What it means
- Any image, video, audio or text your systems generate or materially alter with AI carries:
- a visible label (on-screen text, watermark, or a spoken prefix for audio), and
- embedded provenance (permanent metadata or a unique identifier showing your service made it), where technically feasible.
- Users must not be able to strip or hide the label or metadata.
- Platforms that host user uploads ask uploaders to declare synthetic content and show a label before it goes live.
- ISO 27001 reuse: None. New for AI, not covered by ISO 27001.
Why it matters
- The 2026 IT Rules amendment makes labelling and permanent metadata for synthetically generated information (SGI) binding on intermediaries, with extra declaration duties for significant social media intermediaries (SSMIs).
- The MeitY deepfakes advisory already asked for labels or unique identifiers on content that could be misinformation or a deepfake.
- RBI FREE-AI expects watermarked or clearly labelled digital interfaces for AI in banking channels.
- Non-compliance can cost intermediaries their safe harbour under Section 79 of the IT Act.
Checklist
- Identify every feature that generates or edits media or text with AI (image tools, avatar video, voice cloning, AI copywriting).
- Decide label format per media type: visible overlay, caption, audio prefix. Make it prominent, not a tiny corner mark.
- Embed provenance metadata (for example C2PA content credentials) and a unique identifier pointing to your service.
- Remove any "export without watermark" or "strip metadata" option; test that common edits do not silently drop the label (SC-AI-029 AI-specific security controls (prompt injection, model abuse)).
- If you are an SSMI: add an uploader declaration step, a technical check of the declaration (SC-AI-036 Deepfake and impersonation safeguards), and a label before publication.
- Log which outputs were labelled and with which identifier (SC-AI-032 Logging and record keeping).
- Re-test labels after each model or pipeline change.
Evidence examples
- Sample outputs showing the visible label and the metadata read by a C2PA or EXIF viewer.
- Design document for the labelling and provenance pipeline.
- Test report proving labels survive download, crop and re-encode within reason.
- Upload flow screenshots showing the SGI declaration (SSMIs).
- Change records showing label re-testing after model updates.
Common pitfalls
- Labelling images but forgetting audio and video.
- Metadata only, with no visible label for ordinary viewers.
- Premium tiers sold with "no watermark" as a feature.
- Relying on users to label voluntarily with no technical check.
Satisfies
- BindingEU AI Act
- BindingKorea AI Basic Act
- SupervisoryMeitY AI and Deepfake Advisories
- VoluntaryRBI FREE-AI
Also relevant to
- India AI Governance Guidelinesguidance mapping (deepfake provenance; no obligation IDs)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- A.8.2 system documentation and information for users
- A.8.5 information for interested parties
ISO/IEC 27001
No direct equivalent
SOC 2
- CC2.3
Key takeaways
- Two layers: a visible label people see and embedded metadata machines read.
- Users must not be able to remove either; test it.
- In India this is now binding for intermediaries under the 2026 IT Rules.