Skip to main content
Orilami

Fake or AI Image Checker: before and after

An illustration with sample files; your own files show their own results.

BeforeGenuine, or edited? You cannot tell

  • Receipt: real or edited?

  • Photo: untouched?

  • Picture: original size?

AfterEach one checked, with the evidence

  • Repainted total found

  • No signs of editing

  • Stretched from a smaller original

Fake or AI Image Checker: Editing, AI Markers and Receipt Checks

Checks photos, scanned receipts, IDs and documents for signs of editing, AI generation and re-compression, and reports what it finds as evidence with reasons, never as a percentage. It reads what the file says about itself (metadata, AI-source markers, signed Content Credentials), measures the pixels (error level, JPEG history, frequency, noise, areas copied within the picture) and, on request, tests whether a receipt’s text and numbers add up. All of it runs in your browser; the image never leaves your device.

What it checks

  • Provenance: AI-source declarations such as the IPTC Digital Source Type, generator names and the prompt data local generators write, plus Content Credentials (C2PA) verified offline. A declaration shows what the file claims; its absence shows nothing.
  • Metadata consistency: editing software, timestamps and GPS time that disagree, an embedded preview that no longer matches the picture, and a declared size that is not the real size.
  • Error Level Analysis, quantisation tables and block grid: where a pasted or re-saved region has a different compression history, and whether a JPEG was cropped or saved twice.
  • Frequency and noise: periodic peaks that expose resizing (and some generators), and a map of each area’s grain against the rest of the picture.
  • Copy-move: areas that match each other after a shift, rotation or scaling, drawn with lines joining the matches.
  • Text and numbers (optional): receipt lines against subtotal, tax and total, card, IBAN and passport check digits, impossible dates, QR codes against the printed page, and figures that differ from their neighbours in size, alignment or ink.

How the answer is reached

Every check reports a signal and how far it can be trusted on this particular image. Heavy compression, resizing, screenshots, chat-app copies, small images and printed pages are detected first and lower the reliability of the checks they blind, and the result says so. Correlated checks count once, independent layers are weighed together, and the outcome is one of five categories: declared as AI-generated, strong indicators of synthetic generation, suspected edited region, inconclusive, or no visible indicators.

Limits

Every technique has known false positives (edges and text raise error level, legitimate resizing creates spectral peaks, identical objects match each other) and a screenshot or print-and-scan wipes the compression history. The tool shows where to look; it cannot decide that an image is “real” or “fake”, and “no visible indicators” is not proof of authenticity. There is no bundled AI-detection model: no small in-browser model reliably flags modern synthetic images. Text recognition is English only and misreads digits now and then, so a failed check is a figure to compare with the original.

