An AI image checker’s result is an estimate of how likely an image is to be AI-generated, based on the patterns it found. A high AI score means the checker found strong signs of generation. A low one means it found none, which isn’t the same as proof the image is real. And an “Uncertain” or middle score means there wasn’t enough signal to decide. The percentage isn’t how much of the image is AI, and it isn’t the chance the image is real in any legal or factual sense.
The parts of a typical result #
Different checkers present results differently, but most include some of these:
| Element | What it means | What it doesn’t mean |
|---|---|---|
| Verdict label (“Likely AI”, “Human”) | The checker’s call after applying its thresholds | A certainty |
| AI score or percentage | How strongly the image matches AI patterns | The share of pixels that are AI |
| Confidence | How firmly the checker stands behind its verdict | The probability you’re being lied to |
| Middle zone or “Uncertain” | Not enough signal for a call | “Probably AI” in disguise |
| Generator guess | Which model the pixels most resemble | A confirmed source, unless metadata says so |
| Heatmap or region view | Areas that contributed most to the score | A precise map of edits |
| Metadata or Content Credentials | What the file itself declares | Proof, if the data could have been stripped or faked |
Some online tools, like Hive, also estimate which generator likely produced an image. Illuminarty’s paid plans add localized detection that highlights which regions of an image look generated. Those extras can help, but they’re less reliable than the basic AI-or-not call.
How Expose AI shows its result #
Expose AI keeps the result to one screen with four parts:
- The verdict card. One of three calls, color-coded: Likely real (green), Uncertain (amber) or Likely AI-generated (pink), with a short line such as “No clear AI signals detected.”
- The meter. A bar running from Likely real to AI generated, with a marker showing the AI score and a shaded zone in the middle. A marker below the shaded zone means Likely real; inside it, Uncertain; above it, Likely AI-generated.
- A confidence figure for the verdict, shown alongside the AI score (a shared card might read “98% AI ยท 97% confidence”).
- “How we decided.” Each signal the app checked, which one was used, and a plain-language note on what the verdict can’t promise.
The app runs two passes on the phone. The metadata pass reads Content Credentials (C2PA), XMP and generator tags. If it finds a credible AI claim, such as a file that declares Firefly or Midjourney, that decides the verdict on its own. Otherwise the on-device model decides from the pixels. Camera and editor metadata appear on the card as evidence, never as proof.
What to do after each result #
Likely AI-generated #
- High confidence, decided by metadata: the file declares AI involvement. That’s about as strong as evidence gets, unless you suspect someone deliberately added a fake tag to a real photo.
- High confidence, decided by the model: strong evidence. Still check the source and context before accusing anyone, particularly if the image shows a real, identifiable person.
- Lower confidence: a lean, not a conclusion. Look for artifacts and an original copy.
Uncertain #
This is the result people misread most. It means the checker didn’t find enough signal either way, which is common with screenshots, small avatars and images that have been compressed through several reposts. What to do:
- Find a larger or earlier copy and check again.
- Crop away app interface or captions if you checked a screenshot.
- In Expose AI, run the Pro model on the same image. It’s a larger on-device model that returns roughly half as many Uncertain verdicts as the standard model on the developer’s evaluation set. Where available, you can watch a short ad to try it once on that image instead of subscribing.
- Lean harder on non-technical checks: source, context, reverse search.
Likely real #
The checker found no clear AI signals. For an original, full-size photo that’s reasonably reassuring. For a blurry screenshot it’s weaker, because compression can hide the signal. And it says nothing about whether the caption is true: a real photo from 2015 passed off as today’s news will correctly come back Likely real. Our guide to checking a viral photo covers that case.
Common misreadings #
- “70% AI means 70% of the image is fake.” No. It’s one score for the whole image.
- “Likely real means verified.” No. It means no AI signals were found.
- “Two checkers disagree, so one is broken.” Not necessarily. They’re trained on different data and use different thresholds, so borderline images often split.
- “The generator guess proves who made it.” Only metadata can name the tool with confidence. A pixel-based guess is a resemblance.
- “Uncertain means the app couldn’t work.” It did work. It’s reporting honestly that the evidence is thin.
Why checkers disagree #
Run the same image through three checkers and you may get three different answers. Each was trained on a different mix of generators and real photos, each sets its thresholds differently, and each handles compression and resizing in its own way. Disagreement usually means the image is a borderline case. Treat it as unverified and gather other evidence. How accurate AI image detectors are goes into why.
Sharing a result responsibly #
Expose AI can turn a verdict into a shareable card with the image, the call and the confidence meter, and it keeps a private history of past scans on your phone. If you share a result, share the confidence and the reasoning along with it, and remember the image may show a real person. A detector result supports a conclusion. It shouldn’t be the whole case.
Frequently asked questions #
What does a 50% AI score mean? #
It means the image sits right on the fence: the checker sees about as much evidence for AI as against it. Most checkers treat that as Uncertain. Try a better copy of the image, or rely on source and context checks.
Is a higher AI percentage always more reliable? #
A higher score means stronger AI signals, but the score is only as good as the image and the checker. A 95% score on a heavily filtered selfie can still be a false positive. Check whether metadata or the model decided, and look at the image yourself.
Why did the same image get different results in different apps? #
Checkers use different models, training data and thresholds, and some read metadata while others don’t. Borderline images, such as compressed screenshots or real photos with AI edits, are the ones most likely to get split results.
Can I check which AI generator made an image? #
Only metadata can tell you reliably, when the file still declares the tool. Some online checkers estimate the generator from the pixels, but treat that as a guess. See how to check image metadata for AI.