfies the Act; it produces verifiable evidence that whichever mechanism you used was applied.

Does this send the generated media anywhere?

No. LedgerProof anchors only the SHA-256 fingerprint of the output and the marking record. The media itself never leaves your systems — hash-only, GDPR- and confidentiality-clean.

When does the obligation apply?

Article 50 transparency obligations apply from 2 August 2026. Anchoring evidence as you generate content from now on means the proof already exists when a specific output is later challenged.

See LedgerProof for Article 50 → Verify a real receipt yourself
Primary sources: EU AI Act (Regulation 2024/1689), Article 50 — EUR-Lex · IETF Internet-Draft: SCITT receipts for AI Article 50 · Open verifier

Last updated 1 August 2026 · LedgerProof