What people ask before they wipe an image

SynthID is Google's invisible watermark for AI-generated images. It lives in the pixel data itself, not in the metadata, and survives cropping, compression and filters. Images created with Gemini (and other Google image tools like Imagen) carry SynthID by default, but there is no public detector to confirm its presence on a single image.

Not with certainty, and no tool can. We read the file’s own evidence instead. The tool says “Likely” only when the file names Google AI as the maker, and “Possible” otherwise. A Google camera make on its own doesn’t count, because Pixel photos carry that too.

Metadata and Content Credentials come off images from any generator, including Gemini, Nano Banana, ChatGPT and DALL·E, Midjourney, Adobe Firefly, Microsoft’s tools, Canva and open models like Stable Diffusion. Invisible watermarks are another matter. Google’s SynthID is the one we verified most, using Gemini’s provenance check on three kinds of image. OpenAI’s gpt-image is reported to carry a pixel watermark too; a Maximum disruption wipe of our gpt-image test picture passed OpenAI’s own verify check. Results can vary from image to image. Claude doesn’t generate images, so it has nothing to remove.

We regenerate your image with a published open-source recipe: a diffusion model re-creates the picture at a measured strength, which disturbs the frequency-domain signal SynthID relies on. We verified the output against Google's Gemini provenance check using that recipe. That check is a classifier judgment, not Google's production detector, so treat the result as strongly indicative rather than a mathematical guarantee.

C2PA Content Credentials are the signed manifest that generators like Gemini, DALL·E and Firefly embed, and the file-level “AI info” trigger that platforms read. WipeTheAI removes them losslessly by rewriting the file’s metadata without touching pixels, and the removal is provable: re-scan the output and the manifest is gone.

Because SynthID is designed to survive exactly that. The watermark is spread across the image’s frequency domain, so cropping, resizing, compressing and re-screenshotting all leave it intact. Disrupting it reliably means regenerating the image content itself, which is what Maximum disruption mode does, at a small, measured quality cost.

It removes the file-level trigger those platforms read, but they also run their own classifiers on their own servers, and those may still label the image. Nothing you do to a file controls that, so we don’t promise it.

“Keep it looking the same” runs a light pass on your device and never uploads anything. It changes the picture very little, and it is not the recipe we verified against Gemini. “Maximum disruption” regenerates the image on our servers with that verified recipe. It takes one to two minutes and changes the picture slightly more, and the similarity score tells you by how much.

Scanning your file and stripping its metadata happen on your device. The watermark wipe does upload your image, and the tool tells you on screen before it does.

JPEG, PNG and WebP. Your download keeps the format you uploaded (a JPEG comes back as a JPEG), and removing metadata never touches the pixels.

Yes, slightly, and we show you exactly how much. Every processed image gets a similarity score (SSIM) against the original, and a before/after slider lets you inspect the result before downloading. Metadata removal (C2PA, EXIF, XMP, IPTC) is completely lossless: not a single pixel changes.

Maximum disruption accepts images up to 25 MB. Scanning and metadata removal happen on your device and upload nothing, so that cap doesn’t apply to them.

Small printed text, like product labels, can come back garbled, and fine texture such as knitwear can shift slightly. For small faces in a wider shot, the eyes and face come back from your original so the person still looks like themselves. Large close-up faces are fully regenerated, so check them with the slider before you download.

Only for a while. An image you send for Maximum disruption is stored privately and deleted 7 days after upload; the wiped result is deleted 30 days after it is made. Until then both are in Your images, where you can delete them sooner, or delete your whole account from the Account menu. The lighter pass never uploads anything.

No. Your images and account details are used only to run the wipe you asked for.

Your first 3 images are free and need no card. After that you buy a one-time pack of images; there is no subscription. Payments aren’t live yet, so for now every account gets 3 free Maximum disruption images a day, and Keep it looking the same is free with no limit.

Yes. Everything under Free tools, from compressing photos to signing a PDF, costs nothing, needs no account and shows no ads. There is no daily limit, no paid version for bigger files and no watermark added to what you download. Your files are opened and saved on your own device, so nothing is uploaded. We keep them free because some people who use them later come back to wipe an AI image with us.

A failed job refunds its credit, and the tool tells you what went wrong.

Yes, to wipe images you sign in with Google. Your wipes are linked to your account; Google shares your name, email address, profile photo and account ID with us, and nothing else.

Maximum disruption is limited to 3 images a day per account while payments aren’t live, plus an hourly limit so one person can’t tie up the service for everyone. A wipe that fails doesn’t count. The lighter mode runs on your device and has no limit.

For images you created or are licensed to modify, cleaning them is generally fine. What matters is what you do next: passing AI-generated content off as human-made can violate platform terms, and the EU AI Act’s transparency duties (phasing in through 2026) require AI-generated content to be labelled in many contexts. Use WipeTheAI on your own content and keep disclosing where the law or the platform requires it.

No. Our terms forbid using it to mislead people about where an image came from, or to get around a law or platform rule that requires AI content to be labelled. It is meant for cleaning your own content and for research into how well watermarks hold up.

Because nobody outside Google can prove a SynthID watermark is gone. Our recipe passed Gemini’s provenance check on every kind of test image we tried, but that check is a classifier, not Google’s production detector, and results vary from image to image. Metadata is different: its removal is provable, and a re-scan shows it.