Content Credentials and C2PA: what they are and what to do in 2026
Signed metadata, invisible watermarks and fingerprints: how Content Credentials and SynthID work, where they break, and a checklist for brands.
Content Credentials are digitally signed metadata, defined by the open C2PA standard, that record who created a file, with which tools and how it was edited. Invisible watermarks such as Google’s SynthID add a second layer inside the pixels and audio, which survives better when the metadata is stripped. As of November 2026, Content Credentials, C2PA and watermarks are applied by the leading AI generators and read by Google, YouTube and TikTok: a brand’s job is to keep them intact and pair them with a visible label.
This guide covers how they work, who uses them, where they break and what the rules in Europe and California require. For the label the audience has to see, start with our guide to the AI Act and AI-generated video.
What Content Credentials and C2PA are
The C2PA (Coalition for Content Provenance and Authenticity), founded in February 2021, is the coalition that writes the technical standard. Content Credentials are how it works in practice: a cryptographically signed “manifest” attached to the file, which can record tools, date, location and editing steps, with or without AI. The C2PA FAQ describes them as a cryptographically secure way to express where content comes from.
Version 2.3 of the standard, announced in February 2026, brought credentials to live video for broadcast and streaming. In January 2026 the community promoting the standard had more than 6,000 members, according to the Content Authenticity Initiative.
“Valid” does not mean “true”
A C2PA verifier can return two results. “Valid” means the credential is intact, but the signer may not be on the Trust List; “Trusted” means the signer is recognised, as the C2PA deployment guidance of July 2026 explains. Since October 2025 a conformance programme has certified products, and the first to reach the highest level was Google’s Pixel 10.
A credential tells you who signed what, not whether the scene really happened. A World Privacy Forum report, covered by Fortune in September 2025, stresses that the standard neither detects deepfakes nor flags copyright infringement.
Invisible watermarks: how SynthID works
An invisible watermark is a signal embedded in pixels, frames or audio at the moment of generation. Google DeepMind says SynthID is designed to survive cropping, filters, frame-rate changes and lossy compression. On 19 May 2026 Google reported more than 100 billion images and videos and 60,000 years of audio watermarked, and 50 million checks in the Gemini app.
The same day, SynthID moved beyond Google. OpenAI now applies it to images from ChatGPT, Codex and its API, alongside the Content Credentials it has used since 2024 and extended to Sora. ElevenLabs and Kakao are adopting it as well, and NVIDIA for video from its Cosmos models. TikTok has its own watermark that only the platform can read, in testing since November 2025.
Metadata, watermarks, fingerprints: the differences
The C2PA guidance recommends layering several methods, because no single one is enough.
| Technique | What it does | Strength | Weakness |
|---|---|---|---|
| Content Credentials | Signs the file’s history | Detailed and verifiable | Lost in screenshots and exports |
| Invisible watermark | Signal in pixels or audio | Survives compression and crops | Little data, needs the right detector |
| Fingerprint | Computed from the content and stored | Recovers a lost credential | Matching is probabilistic |
| Visible label | Icon or text in the video | Anyone can see it | Can be cropped out |
Who actually uses them in November 2026
Generators. Google watermarks content from its own models with SynthID, and since 18 December 2025 you can verify videos in the Gemini app, up to 100 MB and 90 seconds; since May 2026 the app also reads Content Credentials. OpenAI is a C2PA “Conforming Generator Product” and is previewing a public tool that checks images made with its products.
Cameras. The Pixel 10 was the first phone to sign photos from its built-in camera app with Content Credentials, and in May 2026 Google announced video support for the Pixel 8, 9 and 10. In professional video, the CAI notes, Sony’s PXW-Z300 applies them at capture.
Platforms
- YouTube. Since May 2026 it applies an AI label automatically when it detects undisclosed photorealistic AI use, and makes it permanent if C2PA metadata shows the content is fully generated. On its own, the label changes neither recommendations nor monetisation. The full rules are in our guide to YouTube and AI content.
- TikTok. It reads Content Credentials on uploads, and by November 2025 said it had labelled more than 1.3 billion videos.
- Instagram. According to Google’s May 2026 announcement, Meta plans to label camera-captured media carrying Content Credentials.
- Search and Chrome. Google started rolling out SynthID checks there in May 2026 and announced Content Credentials for the following months.
Where they break: the limits to know
Metadata is fragile: OpenAI itself says it can be lost through uploads, downloads, format changes, resizing and screenshots. In October 2025 a test published by Indicator showed that saving an AI video to a laptop before uploading it elsewhere was enough to remove the data used to label it. Cropping also defeated the label.
Watermarks hold up better, but each has its own reader. The Gemini app only recognises content made with Google’s AI: if it finds no SynthID, the file may still come from another generator. The SynthID Detector portal, which Google DeepMind says is now open to everyone, also reads content from OpenAI, NVIDIA and Kakao, but not from anyone who skips SynthID. And detectors make mistakes: in July 2026 Lead Stories found that Gemini repeated its verdict on the first file in a chat when asked about a second, a bug Google fixed on 16 July.
Then there is privacy: a credential can hold dates, location and the signer’s identity, so decide what goes in. The working rule: provenance that is present and trusted is useful evidence; its absence proves nothing.
