Skip to main content
Orilami

Step-by-Step: Checking Photo Authenticity and Detecting Manipulation

By 4 min read
ForensicsImagesHow-to

Not sure whether a photo, receipt, or ID has been doctored? You can't always tell by eye, but an image carries traces of how it was made and edited, and the file itself often says something about its origin. This guide walks through checking a picture with on-device analysis. Nothing is uploaded, so even a confidential image stays on your machine.

Open the Fake or AI Image Checker

Head to the Fake or AI Image Checker and drop your image onto the page. It loads and analyses locally in your browser. Keep in mind that forensics is advisory: the tool reports evidence with reasons, never a verdict, and it never calls an image "authentic".

Read the result

The top of the report gives one of five categories: declared as AI-generated, strong indicators of synthetic generation, suspected edited region, inconclusive, or no visible indicators. Below it are the reasons, and a line headed Why confidence is limited whenever the image's condition (a screenshot, a chat-app copy, heavy compression, a small size, a printed page) made some checks less reliable. Inconclusive is a common and honest answer. No visible indicators is not proof of anything: a careful edit or a clean re-render can leave nothing to find.

Check what the file says about itself

The Metadata and provenance panel lists the camera, software, dates and any location, and looks for AI-source markers: an IPTC Digital Source Type, a generator name, or the prompt data that local image generators write into PNG files. If the file carries signed Content Credentials, the tool checks offline that the signature and hashes still match the image. Treat these as claims: anyone can add or strip metadata, and the tool does not check the signer against a trust list. Most genuine photos carry no credentials, so their absence means nothing.

Look at the layers

The viewer lets you zoom, pan and switch layers over the original, with an opacity slider and a before/after slider.

  • Error level re-saves the image and shows where the difference is unusual. Regions that were pasted in or edited often compress differently from the rest, but sharp edges and text raise it too.
  • Noise vs. median compares each textured area's grain with the rest of the picture. Red is grainier, blue is quieter, and flat areas are left uncoloured. A pasted region carries its own grain.
  • Copy-move draws boxes around areas that match each other after a shift, rotation or scaling, with lines joining the matching points. Two identical products or a logo printed twice match too, so ask whether the scene would naturally contain two of them.
  • Frequency spectrum shows the regular peaks that resizing leaves behind.

Select an area and analyse it alone if you want to test one region. It is judged against itself, so a small area gives the statistics less to work with.

For receipts and IDs, check the numbers

Press Check text and numbers to read the printed text on your device (a few megabytes of text engine load the first time, from this site). It tests whether a receipt's lines add up to its subtotal, tax and total, whether card, IBAN and passport-zone check digits are correct, whether dates are possible, whether a QR code agrees with the printed page, and whether a figure differs from its neighbours in size, alignment or ink. A failure means a number does not add up: it may have been altered, or the text recognition may have misread a digit, so compare it with the original. The recognised text is shown on screen only and is not included in the exports.

Weigh the signals together

No single test proves manipulation. Treat the results as evidence: several independent layers pointing the same way is far stronger than one, and each Why this result may be wrong note describes a known innocent cause. When you need a record, the PDF report (English or Arabic) contains the verdict and reasons, the file's SHA-256, the heatmaps and the evidence by layer. Because the whole analysis runs on your device, you can examine sensitive documents without ever exposing them.

Sharing the image afterwards? Use the metadata scrubber to strip the camera, location and software tags first.

Tools for this

Fake or AI Image Checker

Spot edited or AI-made photos, receipts and IDs: AI markers, Content Credentials, ELA, Copy-move detection and Check digits.

90/ imageTry & Preview Free