One hides a signal in the pixels, one staples a signed record to the file, one writes a name across the picture. Take a screenshot and you learn which was attached to the image and which to the file.
A screenshot is the great leveller of provenance: it keeps everything that was in the picture and discards everything that was merely attached to it.
There are three ways to label an image, and they are usually discussed as rivals. SynthID hides a signal in the pixels. C2PA Content Credentials attach a signed record to the file. A visible mark writes a name across the picture. They are not rivals, because they do not say the same thing: the first says a machine made this, the second says this signer vouches for this file, the third says this belongs to someone. Our angle is a single thought experiment, the screenshot, which sorts the three more cleanly than any feature list. A screenshot keeps what was in the image and throws away what was attached to it.
We have reasoned from the published specifications and the developers' own statements, and where those are silent we say so. Checked on 18 November 2026.
What is SynthID?
SynthID is Google's watermarking tool for AI-generated content. Google DeepMind introduced the image version on 29 August 2023 for its Imagen model, and the product page now lists images, video, audio and text. The watermark is applied when the content is generated. For images it is embedded in the pixels, which Google describes as "imperceptible to the human eye, but detectable for identification", and it is produced and read by two deep learning models trained together, one to mark and one to identify.
Because the signal is in the picture and not in the file's metadata, it is meant to outlast ordinary handling. Google says the watermark is "designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression". The 2023 announcement adds that it "remains detectable even when metadata is lost". The same announcement is candid about limits: SynthID "isn't foolproof against extreme image manipulations".
Detection is probabilistic. The original tool reported three outcomes: watermark detected, not detected, or possibly detected. In 2025 Google launched SynthID Detector, a verification portal first tested with journalists and since opened more widely. Its page says the detector checks whether a file was made with AI from Google or from partners, and names OpenAI, NVIDIA and Kakao among them.
Two properties define what SynthID is for. It is applied by the generator, so it exists only on output from participating systems. And it carries one message: this content came from such a system. It does not say who made the image, when, or whether a human edited it afterwards.
What are C2PA Content Credentials?
Content Credentials are the public name for signed manifests defined by the C2PA specification. A manifest lists assertions about a file, such as the device or software that produced it and the edits made, binds them to a cryptographic hash of the content, and is signed with a certificate. A verifier can then tell who signed and whether the file has changed since. Our guide to Content Credentials for photo archives explains the structure in detail.
A credential is richer than a watermark. It can record that a camera captured the image, that particular software edited it, or that a generative model produced it, and it keeps earlier manifests as ingredients so that history accumulates. It can be applied to any kind of image, including a real photograph, which makes it the only one of the three methods that can make a positive statement about camera capture.
Its weakness is its attachment. In a JPEG the manifest lives in a data box inside the file. Anything that rewrites the file without preserving that box removes the credential: many platforms strip it on upload, and an edit in software that does not support the standard breaks the chain. The specification anticipates this with a soft binding, an invisible watermark or a fingerprint that lets a credential stored elsewhere be matched to a copy whose embedded data has gone. That is, in effect, an admission that the signed record needs a pixel-level companion.
What is a visible watermark for?
The overlay is the oldest method and the least fashionable: a name, logo or pattern drawn across the picture. Stock libraries have used it on previews for decades. It differs from the other two in its audience. SynthID and C2PA are read by software. A visible mark is read by people, at a glance, with no tool and no training.
Its message is also different. It makes no claim about how an image was made. It asserts a connection: this picture comes from this source, and this copy is a preview. It deters casual reuse because removing it takes deliberate effort and leaves the person who does so without the excuse of ignorance. Its costs are obvious. It disfigures the image, and a mark confined to a corner can simply be cropped away, while generative inpainting has made removing a central mark far easier than it once was.
For anyone who wants to try one, the watermark tool adds a text or logo overlay to an image in the browser.
How do the three compare?
| SynthID | C2PA Content Credentials | Visible mark | |
|---|---|---|---|
| What it is | Invisible signal in the pixels | Signed record attached to the file | Overlay drawn on the picture |
| What it says | A participating AI system generated this | This signer made these statements about this file, unchanged since | This image is claimed by this source |
| Who applies it | The generator, at creation | Camera, software, platform or publisher | Anyone |
| Who can read it | A detector run by the scheme | Any C2PA verifier | Any person |
| Works on real photographs | Not its purpose | Yes | Yes |
| Survives re-saving as JPEG | Designed to | Only if software preserves the manifest | Yes |
| Survives platform upload | Designed to | Often stripped | Yes |
| Survives a screenshot | Plausibly, not guaranteed | No, unless recovered by soft binding | Yes |
| Survives cropping | Designed to, within limits | No | Only if the mark is inside the crop |
| Main weakness | Removable by regeneration attacks; closed detection | Easily detached; absent from most images | Disfigures; croppable; removable by inpainting |
Which label survives a screenshot?
Consider what a screenshot is. The device renders the image to the display, possibly scaled, and saves the displayed pixels as a new file. Nothing of the original file is copied except its appearance.
C2PA: lost. The new file has no manifest. Nor could the old one be reattached honestly, since the hash that bound the manifest to the content no longer matches. Unless the publisher also registered a soft binding and a verifier looks it up, the provenance record is gone. This is not a flaw peculiar to C2PA. It is true of all embedded metadata, copyright fields included.
