Guide
Can AI Image Detectors Get It Wrong?
Yes. A detector can label a real photo as AI-generated, or an AI-generated image as a real photo. The first is a false positive. The second is a false negative. One score does not settle which picture you have.
Check an imageA false positive is a real photo labeled AI-generated
A false positive, here, is a photograph from a camera that the detector labels Likely AI-generated. Take a phone photo of a lace curtain, or a portrait that has been through a beauty filter. A camera took the picture. The label still says Likely AI-generated. For that photo, the label is wrong.
This site has not published a false-positive rate. The example shows what the mistake is. It does not say how often Real Image Check makes it.
A false negative is a generated image labeled a real photo
A false negative, here, is an AI-generated image that the detector labels Likely a real photo. Take a generated scene that was saved again as a small JPEG, or photographed off a screen. A generator made the picture. The label says Likely a real photo. For that image, the label missed it.
This site has not published a false-negative rate either. A missed image and a flagged real photo are opposite mistakes, and a published rate for some other detector would not describe this one.
Why an automated detector can be wrong
A detector compares the file with patterns associated with generated images and with ordinary photos. It does not know who held the camera, and it does not know the occasion. Ojha, Li, and Lee found that classifiers trained to catch images from one kind of generator often missed images from newer generators they had not been shown. In their study, those unfamiliar images fell into the real class. That finding describes the classifiers they tested. It is not a measurement of Real Image Check.
The kind of picture matters too. A face, a product shot, and a night landscape are different problems. A detector can be more familiar with one of them than with the others. An image from a distribution the detector rarely saw can move the result even when the file itself is clean.
What a changed file can do
The file in front of you may already be a changed copy of some earlier picture.
- Compression, especially a low-quality JPEG, replaces fine edges with blocks and smear.
- Resizing throws away detail.
- Editing can add oddness, or paint over it.
- A crop removes the context around the subject.
- A screenshot is a new picture of a screen, often smaller, and with the display’s own scaling.
Grommelt, Weiss, Pfreundt, and Keuper found that, for the detectors they tested, JPEG compression made generated images more likely to be classified as ordinary photos. They also found that some of those detectors were using compression and image size as clues. Those results describe those detectors and the datasets in that paper. They are not Real Image Check’s accuracy, and they do not say which way a score on this site will move.
For what this checker can estimate from a screenshot, a resized preview, or a compressed copy, see Can an AI image detector check a screenshot?
A score is not a calibrated probability
The score on this site runs from 0.000 to 1.000. At 0.65 or above, the label is Likely AI-generated. Below 0.65, the label is Likely a real photo. Scores on the same side of 0.65 share that label. A result of 0.66 and a result of 0.99 receive the same label.
A result of 0.80 is not a statement that the image has an 80% chance of being AI-generated. The score is an estimate, not a probability or proof. This site has not calibrated it into a probability, and it has not published how often each score is right. The 0.65 line chooses the label. It does not guarantee a particular accuracy. For why another detector, or another version of the file, can show a different number, see Why does the same image get different AI detector scores?
One result is not proof
A single score cannot prove that a camera took the picture, and it cannot prove that a generator made it. The checker does not establish who made the image, prove the image is authentic, name a generator, or read a watermark or Content Credentials. It returns one score for the whole image.
What to check when the result matters
- Keep the earliest file you can. A chat preview, a crop, or a screenshot is a weaker copy. A changed file is a new file.
- Ask where it was published, and who had it first. How to tell if an image is AI-generated walks through that check.
- Inspect Content Credentials when the file still has them. The C2PA explainer describes those notes as signed statements bound to the file. A missing note does not mean the image is a real photo. Real Image Check does not read the note.
- Treat looking as a separate question. A strange hand or a clean face is a suspicion, not a verdict. Can you tell if an image is AI-generated just by looking at it?
- Write down the score, then the other evidence. Do not let the score replace the source.
How the two mistakes differ
- False positive. The image is a real photo, and the label says Likely AI-generated. The mistake accuses a real photo.
- False negative. The image is AI-generated, and the label says Likely a real photo. The mistake misses a generated image.
- Same cutoff, opposite errors. A score just above 0.65 and a score just below it can both be wrong. The label does not say which kind of mistake you are looking at, and it does not say how sure the detector is.
What Real Image Check estimates
The checker scores one JPEG, PNG, or WebP image, up to 20 MB. The score is calculated in your browser. The image is not uploaded for analysis. This browser keeps the preview, file name, and score until you clear that check or leave the page. Those details are not sent to the server. The server stores an operational event for the check, such as whether it finished, the plan, a short error code, the model version, how long it took, and a shortened visitor identifier. That event does not include the image, the preview, the file name, or the score.
The checker does not establish authorship, prove authenticity, identify the generator, or read watermarks or Content Credentials.
Questions
If the label says Likely AI-generated, is the photo fake?
No. That label can be a false positive. A real photo can land at 0.65 or above. The score is an estimate, not proof that a generator made the picture.
If the label says Likely a real photo, did a camera take it?
No. That label can be a false negative. An AI-generated image can land below 0.65. When the answer matters, go back to where the file came from.
Does 0.800 mean the detector is 80% sure?
No. The score is not a calibrated probability. A result of 0.800 and a result of 0.660 receive the same label, Likely AI-generated. Neither number is a chance of being right.
Can a screenshot or a new JPEG change the label?
It can. The file you check is the file that is scored, and a screenshot or a recompressed JPEG is a different file. This site has not published how often that change moves a result across 0.65. See the screenshot guide for what the checker can estimate from a changed copy.
What should I do if the decision matters?
Keep the earliest file, check where it was published, and inspect any Content Credentials it still carries. Real Image Check does not read those credentials, name a generator, or prove the image is authentic. One score is not enough for that decision.