Synthetic media detection
Detection models behind a stable interface, starting with still images, with video and audio to follow.
Synthetic media compliance
Labelling and provenance for AI-generated media.
A compliance layer for platforms that host images, video and audio: detect synthetically generated content, apply the labels and provenance metadata the rules require, and keep a tamper-evident log that shows it was done.
Status: Research and development
The problem
India’s IT Amendment Rules, 2026 came into force on 20 February 2026. Platforms must label synthetically generated audio, visual and audio-visual content prominently and embed permanent metadata or another technical provenance mechanism where feasible, and significant social media intermediaries must obtain user declarations and verify them with technical measures. In the EU, the transparency duties in Article 50 of the AI Act have applied since 2 August 2026. Neither regime tells a platform how to build it.
How we approach it
Declarations, detection and provenance are combined into one policy decision per item. A detector scores each upload, per-platform rules decide the label, and a conflict between a user’s declaration and a strong detection goes to a human reviewer instead of being labelled automatically. Signed outputs carry both an invisible watermark and a C2PA content credential, so provenance survives when metadata is stripped, and every decision is written to a hash-chained compliance log that reports are generated from.
Capabilities
What it is being built to do.
Declaration workflows
Capture of user declarations on whether content is synthetic, recorded against the item they describe.
Policy and labelling engine
Ordered rules with per-platform thresholds, and human review wherever declaration and detection disagree.
C2PA content credentials
Provenance manifests written to the open C2PA standard and verifiable with standard tooling.
Watermark-backed provenance
An invisible watermark alongside the manifest, so provenance can be recovered after metadata is removed.
Compliance log and reports
A hash-chained, anchored decision log from which periodic compliance reports are derived.
Who it is for
Built for.
Social media and content platforms
Meeting labelling and provenance duties for user uploads.
Significant social media intermediaries
Collecting and verifying user declarations at scale.
Banks and fintechs
Screening video-KYC and onboarding media for synthetic identities.
Media and publishing platforms
Marking AI-generated images, video and audio consistently.
Related
Built on this work.
AI & Intelligent Systems
We build applied AI systems - retrieval, agents, forecasting and vision - that sit inside real workflows and are evaluated against your data, not a public leaderboard.
Data Services
We engineer scalable data infrastructure that transforms complex information into reliable, actionable intelligence.
FAQ
Questions we get asked.
Under India’s 2026 amendment, audio, visual or audio-visual information created or altered artificially or algorithmically so that it appears authentic. Text-only content, routine editing and accessibility or translation tools are excluded.
No. They require permanent metadata or other appropriate technical provenance mechanisms where technically feasible, without naming a standard. C2PA is an open industry standard for content credentials, which is why the product writes to it.
Not yet. It is in research and development, and launch is timed to the technical standard for labels and metadata that MeitY is expected to publish. We are speaking with platforms preparing for it now.
Shape it with us.
We are looking for design partners who will use it in production. Tell us about your programme and we will show you where it stands.