Every major AI media provider now marks some of its output, but each uses a different mix of signed metadata, invisible watermarks and visible labels, so knowing what to expect from a given file means checking product by product. This tracker records, from each provider's own documentation, which generated images, video and audio carry C2PA Content Credentials, a SynthID or other invisible watermark, or a visible label, and which verification tools each provider offers. It covers Adobe, ElevenLabs, Google, Meta, Microsoft and OpenAI, listed alphabetically; order is never a ranking. It is written for trust and safety teams, newsrooms, compliance teams preparing for labelling laws, and developers who need to keep provenance intact through their pipelines. Every row is dated and sourced, and a field is filled only where the provider states it.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Key takeaways
- Two technical layers, used together. C2PA Content Credentials are signed metadata that can be stripped by editing or re-encoding; SynthID is a watermark embedded in the media itself that may survive some transformations, according to OpenAI's provenance documentation.
- SynthID is spreading beyond Google. Google's I/O 2026 post (20 May 2026) says OpenAI, Kakao and ElevenLabs are bringing SynthID to their content, and OpenAI's provenance API already checks SynthID on images and audio.
- Cloud APIs sign by default. Vertex AI adds Content Credentials signed by Google LLC to supported Veo, Gemini image and Lyria models, and Azure OpenAI applies them to all images from DALL-E and GPT-image-1 series models.
- Verification is moving into consumer tools. The Gemini app verifies SynthID for image, video and audio and, from 20 May 2026, C2PA Content Credentials; OpenAI runs openai.com/verify and a provenance check API.
- A negative check proves nothing. OpenAI states that a not_detected result does not rule out AI-generated content, and each provider's tool checks only its own or partner signals.
How AI-generated media is labelled
Three kinds of signal appear in the tables below. They answer different questions and fail in different ways.
| Signal | What it is | Who defines it | Main weakness |
|---|---|---|---|
| C2PA Content Credentials | A cryptographically signed manifest attached to the file, stating issuer and actions such as AI generation | Coalition for Content Provenance and Authenticity (open standard; spec page labelled version 2.3 when checked) | Removed by editing, conversion or sharing through tools that do not preserve it; any change by a non-C2PA tool fails validation |
| Invisible watermark (SynthID and others) | A signal embedded in the pixels, frames or audio, imperceptible to people | Each provider; SynthID from Google DeepMind, licensed to partners | Detection only by the owner's or partner tools; a missed detection is not proof of human origin |
| Visible label or mark | Text or a mark on the content or in the platform interface, such as "Imagined with AI" | Each platform | Can be cropped or not shown off-platform |
The C2PA describes Content Credentials as a "nutrition label for digital content". Its steering committee, as listed on c2pa.org when checked, includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok and Truepic. The mechanics of manifests, signing certificates and validation are covered in the AI Content Provenance and C2PA Guide, and the two technologies are set side by side in SynthID vs C2PA Content Credentials. This page does not repeat them; it tracks what each product actually applies.
Labelling by provider and product
As of 2026-11-15. Providers listed alphabetically. "Not stated by <provider>" means the provider documentation checked does not say either way.
