Does an invisible AI watermark like SynthID survive cropping and compression?

What the vendor claims, in its own words

Google DeepMind's SynthID page says the watermark is added the moment an image is created, does not change image quality, and is "designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression". The Gemini verification help page is more cautious. It says the watermark "will usually still exist" when an image is re-scaled, re-colored, compressed or altered in other ways, and adds that "after many alterations" it may not be detected.

Neither page publishes a survival rate for a particular crop size or JPEG quality, so any specific percentage you see quoted elsewhere is not coming from these sources. If you need numbers, test with your own files.

Why a watermark can survive edits that kill metadata

Metadata, including a Content Credential, sits beside the pixels in the file and is lost when an app re-encodes without copying it. A watermark of the SynthID kind is spread through the pixel values themselves. The research paper SynthID-Image describes it as a deep learning system for invisibly watermarking AI-generated imagery, reports use on over ten billion images and video frames across Google's services, and frames the design around effectiveness, fidelity, robustness and security. The C2PA explainer lists this idea as a "soft binding": an invisible watermark can help rediscover a credential that was stripped from the file.

What can still make detection fail

Situation	What the sources say
Moderate crop, resize, filter, compression	Designed to survive (DeepMind), usually survives (Gemini help)
Many stacked alterations	Gemini help: a chance the watermark is not detected
Very simple or abstract content	Gemini help: "not enough details to watermark" can come back as unclear
Very small edit to an AI base image	Gemini help: "likely too small an edit" may not carry a detectable mark
Screenshot with borders or several images	Gemini help advises cropping tight around the image and not uploading a collage
Image from a tool that does not use SynthID	Gemini can only recognise content created by Google AI tools; the DeepMind page says the separate SynthID Detector portal also covers some partners
Deliberate attack on the watermark	The SynthID-Image paper's threat model names removal and forgery, using adversarial techniques

Note the last row. The paper does not claim the mark is unbreakable. Its limitations section says watermarking alone will not solve misinformation, impersonation or copyright tracking, and that it has to sit alongside other tools such as C2PA.

How to read a result

1. Detected means all or part of the image was created or edited by the watermarking vendor's AI. It says nothing about how much of the picture is AI, or about the intent behind it.
2. Not detected means the image was not created or edited by that vendor's tools, or that the mark did not survive, or that the image is too plain to carry one. Google's own page says an image without the mark "could have been created by other AI systems".
3. Unclear is a real answer. Treat it as "use another method".

Gemini's help also mentions a usage quota of roughly ten image checks in a rolling 24 hours, so batch verification is not what the consumer tool is built for.

What to do with this as a creator

If you publish AI-assisted work, do not try to scrub watermarks or metadata so a check comes back clean. That fails more often than it works, damages image quality, and if a client or platform later asks how the image was made, it leaves you with nothing to show. Disclose the process instead, keep your layered files, and attach a credential where your tools support it.

If you are verifying someone else's image, run the checks in this order: provenance first (the Content Credentials checks), then watermark detection, then context. For the screenshot problem in particular, the detector discussion on reposted images shows what members saw, and the verification start-here guide covers the full sequence.