SynthID: probably present, with caveats. A screenshot amounts to resampling, perhaps cropping, and recompression, which are the kinds of change the watermark is designed to withstand. Google's published list names cropping, filters and lossy compression and does not name screenshots, so we would not state it as a guarantee, and Google itself says the method is imperfect.
Visible mark: intact. It was part of the picture, so it is part of the screenshot. This is the one test the humble overlay wins outright.
The screenshot test rewards whatever is carried by the pixels. It is why the provenance community now talks about combining methods: a signed record for detail, a pixel-level signal to find the record again, and something a person can see.
Can these labels be removed or faked?
Removal first. In a paper first posted on 2 June 2023, Xuandong Zhao and colleagues described regeneration attacks: add random noise to an image to disrupt the hidden signal, then reconstruct the image with a denoiser or a pretrained generative model. They report proofs and experiments showing that pixel-level invisible watermarks are vulnerable, with results across four watermarking schemes, and argue that research should move towards watermarks tied to the semantic content of an image. That paper did not settle the question for every scheme, and developers continue to harden their methods, but it established that an invisible mark is a speed bump for a determined adversary and not a wall.
Stripping a credential requires no attack at all, only a re-save in the wrong software. Removing a visible mark requires editing skill or a generative tool.
Faking is the mirror problem. A C2PA signature is hard to forge without the signer's key, so a false credential is mostly a matter of a signer making false statements. A visible mark can be added to anything by anyone. A pixel watermark's resistance to being planted on a real photograph, to discredit it, depends on details of each scheme that are not all public.
What does the law now require?
In the EU, Article 50 of the AI Act obliges providers of systems that generate synthetic images, audio, video or text to ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated", using solutions that are effective, interoperable and robust as far as technically feasible. Systems that perform an assistive function for standard editing are excepted. Deployers who publish deep fakes must disclose that the content is artificial. The Commission states that these obligations apply from 2 August 2026. Its Code of Practice on marking and labelling, a voluntary route to showing compliance, went through drafts on 17 December 2025 and 3 March 2026 and was published in final form on 10 June 2026; the Commission says about 190 organisations had signed by the end of July 2026, and it has issued a set of icons that deployers may use as labels.
The UK has no equivalent duty. The government's report of 18 March 2026 notes that there are no obligations in the UK for AI-generated content to be labelled and proposes working with industry on best practice. Labelling rules of this kind concern transparency, not ownership, which we cover in who owns an AI-generated image.
What do labels mean for a photograph archive?
An archive stands on the other side of this problem. The generators' task is to mark what is synthetic. An archive's task is to show that its pictures are not, and no detector can do that, because a negative result is nearly worthless. The evidence has to be positive: the film original, the raw file, dated publication, a documented chain of custody. We set out why that matters to model builders in model collapse and generated training data.
Within that, each method has a place. A signed credential on a scan can record who digitised it and when. Embedded copyright and licensing fields say whom to ask. A visible mark protects a preview. A file can be inspected for a manifest with the Content Credentials checker, which reads it in the browser without uploading the image.
The Robert Harper Photograph Library is largely film: 52,684 analogue photographs by its own inventory, and more than half of the archive is still on film and has yet to be scanned. As scanning proceeds, how each file is marked and documented is a decision to take with partners. Agencies, platforms and data teams with a view on which signals they need can tell us through the contact page.
Frequently asked questions
What is the difference between SynthID and C2PA?
SynthID is an invisible watermark that Google embeds in the pixels of AI-generated content, read by a detector. C2PA Content Credentials are a signed record attached to a file that states who signed it and what was done to it. SynthID answers whether a supported AI tool made the image; C2PA answers who vouches for the file and whether it has changed.
Does taking a screenshot remove a watermark?
It depends on the kind. A visible overlay is part of the picture and is captured intact. Embedded metadata, including a C2PA manifest, is not copied into a screenshot. A pixel watermark such as SynthID is designed to survive cropping, filters and lossy compression, though Google's published descriptions do not name screenshots and say the method is not foolproof.
Can invisible AI watermarks be removed?
Research says yes, with effort. A 2023 paper by Zhao and colleagues described regeneration attacks, which add noise to an image and then rebuild it with a generative model, and reported that pixel-level invisible watermarks were vulnerable across four schemes. The authors recommended watermarks tied to an image's semantic content instead.
Does a missing watermark mean an image is a real photograph?
No. Many generators add no watermark, credentials are stripped by many platforms, and marks can be removed. Detection results are evidence that an image was generated when positive, and close to no evidence when negative. Showing that a photograph is camera-made relies on positive proof such as raw files, signed capture credentials or a film original.
Which labelling method should a photo archive use?
They do different jobs, so most archives would combine them. A visible mark deters casual reuse of previews. Embedded copyright metadata and, where software supports it, a signed credential record who issued the scan. Pixel watermarks for AI output are a generator's tool, not an archive's, though similar invisible marks exist for tracking licensed copies.
Sources
- Google DeepMind: SynthID
- Google DeepMind blog: Identifying AI-generated images with SynthID (29 August 2023)
- C2PA Technical Specification
- Zhao et al., Invisible Image Watermarks Are Provably Removable Using Generative AI (arXiv:2306.01953)
- EU Artificial Intelligence Act, Article 50: transparency obligations (text of Regulation (EU) 2024/1689)
- European Commission: Code of Practice on marking and labelling of AI-generated content (updated 19 October 2026)
Published by Archive.London. Corrections are welcome under our editorial policy.

