Privacy for File Checks

Maybe the file is a client draft. Maybe it shows your home, your child, a private workplace, or a source you need to protect. Before you drop that file into a checker, you should not have to decode a policy written for lawyers. Most current checks run on your device. The Stable Diffusion watermark check and the video and audio metadata check are the exceptions, and each page clearly says when it sends a file to our server.

What stays on your device

The current EXIF metadata viewer and Content Credentials checker are built to inspect the file locally in your browser. You choose the file, the page reads the relevant information it can access, and the result appears without requiring the file to be uploaded to Check AI Media. That design is intentional, because the safest file transfer is the one that does not happen.

Local checking does not mean every privacy risk disappears. Your own device, browser, extensions, network, workplace software, or screen-sharing session may still affect your privacy. It does mean this site does not need a copy of the file to run those current checks. We do not receive the image for the browser tools, so we do not store it, resell it, train on it, or build a private gallery from it.

If you are checking sensitive media, use the same care you would use with any local file. Avoid checking private files on a shared computer. Be careful in a browser profile full of extensions you do not trust. Do not paste results into public chats unless you mean to share them. Privacy is partly a site promise and partly a handling habit.

What the result may show you

A file can reveal more than people expect. Metadata may include camera settings, software names, dimensions, dates, or other fields depending on how the file was created and exported. Content Credentials may describe provenance information attached to the file. Those details can be useful for accountability, but they can also be sensitive. The tool shows them so you can understand the file, not because every field should be broadcast.

Think before sharing a screenshot of a result. If the result includes a date, creator name, device detail, or editing history, that information may matter to someone else. If you need to report a finding, describe only the relevant part. For example, “The file carries Content Credentials indicating AI involvement” is often safer than pasting every visible field into a public thread.

When no signal appears, that lack of a finding is also not something to overstate. The privacy page is not a detection promise. It explains handling. For result interpretation, the what can be detected guide is the better place to understand why missing metadata or credentials cannot prove a file is human-made.

The server-side watermark check

The Stable Diffusion watermark check needs server-side image processing. When you use it, the selected PNG or JPEG is sent over the same encrypted connection as the page. The service reads the image in memory, returns the result, and discards the request. It does not write the image to disk, keep a gallery, use it for training, or reuse it for another purpose.

If a file is highly sensitive, do not upload it. Prefer the local browser tools, or run an appropriate decoder in an environment you control. The watermark checker labels the upload before you choose a file so this tradeoff is not hidden.

The video and audio container check

The video and audio metadata checker also needs server-side processing, because reading a container requires ffprobe. When you use it, the selected video or audio file is sent over the same encrypted connection as the page. The service reads it with ffprobe, returns the container information, and deletes the copy as soon as the result is ready. It does not keep the file for later review, reuse it for another purpose, or add it to any library.

The read is one-way. ffprobe inspects a file without changing it, and this site does not offer re-encoding, rewriting, or watermark removal. If you would rather not send a file at all, use the browser tools where possible, or inspect the file with ffprobe yourself in an environment you control.

What we do not use your files for

We do not use files from the checks to train models. We do not sell them. We do not keep a copy for later review. We do not publish examples from user uploads. If you see an example image in an article or guide, it should be treated as site content, not as a file silently taken from a visitor.

We also do not use privacy as a cover for stronger detection claims. Keeping a file local does not make the result more certain. It simply reduces what the site needs to handle. The same honesty boundary still applies: we can say when a file carries readable AI credentials or other checkable signals, and we cannot say a file is human-made just because the check found nothing.

If your main concern is affiliation, watermark limits, or SynthID naming, the disclaimer covers those boundaries separately. Privacy answers the handling question. The disclaimer answers the authority question. The method page answers the evidence question. Keeping those separate helps prevent one kind of trust from being mistaken for another.

A practical privacy habit

Before checking sensitive media, ask two questions. First, does this page say the file stays in the browser or requires upload? Second, do I actually need to share the result, or only learn from it? Those two questions prevent many avoidable mistakes. They also make the tool more useful, because you stay focused on the evidence you need rather than giving away more context than the situation requires.

Use the checker when it helps, keep originals when provenance matters, and be cautious about publishing metadata that belongs to someone else. Privacy is not a banner at the bottom of a page. It is a promise about how the tool behaves and a practice in how the result is handled afterward.