Guides · 2026-10-11
AI Content Labeling: C2PA, Watermarks, Visible Labels
Every AI disclosure regime — the EU AI Act, Google Ads, China's labeling measures — demands roughly the same two things: a marking machines can read and a disclosure humans can see. But the technologies behind those markings differ wildly in strength, durability and cost. Here's an honest comparison.
The two layers, and why you need both
Regulators split disclosure into machine-readable marking (metadata, watermarks, cryptographic manifests embedded in the file) and visible disclosure (a label the viewer actually sees). The machine layer lets platforms and regulators detect synthetic content at scale; the human layer protects the viewer from deception. One without the other fails most rule sets — the EU AI Act's Article 50 explicitly requires both.
C2PA content credentials: the gold standard
The Coalition for Content Provenance and Authenticity (C2PA) standard embeds a cryptographic manifest in the file — a signed claim describing who created the asset, with what tool, and what was done to it. It's tamper-evident: edit the pixels and the signature breaks. Adobe, Microsoft, Meta and major camera makers support it, and regulators name-check it as the preferred provenance form.
The catch: C2PA only works if your entire pipeline preserves it. Export from an editor that doesn't write C2PA, upload to a platform that strips manifests on recompression, and the credential is gone. It's the strongest provenance available — and the most fragile in real publishing workflows.
Invisible watermarks: durable but proprietary
AI generators increasingly embed invisible watermarks (Google's SynthID is the best-known) — signals hidden in the pixels or audio that survive cropping, compression and screenshots far better than metadata. Detection usually requires the vendor's own tooling, which makes watermarks a vendor-locked layer: great for proving your own content is yours, weak as an industry-wide standard.
XMP and EXIF: the minimum viable marking
XMP packets and EXIF fields are the oldest, simplest machine-readable layer — plain metadata fields like "generated by" or generator identifiers. Every image tool reads them; nearly every social platform strips them. Treat XMP/EXIF as a baseline, not a strategy: cheap to add, trivially lost.
Visible disclosure: the layer humans actually see
Visible disclosure is the label on the content itself — "AI-generated" overlays, "Created with AI" captions, platform-native tags (TikTok's and Meta's AI labels, YouTube's altered-content disclosure). Rules converge on one principle: disclosure at first exposure. A viewer should know content is AI-generated before or as they consume it, not after hunting through a description.
Placement matters more than wording. An "AI-generated" badge burned into the first three seconds of a video beats a perfect caption nobody opens. For ads, Google's rules require the disclosure in the ad itself for synthetic content that could mislead — check the platform's current policy text, because these rules move.
What survives publishing: a durability ranking
From most to least likely to survive a real publish pipeline:
- Burned-in visible labels — part of the pixels; nothing strips them short of cropping.
- Invisible watermarks — survive compression and screenshots; need vendor tools to verify.
- C2PA manifests — strongest proof, but stripped by many platforms on upload.
- XMP/EXIF metadata — stripped by almost every social platform.
The practical lesson: never rely on a single layer. Ship visible labels for the human, C2PA or watermarks for the machine, and verify after publishing that at least the visible layer survived.
A labeling workflow that actually works
For each asset: (1) generate with tools that emit provenance; (2) add a visible label at first exposure; (3) export preserving C2PA where supported; (4) record what you shipped; (5) re-fetch the published URL and check what survived. Step 5 is the one everyone skips — and it's the one that catches Instagram recompression eating your manifest. DisclosureGuard automates the verification: deterministic provenance reads, per-asset compliance scores, and drift monitoring that diffs published files against your baseline daily.
Informational explainer, not legal advice. Check a specific asset against the current rules with a free DisclosureGuard scan.