In a world without privacy, we create it.

Veldrinby CounterCustody
Download on the App StoreDownload on the App Store
Scan to download VELDRIN on the App Store.
TL;DRWant the short version? Click here.
The problem
Private files sit in a camera roll anyone can scroll, an album anyone can open, or a cloud you don't control.
The fix
Veldrin keeps them in a pattern-locked vault on your iPhone. No account, no vault list, and a red emergency button for when you're pressed to open it.
In short
A label inside the file is the strongest clue an image is AI made, but only if it survives.

Why you need this

In a world where any picture can be faked, your real photos are worth more, and so is the risk if they leak. Someone seeing a private image of yours, or a copy of it turning up where it should not, is hard to undo. Veldrin keeps your genuine photos in an encrypted vault on your iPhone, behind a pattern, with no account. Protect what is real. The app is waiting when you want it.

Tool · 5 min read

How to check if an image is AI generated

Start with the file itself: open it in the metadata viewer and look for a Content Credentials marker, an AI digital source type, or generator settings. If one is there, you have a statement from the software that handled the image. If none is there, you've learned very little.

Three kinds of evidence, and how far each one goes

No single test settles it. Labels written into the file are direct statements but easy to lose. Watermarks hidden in the pixels survive more handling, but only the company that made them can read them. Judging from how the picture looks is the weakest of the three, and getting weaker as the generators improve.

What an AI label inside an image file looks like

Several things can appear in a file that came from a generator, or from an editor that used one.

  • A Content Credentials (C2PA) manifest. A signed record of the tool that made or edited the image and the steps taken. The Content Credentials guide explains the structure.
  • An IPTC digital source type. A standard vocabulary includes a value for media created by a generative model trained on real content (trainedAlgorithmicMedia), one for real media edited with generative AI such as inpainting (compositeWithTrainedAlgorithmicMedia), and one for mixed material (compositeSynthetic). It's stored in the XMP block.
  • Generator settings in PNG text. Some open-source image tools write the prompt and sampler settings into PNG text fields with names like parameters or workflow.
  • A creator-tool string. XMP and EXIF can both name the software. An editing program named there tells you which program saved the file, not whether AI was used.

Inspecting an image file in three steps

  1. Open the metadata viewer and choose the file. Nothing is uploaded.
  2. Read the Provenance rows first, then Device. A source-type value or a C2PA marker is the useful finding.
  3. Weigh the answer by what you found. A label is a claim by whatever signed or wrote it. No label means the information is absent, not that the image is real.

The viewer prints strings it can read from the C2PA data, but it doesn't check the cryptographic signature. To validate a credential, use a verifier built for the purpose, such as the Content Authenticity Initiative's Verify page, and read its privacy terms first, since it's a web service and your image goes to it.

Why AI labels vanish from images

Labels sit in the file wrapper, not in the picture, so anything that rebuilds the file from pixels drops them: a screenshot, a re-save in many editors, a format conversion, an upload to a service that recompresses images. OpenAI says C2PA data can be removed and that platforms and other handlers need to retain it. An image that was labeled the day it was made can arrive in your inbox with nothing at all.

Why AI image detectors disagree with each other

Programs that look at the picture and return a percentage are estimating. They're trained on samples from particular generators and can be fooled by a new model, a heavy edit, or an ordinary phone photo that has been compressed several times. Two detectors can give opposite answers on the same file, and a confident number isn't the same as a correct one. Treat the number as a reason to look further.

Invisible watermarks such as SynthID

Some generators embed an imperceptible watermark in the image itself. Google describes SynthID this way: the mark is meant to stay detectable after edits like cropping, filters and compression, and it's read with Google's own detection tools rather than with a general-purpose viewer. This page can't read it. Only the company that applied a watermark can say whether it's present, and a missing watermark proves little as well, since many generators don't add one. If you're wondering what can and can't be stripped, removing an AI watermark draws that line.

Checks that don't depend on the file at all

  • Find the original source. Who posted it first, and with what caption?
  • Run a reverse image search to see whether the picture appears earlier, or in a different context.
  • Ask the poster for the original file rather than a screenshot.
  • Look for consistency across several images from the same event, not for a single odd finger.

Hands, text and reflections used to be reliable tells, and they're less so every year, which is why source-based checks now carry more weight than squinting. A related thing to look at is where a photo was taken, because a camera file that claims a location and a time can be compared with the story that came with it.

Questions people ask

Is it possible to tell whether an image was made by AI?

Sometimes. A valid credential or a clear source-type label settles it. Without one you're weighing probabilities, and you often can't be sure from the file alone.

Are AI image detectors accurate?

They vary, and they disagree with each other. A detector result is a clue that falls short of proof, and a newer generator can defeat an older detector without any visible change in the image.

Does a screenshot keep AI metadata?

Usually not. A screenshot is a new image made from pixels, so file labels don't come along. Open the result in the viewer to be sure rather than assuming.

Does Instagram remove EXIF data?

Platforms re-process uploads and policies change, so don't rely on it either way. Download what the platform serves and read that file instead of the one you uploaded.

Which AI generators can the viewer identify?

It doesn't guess from the picture. It prints only what the file states, such as a C2PA claim generator, an XMP creator tool or a PNG settings field.

Does the image leave my device?

No. The viewer reads the file in the tab. The separate Verify service mentioned above is different: it's a web service, so read its terms before you use it.

Sources

VELDRIN

Collect the files you are verifying in one organised vault

See Veldrin on the App Store