Google recently announced a big change to how its newest Gemini Nano Banana models will handle image generation. Users can now skip the visual watermark on their AI-generated pictures.

The feature is still rolling out, so you may not have it yet, but if it has reached your Google account, you'll find it as a new Media Watermark option in the Settings menu.

It's a global toggle, so turning it off once will prevent the watermark from appearing on all images you generate using Gemini until you turn it back on.

It may even be off by default — it was for me — so if you see new images lacking the watermark and want it back, you'll have to toggle it back on.

I'm thrilled that it's gone, as it wasn't only ugly — it was rather pointless, especially since Google has a more powerful watermark in play.

Gemini's visible watermarks were security theater

Google simply admitted the truth we already knew

Those little Gemini logos in the lower-right corner of AI-generated images may have looked like a good way of setting them apart, but they were an entirely unreliable narrator.

They were fine for casually generated images that folks shared without giving them much thought, but they're also trivial to remove.

A 30-second generative fill or repair in Photoshop was all it took. A cottage industry of reverse alpha-blending tools also sprang up explicitly to remove Gemini watermarks.

Plus, many folks removed the watermark for aesthetic reasons, not to hide the image's provenance.

I've done this myself. I'll never use an AI-generated image without attributing it as such in the caption, but I despise the watermark.

Google has been building deeper protections into its AI-generated images for a long time.

C2PA: The fragile letter of provenance

A digital passport can show you where an image came from — if it survives

The first hidden gem inside an AI-generated image comes from an open industry standard by the Coalition for Content Provenance and Authenticity (C2PA).

This is cryptographically signed metadata, separate from standard EXIF data. It can be embedded in any image, and not only by generative AI tools.

Major camera manufacturers like Canon, Nikon, and even Google's Pixel also embed C2PA metadata.

It's not there solely to say an image is AI-generated, but to identify where it came from, whether that's a Canon R5 Mark II, Nano Banana, or even Photoshop.

C2PA data is tamperproof, but it's not foolproof. The digital signature means you can't fake it, but it's about as trivial to remove as a watermark.

This data is so fragile that many folks remove it unintentionally. Almost any image editing app will strip it out during export, as will many places you share photos online.

Even Photoshop, which supports C2PA — Adobe was a founding member of the coalition — only preserves it if Content Credentials are enabled in the export settings.

In other words, you can trust whatever C2PA metadata an image has, but you can't infer anything from the absence of C2PA metadata.

You can check for C2PA metadata using websites like C2PA Viewer or Chrome extensions like C2PA Content Credentials, which adds an option to the menu that comes up when right-clicking on an image.

However, don't be surprised if most images you find online lack C2PA metadata. That's why Google offers an even deeper watermark.

SynthID: The watermark baked into the pixels

More work to check, but nearly bulletproof

Google's ultimate watermark doesn't live in a mathematically subtractible transparent overlay or fragile metadata. It's woven into the images' pixels.

Dubbed SynthID, it's a system that Google DeepMind announced in 2023 to solve the common fear that AI images are becoming too "real."

While the folks at DeepMind will likely need to play a cat-and-mouse game to stay ahead of bad actors, the watermark is remarkably resilient.

Nearly every practical image manipulation technique leaves the SynthID watermark intact. Even a heavy VSCO filter won't eliminate it, much less simple Instagram crops or compression algorithms.

That's not to say it's indestructible, but researchers have found that you'd have to crop an image down to about 10% of its original size to degrade the SynthID watermark.

That's a Pyrrhic victory, though, as the resulting image would be effectively useless.

While there are other ways to hack it out, such as feeding it through another AI to generate a new image, a person must deliberately try to do so.

Here's the real strength of SynthID: While it's not unhackable, it's virtually impossible to strip out by accident.

How to check images for SynthID

Going back to the source

While Google recently launched a dedicated "SynthID Detector" portal, it is currently locked by an early-access waitlist that's restricted to journalists and other media professionals.

However, the simplest way to check an image for a SynthID watermark is to ask the same tool that would have generated it: Gemini.

Open a Gemini session, upload the image, and ask something like "Was this AI-generated?" or "Did Google AI create this?"

For the best results, tag your request with @SynthId to target the verification engine. This forces Gemini to look at the pixels rather than guess based on the image's visual composition.

You'll need to upload the image, as Gemini can't check an image from a pasted URL. However, Chrome extensions like Is This Image Real? can do some of the heavy lifting for you.

Similar to the C2PA extension, this shows up on the right-click menu. Choosing it for an image opens Gemini in a new tab, uploads the image, and asks the question.

Note that SynthID detection also isn't a zero-sum game. The tools used to detect these watermarks don't return a simple "yes" or a "no."

A SynthID watermark can only be degraded — it can't be stripped without generating a new image.

Instead, these tools operate on a Bayesian probability scale. That means there's lots of room for a "Maybe."

That may not be ideal, but it may be enough to confirm your suspicions if you're not sure about an image you're looking at.

While it's still unfortunate that we can't spot AI-generated images at a glance, it doesn't matter for most of the images we see online. For the ones where it does, the tools are there to check it out.