How AI Image Detectors Work (and How to Judge Them)

Someone sends you a striking photo and you wonder: did a camera take this, or a generator? A good AI image detector helps you answer that, but only if you know what kind of evidence it produces. Most of the confusion around these tools comes from treating one number as a verdict. This guide explains how detectors work, where each method breaks, and how to compare tools without being fooled by a confident percentage.

The question a detector really answers

An AI image detector cannot prove that a photo is real. At best it answers a narrower question: does this file carry signs that it was created or heavily rewritten by a generative model? "Signs" can mean three very different things:

• Provenance: a label that the file itself carries, written by the tool that made it.
• Forensic traces: statistical oddities in the pixels, such as compression patterns or noise that does not match a camera.
• A learned guess: a neural network trained on real and generated pictures that outputs a probability.

These are different kinds of evidence, and they fail in different ways. A trustworthy tool tells you which one it is showing.

Method 1: provenance and metadata

This is the most reliable evidence when it exists, and the easiest to check.

Content credentials (C2PA). C2PA is an open standard for attaching a signed record to a file: which tool created it, which edits followed, who signed it. Several image generators and editing apps now write these records. If a file carries one that says "created by a generative model," that is close to a direct statement from the source.

Generator fingerprints. Many tools leave readable traces in the file: a text chunk in a PNG with the prompt and sampler settings, an XMP field naming the software, or an IPTC "digital source type" flag. They are easy to read and, when present, hard to argue with.

The weakness is just as clear. Metadata disappears. Taking a screenshot, re-saving through a messaging app, or uploading to most social networks strips it. So the rule is asymmetric: finding a tag is strong evidence, and not finding one tells you almost nothing.

Method 2: forensic analysis of the pixels

When there is no label, you can look at the image itself.

Error Level Analysis (ELA) re-saves the image as a JPEG at a known quality and compares it with the original. Regions that were edited or pasted in at a different time often compress differently and show up as brighter patches. In a plain camera JPEG the error tends to look fairly even.

ELA is useful and easy to misread. A photo that was resized, filtered, or shared several times can light up everywhere, and a fully generated image saved once as a JPEG can look perfectly uniform. Treat an ELA map as a prompt to look more closely at a region, never as a result. Other forensic views, such as noise residuals and sensor-pattern checks, have the same character: they describe the file's history, and the history of a real photo can also be messy.

Method 3: a classifier that outputs a probability

Most hosted detectors that print "93% AI" use a trained network. The idea is simple: show a model many real photos and many generated ones, and let it learn the difference. It then scores new images.

This can work well on the kinds of images it was trained on. It has three well-known limits:

1. Generators move faster than training sets. A model trained last year may not recognize output from a new generator.
2. Everyday processing hurts it. Heavy compression, screenshots, cropping, and re-upload all change the statistics the model relies on.
3. The number looks more exact than it is. A score of 87% is a model's confidence, not a measured chance of being right. Two tools can disagree on the same image and both be reasonable.

Hosted detectors also usually require uploading the file to their servers, which matters if the image is private.

What our detector does, and what it does not

The AI image detector on RestorePix AI is deliberately modest. It runs entirely in your browser, so the photo is not uploaded. It does two things:

1. Metadata and provenance scan. It looks through the start of the file for C2PA markers and the names of well-known generators and tools. It does not verify a C2PA signature chain, so treat a hit as "this file mentions a known source," then check the details yourself.
2. Error Level Analysis. It renders an ELA map you can switch to and inspect next to the original photo, and download.

It does not output an AI probability score. We have not connected a classifier, and the page says so. That means a result of "no signature found" is not the same as "this is a real photo." It only means the file did not contain the markers we look for. We would rather give you two honest signals than one invented percentage.

How to evaluate any detector

Whether you use ours or another, a few questions separate useful tools from marketing:

• Does it show its evidence? A tool that says which tag it found, or lets you see the forensic map, is easier to trust than a bare verdict.
• Does it say what it cannot do? Look for notes about screenshots, compressed files, and new generators. Silence about limits is a warning sign.
• Where does the image go? If the file is sensitive, prefer a tool that processes it locally, or read the privacy terms.
• How does it treat uncertain cases? A good tool has an "unsure" outcome. A tool that always picks a side is guessing.
• Can you repeat the result? Run the same file twice, then a re-saved copy. If a trivial re-save flips the answer, do not rely on it.

Practical ways to use a detector

Use it as one step in a short check, not the whole check:

1. Look for provenance first. A real content-credential record or generator tag settles more than any score.
2. Ask where the image came from. The original post, the first upload, and the account behind it often say more than the pixels.
3. Search for earlier copies. A reverse image search can show a photo predates the story attached to it.
4. Look at the details. Hands, text, reflections, and repeated textures are still common weak spots, though the best generators now handle many of them.
5. Keep the stakes in mind. For anything that affects a person's reputation, job, or legal situation, no automated tool should be the deciding evidence.

When the image is yours

If you edit your own photos with AI, such as restoring a damaged print, upscaling, or adding a scene, the result may legitimately carry AI-related metadata or look "processed" under ELA. That is expected and not a problem. Our editing tools, like Photo restoration and the AI Christmas photo tool, are meant for changing your own pictures, and the detector is there when you want to check where a file came from, not to hide or strip provenance.

Bottom line

An AI image detector is a set of imperfect tests. Provenance tags are strong when present, forensic maps are hints, and classifier scores are guesses that deserve their own caution. Read the evidence a tool gives you, remember that "not detected" is not "real," and try the free AI image detector whenever you want a quick, private look at what a file carries.