| Provider and product | Media | C2PA Content Credentials | Invisible watermark | Visible label | Source |
|---|---|---|---|---|---|
| Adobe Firefly | Image and other Firefly-generated assets | Applied automatically where 100% of pixels are generated with Firefly, such as Text to Image | Not stated by Adobe on the page checked | Not stated by Adobe; credentials are metadata | Adobe Help |
| ElevenLabs | Speech, music, sound | Safety page states ElevenLabs incorporates C2PA | SynthID adoption announced in Google's I/O 2026 post; rollout detail not stated by ElevenLabs | Not stated by ElevenLabs | ElevenLabs Safety |
| Google consumer apps (Gemini app, Omni video, Lyria, NotebookLM) | Image, video, audio, text | Gemini app verifies Content Credentials from 20 May 2026 | SynthID added at creation for images and video in Google generative AI consumer products; on Omni videos; on Lyria music and NotebookLM podcasts; on Gemini app text | Not stated by Google on the pages checked | Google DeepMind SynthID |
| Google Vertex AI (Imagen) | Image | Not listed among Content Credentials models on the page checked | SynthID added by default (addWatermark true); cannot be disabled in the console; can be set false in the API | Not stated by Google | Vertex AI image generation |
| Google Vertex AI (Veo, Gemini image, virtual try-on, Lyria) | Video, image, music | Added automatically and signed by Google LLC, app or device "Google Media Processing Services"; models include veo-3.1-generate-001, gemini-2.5-flash-image, virtual-try-on-001, lyria-3-pro-preview | Not stated on the Content Credentials page | Not stated by Google | Vertex AI Content Credentials |
| Meta AI (Facebook, Instagram, Threads) | Photorealistic images (as described in Meta's post of 6 February 2024, updated 1 April 2025) | IPTC metadata embedded; Meta detects C2PA and IPTC "AI generated" signals from other companies' images | Invisible watermarks on Meta AI images | "Imagined with AI" label and visible markers | Meta Newsroom |
| Microsoft (Azure OpenAI in Foundry) | Image | Applied automatically to all images from DALL-E and GPT-image-1 series models; signed by a certificate tracing to Azure OpenAI | Not stated by Microsoft on the page checked | Not stated by Microsoft | Microsoft Learn |
| OpenAI | Image, audio | Images: signed metadata naming OpenAI as issuer with AI-generation action | SynthID on supported images and audio | Separate from the checked signals; not described on the provenance page | OpenAI content provenance |
OpenAI's Sora video product is no longer available; for what that means for existing files, see OpenAI Sora Shutdown Explained. Retired media models generally are listed in the AI Media Model Retirements Tracker.
What each manifest says
Where providers document the content of their C2PA manifest, the fields tell a verifier who signed and how the file was made.
| Provider | Signer shown | Documented fields |
|---|---|---|
| Google Vertex AI | Google LLC; app or device "Google Media Processing Services" | Not stated beyond signer and app |
| Microsoft Azure OpenAI | Microsoft Corporation (shown in Content Credentials Verify) | description "AI Generated Image"; softwareAgent "Azure OpenAI DALL-E" or "Azure OpenAI ImageGen"; when (timestamp) |
| OpenAI | OpenAI as issuer | Issuer and AI-use details; a detected result needs a trusted or valid manifest with an AI-generation action |
As of 2026-11-15. Adobe, ElevenLabs and Meta do not publish manifest field lists on the pages checked.
How to verify a file
Verification tools check only the signals their owner knows about. Use more than one.
| Tool | Owner | Checks | Stated limits |
|---|---|---|---|
| Adobe Inspect (Adobe Content Authenticity) | Adobe | Content Credentials on an asset; recovers copies stored in Adobe's Content Credentials cloud | Not a gallery or search service |
| AI Speech Classifier | ElevenLabs | Probability a clip was generated with ElevenLabs | First minute only; does not reliably classify Eleven v3 audio |
| Content Credentials Verify (verify.contentauthenticity.org, contentcredentials.org/verify) | Content Authenticity Initiative | Any C2PA manifest | Cannot validate files modified by non-C2PA tools |
| Gemini app, Search, Chrome | SynthID in image, video and audio; Content Credentials in the Gemini app from 20 May 2026 | Search verification from 20 May 2026, Chrome in the following weeks; C2PA in Search and Chrome "in the coming months" | |
| SynthID Detector | Google DeepMind | Image, video or audio made with AI from Google or partners (OpenAI, NVIDIA, Kakao; Apple listed as coming soon) | Partner signals only |
| openai.com/verify and POST /v1/content_provenance_checks | OpenAI | C2PA and SynthID on images; SynthID on audio | PNG, JPEG, WebP, GIF; MP3, Opus, AAC, FLAC, WAV, PCM; 50 MiB max; audio 60 seconds or less; strict rate limits; text verification for approved organisations only |
As of 2026-11-15. Google says SynthID verification had been used 50 million times when it announced the Search expansion on 20 May 2026.
A practical verification order
- Read the metadata first. Run the file through a C2PA validator. A valid manifest from a known signer is the strongest positive evidence.
- Check watermarks next. Use the SynthID-capable tools (Gemini app, SynthID Detector, OpenAI verify) for images and audio, and the provider's own classifier for voice.
- Look for platform labels. Note any visible label in the source platform before the file was downloaded.
