When AI Image Detectors Get It Wrong: False Positives and Screenshots

Real photos get flagged as AI when screenshots, compression or AI upscaling strip the fingerprints of a camera capture. Here is how to spot a false positive.

An AI image detector false positive usually means a real photo went through so much processing that it no longer looks like a camera capture. Screenshots, heavy compression, AI upscaling and phone computational photography are the four biggest causes. GPTOne's free image detector reports a confidence score plus a region heatmap, and the heatmap is what lets you spot this happening.

Detectors do not look for weird hands. They measure how an image was rendered. Change how a real photo was rendered enough times and the measurement drifts toward AI.

Key Takeaways

• 4 processing steps cause most false positives: screenshotting, re-compression, AI upscaling or denoise, and aggressive phone processing.
• AI upscaling is the trickiest case, because the extra detail genuinely is generated. The detector is not fully wrong.
• Diffuse heat across the whole frame points to a processing artifact. Heat concentrated in 1 or 2 regions points to a real local edit.
• False negatives run the other way: downscaling, added grain and print-then-rescan all weaken the signal an AI image leaves.
• The original file beats every argument. Ask for the camera original before accepting or disputing any score.

Why an AI image detector false positive happens

Every image carries traces of how it was produced. A camera sensor leaves a particular noise pattern. A lens leaves particular optics. JPEG encoding leaves a particular block structure. Detectors read those traces, as our explainer on how detectors work covers in detail.

The problem is that generated images are statistically cleaner than camera captures in specific ways. Their noise behaves too regularly. Their frequency content is too smooth in places a lens never is.

Now think about what processing does to a real photo. Compression smooths out noise. Denoise removes it deliberately. Upscaling invents plausible detail. Each of those pushes a real photo in exactly the direction that reads as generated.

The detector is not malfunctioning. It's reporting that the image no longer has the fingerprints of a raw capture, which is true. The inference you draw from that is where it goes wrong.

The four processing steps most likely to trigger a flag

Screenshotting. The single most common cause. A screenshot re-renders the image through your display pipeline and re-encodes it, discarding the original noise structure and the metadata with it. Most suspicious images reach people as screenshots, which is the worst possible starting point for a check.

Re-compression. Every save, upload and download re-encodes. An image that has been through three platforms carries three generations of compression artifacts and very little of its original signal.

AI upscaling and denoise. Phone super-resolution, Topaz, "enhance" buttons. This one deserves care, which is why it gets its own section below.

Computational photography. Modern phones do not take one photo. They stack multiple exposures, apply learned sharpening, smooth skin and reconstruct faces and skies with models. The output is a real scene reconstructed by software. Portrait and night modes are the most affected.

Add beauty filters and heavy editing on top and you have a real photograph that has been through more generative processing than some images people would call AI.

AI upscaling is the honest edge case

If you ran a 2010 photo through an AI upscaler, part of that image is now genuinely AI-generated. The upscaler invented pixels that were never captured. It guessed at texture, edges and detail based on training data.

So when a detector flags it, the detector has a point. The subject is real, the composition is real, and a meaningful fraction of the pixels are synthetic.

This is where region output earns its keep. An upscale typically produces heat spread fairly evenly, because the whole frame was reconstructed. A face swap produces a hot patch on a face. Our post on reading which parts of a photo are AI covers that distinction.

Worth deciding in advance: does your workflow care whether an image was generated or whether it was altered? For insurance and journalism the second question usually matters more.

False negatives: when AI images slip through

The failure runs both ways, and this direction gets discussed far less.

Downscaling. Shrinking a generated image throws away the high-frequency detail detectors rely on. A 4K generation reduced to 600 pixels wide is a much harder call.

Added grain. Film grain overlays are a one-click filter and they bury the too-clean noise signature that gives diffusion output away.

Print and rescan. Printing a generated image and photographing it produces a genuine camera capture of a fake image. The physical layer is real.

Heavy crop. A tight crop removes most of the frame, and with it most of the evidence, including any of the artifacts around hands, text and backgrounds that our 12 signs guide describes.

Anyone claiming a detector that cannot be defeated is selling something. Our 60-image benchmark includes the categories where our own results were weakest.

How to tell a false positive from a real detection

Work through this in order.

1. Look at the heat pattern. Even spread across the frame, including sky and flat areas, suggests processing. Concentrated on a face, an object or a background seam suggests a real edit.
2. Ask how the file reached you. Screenshot, forwarded chat image or platform download means the signal is degraded and the score deserves less weight.
3. Get the original. A camera original with intact metadata settles most disputes in seconds. Our guide to reading photo metadata covers what to look for, and note that metadata absence proves nothing on its own.
4. Check for provenance. Intact C2PA Content Credentials or a SynthID mark outrank any inference from pixels.
5. Look at the image yourself. If a detector says 80% and you find no visual tell anywhere the heat sits, weight your own eyes more heavily.

What to do if your own real photo gets flagged

It happens, and it is not an accusation you have to accept.

Send the original file rather than a screenshot or a chat forward. Include the EXIF block if it survived. Say what processing the image went through, because "shot on a phone in night mode, then upscaled" explains a high score completely. If the photo was edited, say which parts, since a heatmap that agrees with your account of the edit supports you rather than contradicting you.

And push back on any process that treats a single percentage as proof. A score is a signal, and a signal is a reason to look closer, not a reason to conclude.

Why this matters more in a queue than in a one-off check

Checking one image, you can afford to be careful. Checking two hundred listing photos a day, your error rate is a policy.

At volume, false positives are not an inconvenience, they're a cost. Every wrongly flagged real photo becomes an appeal, a support ticket or an angry seller. Moderation and claims teams that route on a raw threshold, say flag everything over 60%, end up with a queue full of night-mode phone photos and AI upscales.

The fix is not a higher threshold, since that lets real fakes through. It's routing on the heat pattern rather than the number: whole-frame heat goes to a light-touch review, concentrated heat on a claim-relevant region goes to a human immediately. Same detector, far fewer bad decisions.

FAQ

Why did an AI detector flag my real photo?

Most likely because it was screenshotted, compressed repeatedly, upscaled or shot in a heavy computational mode like night or portrait. All of those remove the noise signature detectors use to recognise a camera capture.

Does a screenshot make a photo look AI-generated?

It makes it harder to judge in both directions. Screenshotting strips metadata and re-encodes the pixels, which weakens the evidence for and against, and pushes scores toward the middle.

Can editing a photo in Photoshop cause a false positive?

Ordinary edits like cropping and colour correction rarely matter much. Generative fill, content-aware removal and AI denoise are different, since they insert generated pixels and a detector flagging that region is technically correct.

How do I prove my photo is real?

Produce the camera original with its metadata, and the provenance credentials if the file has them. A raw file from a real camera is far stronger evidence than any detector score.

Do AI image detectors get better over time?

They improve as new generators are added to training, but the underlying difficulty does not vanish. Any detector will always have inputs it reads wrongly, which is why the score comes with a heatmap rather than a verdict.

The short version

Detectors read how an image was rendered, so anything that re-renders a real photo can push it toward a false flag. Screenshots and upscales are the usual culprits. Read the heatmap, ask where the file came from, and get the original before anyone draws a conclusion.

Check an image free with GPTOne, just a free account, and use the region view rather than the score alone. For written work, the free AI text detector runs at 99.99% accuracy on every major model. Both at gptone.me.