Frequently Asked Questions

Fake Image Detector is a free online image forensics tool that has been running since 2020. Upload a photo and it is analysed through four independent layers: AI detection, metadata analysis, Error Level Analysis (ELA), and watermark and provenance checks. Within seconds you get a plain-language report on whether the image shows signs of manipulation or AI generation.

Open the home page, drag your image into the upload box or click browse to choose a file, then press Scan Now. The analysis runs immediately and the results page shows your original image beside its ELA map, together with the metadata verdict and any AI watermark or provenance flags.

JPG, PNG, BMP, TIFF, and WebP images up to 10 MB. Other formats such as HEIC or AVIF are not supported yet, so convert those to JPG or PNG first. Keep in mind that every conversion is itself an edit: whenever possible, scan the most original file you have.

Yes, it is completely free and there are no accounts or sign-ups. The site is funded by advertising and voluntary donations through Ko-fi, so image verification stays available to everyone.

Yes. The scanner runs in any modern browser on mobile, tablet, or desktop. There is nothing to install: open the site, upload an image, and scan.

No. Your image is analysed entirely in memory, is never written to disk, and is discarded the moment your results are generated. We do not keep, share, or reuse uploaded images for any purpose, and they are never used to train our models.

The scanner itself asks for nothing: no name, email address, or account. Like most websites, we use cookies for analytics and advertising; what they collect and how to opt out is described in our privacy policy.

No. Results exist only in your browser session. They are not saved on our servers, published anywhere, or shared with anyone. Once you leave the results page, the report is gone for good.

Four independent layers: AI detection with machine-learning classifiers, metadata analysis of the file's EXIF data, Error Level Analysis of its compression history, and watermark and provenance detection including C2PA Content Credentials. Each layer looks for a different kind of evidence, so together they give a far more reliable picture than any single technique alone.

ELA recompresses your image and measures how much each region changes. Areas that were pasted in, retouched, or generated separately often carry a different compression history from the rest of the picture, so they stand out in the ELA map with a noticeably different error level.

Digital photos carry hidden EXIF metadata: camera make and model, timestamps, GPS coordinates, and software tags. Editing programs such as Photoshop or GIMP often leave a recognisable software signature when a file is saved. This layer inspects those fields for evidence that the file has passed through editing software.

Machine-learning classifiers, including Local Binary Pattern Histogram (LBPH) models trained on curated datasets of authentic and manipulated images, compare the statistical texture of your image against learned patterns. AI generators tend to leave subtle fingerprints, such as unnaturally uniform noise, that these classifiers are trained to recognise.

It combines two checks. A frequency-domain analysis searches for imperceptible watermark patterns of the kind AI generators embed (such as Google's SynthID) along with other synthetic tells like missing sensor noise. A provenance check then reads any C2PA Content Credentials manifest in the file and reports whether it declares the image as created or edited by generative AI.

C2PA is an open provenance standard backed by Adobe, Google, Microsoft, OpenAI, and major camera makers. Compatible tools cryptographically sign a record into the image describing how it was made and edited. When your image carries one, we read it and report what it declares, including which tool generated the file.

Look at the layers together rather than any one in isolation. The results page shows your original image beside its ELA map, plus the metadata verdict and watermark flags. Agreement between layers is what matters: an image that trips several independent checks deserves far more suspicion than one that trips a single check.

Brightness in the ELA map shows where the error level differs. Edges, text, and high-contrast detail are naturally brighter, so brightness alone is not proof of editing. What matters is contrast between similar surfaces: if one face, object, or patch of sky glows while identical textures around it stay dark, that region may have a different compression history from the rest of the image.

No. A software signature proves the file was saved by that program, not that it was deceptively altered. Perfectly honest operations such as cropping, resizing, or colour correction leave the same trace. Treat it as one signal that becomes meaningful when other layers point the same way.

Not by itself. Social networks and messaging apps routinely strip metadata from every upload, and screenshots never had camera metadata to begin with. Absent metadata mostly means the image did not come straight from a camera; it is weak evidence in either direction.

Automated image forensics is probabilistic by nature. Combining four independent layers makes the result considerably more reliable than any single technique, but false positives (authentic images flagged) and false negatives (fakes missed) both happen. Stripped metadata, heavy recompression, and brand-new generation techniques all affect individual scans. Treat results as strong signals to inform your own judgement, not conclusive proof; our disclaimer explains how to interpret them.

No detector can. New models appear constantly and each leaves different traces; watermarks in particular exist only when a generator chooses to embed them. We retrain our models and update the pipeline continuously as generation techniques evolve, but a clean result never guarantees an image is authentic.

Each step away from the original destroys forensic evidence. A screenshot is a brand-new image with none of the original's metadata or compression history, and messaging apps aggressively recompress photos and strip their EXIF data. Whenever possible, obtain and scan the most original version of the file.

No. Results are informational and are not certified forensic evidence. They should not be the sole basis for decisions in legal, insurance, employment, or publishing contexts. If you need analysis that stands up in court, consult a qualified digital forensics professional. See our disclaimer for details.

First check the basics: the file must be a JPG, PNG, BMP, TIFF, or WebP under 10 MB. If it is, try uploading again, and then try another browser or device. Still stuck? Contact us with the image format, its rough file size, your browser, and what happened, and we will investigate.

Email contact@fakeimagedetector.com or use the contact page. Bug reports, feature ideas, and corrections to anything on the blog are all genuinely welcome; if we got something wrong, we want to fix it.

Three easy ways: a donation through Ko-fi helps cover the cost of running and improving the tool, sharing the site helps more people verify what they see, and following us on X or Facebook keeps you posted on new features. Neither ads nor donations ever influence analysis results.