- Record results, not conclusions. Log each tool, version and result. No result is proof of human origin.
- Escalate high-impact cases. For fraud, election or newsroom decisions, combine provenance checks with source verification and call-back.
Why labelling is becoming mandatory
Provider choices are being overtaken by law. Article 50 of the EU AI Act, which applies from 2 August 2026, requires providers of systems generating synthetic audio, image, video or text to mark outputs in a machine-readable format, and deployers to disclose deepfakes (Article 50 text); see the EU AI Act Article 50 Transparency Obligations Guide and the EU Code of Practice on Marking and Labelling AI-Generated Content Guide. China requires explicit and implicit labels on AI-generated content, covered in the China Generative AI Measures and AI Labelling Rules Guide. Laws by jurisdiction are mapped in the Deepfake and Synthetic Media Laws Map.
The practical effect for organisations is that relying on the provider's default is not enough: if you edit, compress or republish generated media, your own pipeline must preserve or re-apply the marking, and your publishing systems must show a label where the law requires one.
Decision guidance for teams
- If you publish generated media: choose tools that sign Content Credentials by default, keep originals, and test that your CMS and image CDN do not strip manifests.
- If you moderate user uploads: read C2PA and IPTC "AI generated" signals at ingest, as Meta describes doing, and ask users to self-disclose audio and video where signals are unreliable.
- If you run a contact centre or payments desk: do not rely on detection of cloned voices; use call-back and out-of-band approval. See the Spotting Deepfake Scams Guide.
- If you build on cloud APIs: check the documented default (Vertex AI Imagen adds SynthID unless addWatermark is set false) and record any decision to change it.
- Small organisations: use the free web verifiers and a written labelling rule. Large organisations: automate checks at ingest through APIs and log results for audit.
Common failure patterns
- Stripping metadata in the pipeline. Resizing, format conversion or social upload drops the C2PA manifest, so the published file carries no credential. OpenAI and Google both warn that editing or non-C2PA tools remove or invalidate it.
- Treating a negative as proof. A not_detected watermark result is read as "human made", although providers state it does not rule out AI generation.
- Checking only one signal. Teams run a single provider's detector, which checks only that provider's or its partners' marks.
- Turning off defaults without a record. Disabling a default watermark for deterministic seeds or other reasons, with no approval or log, leaves unmarked output in circulation.
- No label at publication. Machine-readable marks exist, but the website or app shows no human-readable disclosure where the law requires one.
Detection and measurement
- Share of published AI-generated assets that still carry a valid C2PA manifest after your publishing pipeline.
- Share of inbound media checked at ingest for C2PA and watermark signals.
- Number of generation tools in use whose labelling behaviour is documented in your register.
- Count of approved exceptions where a default watermark was disabled, with owner and reason.
- Time from a suspected deepfake report to a recorded provenance result.
For the governance side of model provenance and watermark programmes, see the AI Model Provenance and Watermarking Guide.
Practitioner checklist
- List every AI media tool in use and record, from this tracker or the provider's docs, which signals it applies.
- Test your publishing chain with a signed file and confirm the manifest survives to the live page.
- Keep generation defaults on (SynthID, Content Credentials) unless a documented exception is approved.
- Run at least two verifiers on any high-impact file: a C2PA validator and a watermark checker.
- Log verification results with tool, date and outcome, never as a yes or no verdict on authenticity.
- Add a visible AI disclosure where EU, Chinese or other applicable rules require it.
- Re-check provider documentation before relying on a row here; labelling features change frequently.
Standards and references
- EU AI Act, Regulation (EU) 2024/1689, Article 50: Transparency obligations (text)
- C2PA: Coalition for Content Provenance and Authenticity
- C2PA: Specifications 2.3
- Google DeepMind: SynthID
- Google Cloud: Content Credentials in Vertex AI
- OpenAI: Content provenance
- Microsoft Learn: Content Credentials in Azure OpenAI
- Adobe: Content Credentials overview for Firefly
- Meta: Labeling AI-generated images on Facebook, Instagram and Threads
Related NHI Mgmt Group resources: AI Content Provenance and C2PA Guide · SynthID vs C2PA Content Credentials · AI Image, Video, Voice and Music Models Hub · ElevenLabs Explained