Frequently asked questions

Can this tell me for certain whether a photo was edited?
No, and no honest tool can. It surfaces traces that editing tends to leave: mixed compression levels, a broken 8×8 JPEG block grid, resampling periodicity, mismatched sensor noise, areas copied within the picture and metadata that disagrees with itself. Agreement across several independent signals is good evidence; a single anomaly often has an innocent cause.
What does "No visible indicators" mean?
None of the checks found a visible sign of AI generation or local editing, which is different from proof that the image is authentic. A careful edit, a clean re-render or a screenshot can leave nothing to find, so the result is never labelled "authentic". When the image has been compressed or resized too much for the pixel checks to say anything, the answer is "Inconclusive" instead.
Does it detect AI-generated images?
Partly, and the two parts are kept apart. A file can declare its origin: an IPTC Digital Source Type, a generator name in its software tags, the prompt data local generators such as Stable Diffusion front-ends write, or signed Content Credentials. A declaration is reported as a declaration, since anyone can add or strip one. The pixel-level signals are statistical hints only, and no detector model is bundled.
What are Content Credentials, and what does the check show?
Content Credentials (C2PA) are a signed record of how an image was made and edited. The tool reads one and checks, offline, that the signature and hashes still match the image: a record that fails means the image changed after signing. Because no trust list is consulted, it does not confirm that the signer is who they claim, and most images, genuine ones included, carry no credentials at all.
What does the copy-move search find?
Areas of a picture that match each other after a shift, rotation or scaling, which is what a pasted copy of part of the same image looks like. Matches must agree on one transformation and one set of pixels, and regular lattices such as tiles or fences are set aside. Genuinely repeated objects (two identical products, a logo printed twice) match too, and copies that were blurred or recoloured leave nothing to match.
What do the text and number checks do for receipts and IDs?
On request, the tool reads the printed text on your device and tests whether it adds up: a receipt's item lines against its subtotal, tax and total, card-number, IBAN and passport-zone check digits, impossible dates, a QR code against the printed page, and whether a figure differs from its neighbours in size, alignment or ink. A failure means a figure does not add up, which can be an altered number or a misread digit; it is not proof of forgery.
Is the image uploaded to a forensics service?
No. Every analysis runs in your browser on your own machine, including the optional text recognition, whose engine and English language data are served from this site (about 4 MB, fetched the first time you use it) rather than from a third-party host. The recognised text is shown on screen only and is left out of the PDF and JSON reports.
What is Error Level Analysis actually showing?
The image is recompressed at a known quality and compared with itself. Regions with a different compression history respond differently and stand out. An even response is expected for an untouched photo; a sharply different patch deserves a closer look.
Why does a screenshot, a chat-app copy or a scanned page look suspicious?
Because it was re-rendered or recompressed. Saving, resizing or sending through WhatsApp or Instagram rewrites the whole image's compression history and wipes the camera's grain, and a printed page has no sensor grain and repeats its own letters. The tool detects these conditions, lowers the reliability of the checks they blind and says so, instead of calling the file manipulated.
Why does an unedited photo show ELA hotspots?
ELA depends on edges and texture: sharp edges, text and saturated colours legitimately produce higher error. That is why the tool flags only clustered outliers well above the frame median, weighs them as weak evidence and labels them suggestive.
What does a misaligned block grid mean?
The 8×8 JPEG grid should sit at offset (0,0). A different offset means the image was cropped or shifted after compression; a strong competing grid suggests double compression or a spliced region. The check applies to JPEG files only, since a PNG or WebP has no JPEG grid to shift.
Can I analyse only part of an image?
Yes. Drag a rectangle over the area in the viewer and analyse it alone; the area is snapped to the JPEG block grid and judged against itself, so a small area gives the statistics less to work with. The file's metadata and Content Credentials are always read for the whole file, and the text checks need the whole image.
What does the PDF report contain?
The verdict category and its reasons, the file's SHA-256 so the exact original can be identified, the analysis time and tool version, the image with flagged areas outlined and the heatmaps laid over it, every piece of evidence by layer, what the metadata and Content Credentials say, the methodology and the limitations. It is available in English and Arabic; GPS coordinates and the recognised text are not printed.
Which image formats work?
JPEG, PNG and WebP, plus TIFF and HEIC/HEIF through in-browser decoders. Quantisation-table and block-grid analysis need a JPEG; error-level, noise, frequency, copy-move and text analysis work on all of them, and metadata reading covers JPEG, PNG, WebP and TIFF.
More questions about Orilami

Fake or AI Image Checker at a glance

Key facts about Fake or AI Image Checker: where it runs, what it costs, what it accepts and produces, and its limits.
Where it runsEntirely in your browser. Files are never uploaded.
Cost90 tokens per image (about $0.90 at the base rate)
AcceptsJPEG, PNG, WebP, TIFF, HEIC/HEIF
ProducesA verdict category with its reasons (never a percentage), the evidence behind it grouped by layer, an interactive viewer with error-level, noise, copy-move and frequency layers, what the file's metadata and Content Credentials say about it, optional receipt and ID text checks, and a PDF or JSON report.
BatchOne at a time.
Account neededOnly to spend tokens. Browsing and previewing are free.
What it does not doThe result is advisory evidence, never a verdict: nothing here proves an image is authentic, AI generation is judged from markers a file declares and statistical hints (no detector model is bundled), and text recognition is English only and misreads digits now and then.