What the rules require in Europe and California
Article 50 of the AI Act requires providers of generative systems to mark their outputs in a machine-readable format. The code of practice published by the Commission on 10 June 2026 turns that into technical commitments, summarised by the law firm Freshfields:
- at least two machine-readable layers of marking where a single method is not enough;
- metadata digitally signed and time-stamped in a tamper-evident way;
- interoperable watermark detection by 2 February 2027;
- detection free of charge in principle, and always free for authorities, media, fact-checkers and researchers.
Systems already on the market before 2 August 2026 have until 2 December 2026. These are duties for providers; what stays with whoever publishes is the visible label.
In California, the AI Transparency Act (SB 942, as amended by AB 853) has applied since 2 August 2026 to providers of systems publicly accessible in the state with more than one million monthly users, as OneTrust summarises: “latent” disclosures embedded in files and a free detection tool. From 1 January 2027, platforms with more than two million monthly users must detect provenance data, show it and not knowingly strip it; capture devices follow from 2028. The penalty is USD 5,000 per violation, per day.
What a brand should do: a checklist
- Put it in the brief. Which tools will generate the images and video, whether they apply Content Credentials or watermarks, and who is accountable if they get stripped.
- Protect the chain. Work from original files, with software that appends to the history rather than erasing it.
- Choose conforming tools. Check the list of C2PA conforming products before adopting new software.
- Verify before you publish. Run masters through a C2PA verifier and, for Google or OpenAI content, through their own tools.
- Label anyway. Nobody sees metadata or watermarks: a deepfake needs a visible label.
- Explain it on your site. The C2PA guidance suggests the “Cr” badge next to AI assets, a link to a verifier and a plain-language note.
- Archive the originals. Masters with credentials, dates and versions are your evidence if anyone disputes or fakes your content.
Provenance and delivery: how we work at Kult
At Kult every format comes from a single master, delivered in post-production at up to 4K. Usage rights, revision rounds and delivery are written into the proposal before work starts, which is the right moment to add a line on metadata and labels. If you work with tools from Adobe, one of our partners, read our guide to Adobe Firefly for video, or get in touch.
This article is general information, current as of November 2026, and is not legal advice.
Frequently asked questions
How do I check the Content Credentials of an image or video?
Upload the file to a C2PA verifier such as verify.contentauthenticity.org, or to the Gemini app, which since May 2026 shows a file’s origin and edit history when credentials are present. Finding nothing does not mean the file is authentic: the metadata may have been stripped.
Can Content Credentials be removed from a file?
Yes, and it often happens by accident: a screenshot, a format conversion or a fresh export can wipe them. That is why providers pair them with invisible watermarks such as SynthID, which stay in the content, and why the EU code of practice asks for at least two layers of marking where one is not enough.
Does SynthID detect video made with other AI models?
Only if that model applies SynthID. Besides Google, it is used or being adopted by OpenAI for images, ElevenLabs, Kakao and NVIDIA for Cosmos video. A video from another generator will carry no watermark, so the check comes back negative even if the video was made with AI.
Sources
- FAQ · C2PA (Coalition for Content Provenance and Authenticity), 2026-11
- The C2PA launches Content Credentials 2.3 and celebrates 5 years of impact across the digital ecosystem · C2PA, 2026-02-09
- Content Credentials Deployment Guidance, version 1.0 · C2PA, 2026-07-08
- The State of Content Authenticity in 2026 · Content Authenticity Initiative, 2026-01-18
- Raising the bar for trust: introducing the C2PA Conformance Program · Content Authenticity Initiative, 2025-10-09
- Big Tech’s C2PA Content Credentials standard for fighting deepfakes puts privacy on the line · Fortune, 2025-09-18
- SynthID · Google DeepMind, 2026-11
- Making it easier to understand how content was created and edited · Google, 2026-05-19
- Verify Google AI-generated videos in the Gemini app · Google, 2025-12-18
- Verify Google AI-generated images and videos with SynthID · Gemini Apps Help, 2026-11
- Advancing content provenance for a safer, more transparent AI ecosystem · OpenAI, 2026-05-19
- More ways to spot, shape and understand AI-generated content · TikTok Newsroom, 2025-11-19
- Improving AI labels for viewers and creators · YouTube Official Blog, 2026-05-27
- The Indicator guide to AI labels · Indicator, 2025-10-02
- Google Gemini SynthID detector confuses results within same chat · Lead Stories, 2026-07-13
- Code of Practice on Transparency of AI-generated Content · European Commission, 2026-10-19
- Code of Practice on Transparency of AI-generated Content: FAQ · European Commission, 2026-06-10
- EU AI Act unpacked #33: the final Code of Practice on Transparency of AI-generated Content · Freshfields, 2026-06-25
- California AI Transparency Act: requirements, deadlines and compliance steps · OneTrust, 2026-07-30
Keep reading.
AI Act and AI-generated video: what brands and agencies must do
Article 50, the digital omnibus, the EU code of practice and Italian law: who must label AI-generated video, when and how. With a checklist for your brief.
AI video copyright: who really owns a video made with AI?
An AI-made video is protected only where human intellectual work sits behind it. Italian law, the EU view, tool terms and the clauses your contract needs.
AI Christmas ads in 2026: who uses AI, who refuses it and why
Prada draws by hand, Levi’s goes back to claymation, Dior faces AI accusations: Christmas 2026 shows what audiences punish and accept from AI.