Is This Image AI? We Checked 1,331 Web Images for Content Credentials After the EU AI Act Deadline

1. What is C2PA? Content Credentials explained

Content Credentials are a signed record of where an image came from and what was done to it, embedded in the image file itself. The format is the C2PA standard, published by the Coalition for Content Provenance and Authenticity. A camera, an editing app or an AI generator writes a manifest that lists the actions taken (captured, cropped, generated), signs it with a certificate, and records a SHA-256 hash of the image so that a verifier can tell whether the file changed afterward.

A second, simpler label lives in the image metadata. The IPTC digital source type vocabulary has a term that, in IPTC's own words, describes "Digital media created algorithmically using an Artificial Intelligence model trained on captured content." Generators and editors can write that term into an image, inside a C2PA manifest or on its own.

Both labels share one weakness: they travel inside the file. Anything that rewrites the file, such as a content management system, an image CDN or a social platform, can drop them or leave them out of date.

2. What the EU AI Act requires

Since August 2, 2026, providers of generative AI systems in the EU have a labeling duty. Article 50(2) of the AI Act reads:

"Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."

The Act does not name C2PA, but a signed, machine-readable manifest is the most widely implemented way to meet that requirement. Systems already on the market before August 2, 2026, were given until December 2, 2026, by a 2026 amendment. The duty falls on the provider of the AI system, so the Act says nothing about whether a website that later publishes an image keeps the label intact. That gap is what we measured.

3. How we tested 1,331 images

On November 8, 2026, we visited 105 large websites in six groups: news (30), stock photo sites (15), AI image generators (15), social and user-generated content platforms (15), ecommerce (15) and companies that build or promote content provenance (15). From each site we took the homepage and up to three inner pages, and, for the 30 news sites, their public RSS feeds as well. We collected the images those pages publish, choosing the largest variant offered, and downloaded each file exactly as a visitor's browser would receive it.

For every file, we inspected its bytes directly to see whether it carried a C2PA manifest, the IPTC label for AI-generated media, and EXIF or XMP metadata. Where a manifest was present, we recomputed the image's SHA-256 and compared it with the hash recorded in the manifest. Of the 105 sites, 55 returned images we could analyze; the others refused automated requests or loaded their images only through JavaScript. The full method and its limits are in the methodology section.

4. How many images carry Content Credentials

Almost none. In ProofSnap's November 2026 test, 3 of the 1,331 images carried a C2PA manifest and 1 carried the IPTC AI label, in the same image as one of those manifests. Ordinary metadata survived far more often: 233 images (17.5%) still had EXIF data, and so did 126 of 914 news images (13.8%), compared with the 20% that Imatag reported for editorial sites in 2018. That gap suggests most of these images never carried a manifest, though our data cannot show what the original files contained.

Read the AI generator row with care. The images we reached on Ideogram, Playground, Krea, Runway, Stability AI, DeepAI and Lexica are what those sites publish on their own pages, which mixes generated showcase images with ordinary marketing graphics. None of the 98 we checked in November 2026 carried a manifest. That does not show what a user receives when downloading from the generator itself; it shows that the images these sites put on the open web arrive without one.

5. The three labeled images we found

Each of the three tells a different part of the story:

1. Two AP photos on politico.com. Both carry a full chain: signed inside a Sony camera, then edited in Capture One, with the actions recorded and the IPTC source type set to digital capture. The provenance traveled all the way from the camera to Politico's image server. But when we recomputed the hash, neither file matched the hash in its manifest, so something changed the file after it was signed. A C2PA verifier checks that hash, so these manifests would not verify as served.
2. One ad on corriere.it. A 300 x 250 banner served from an advertising partner's server carried a manifest signed by Canva, with the action "created" by "Canva AI" and the IPTC source type for images composited with AI. Its hash still matched. This was the only intact, AI-labeled image in the whole sample, and it was an ad.

A fourth image, on Google's blog, carried no manifest but referenced one stored elsewhere. Checking such a remote manifest requires a verifier that follows the link.

6. Where the label gets lost

Most images in our sample were not the files their creators uploaded: in November 2026, 784 of 1,331 images (58.9%) arrived as WebP or AVIF, formats that image CDNs usually produce by re-encoding the original. To see what such services do with Content Credentials, we passed two signed images through three public image services: one of the AP photos above, and the official C2PA test file from the standard's own code repository.

The WebP result is the encouraging one. Photon added its own signed manifest, identifying itself as "Automattic C2PA Signer - Photon", recorded that the image was resized and transcoded, and kept the original manifest as an ingredient, so the provenance chain continued. When Photon returned a JPEG instead, it copied the old manifest unchanged, and the hash no longer matched the new file. The same service, two outcomes, depending only on the output format the browser asked for.

The practical conclusion is short: a label survives only when every system between the creator and the reader is built to carry it. On today's web, few are.

7. How to tell if an image or photo is AI

No single test answers the question, so work from the strongest signal to the weakest:

1. Check for Content Credentials. Upload the image, or paste its address, into a C2PA verifier such as Content Credentials Verify. A valid manifest that names an AI tool is strong evidence. A manifest whose check fails tells you the file changed after signing.
2. Check for a Google AI watermark. Upload the image to Gemini and ask whether it was created or edited by Google AI. Gemini looks for SynthID, an invisible watermark that, unlike a C2PA manifest, sits in the image content rather than in the file's metadata. It only recognizes images made with Google's AI tools.
3. Read the metadata. An EXIF or XMP viewer shows the camera model, the editing software and, if present, the IPTC digital source type. Remember that metadata can be edited by anyone.
4. Find earlier copies. A reverse image search often surfaces the original source, an earlier date or the photographer's credit.
5. Read nothing into a missing label. In our 2026 sample, 99.8% of images carried no manifest at all, including real news photos. The absence of a label does not make an image real, and an AI image stripped of its label looks exactly like any other.
6. Use AI detectors with care. Detection tools give a probability, not a proof, and their results can change when an image is resized or compressed.