Five eyeball checks
Before reaching for any tool, scan the photo with these in mind:
Most manipulations are caught by the eye before any forensic tool runs. Photo editors are good at making individual elements look real but bad at matching every aspect of the surrounding scene — lighting direction, shadow softness, lens distortion, and chromatic aberration all have to agree across the composite, and inserted elements usually fail at least one of those.
- Shadows that don't match the light source.
- Reflections that show the wrong objects.
- Lens distortion that flips between parts of the image.
- Fonts on signs that look subtly wrong (kerning, weight).
- Pixel-level smudges or repeating patterns in skin or hair.
Error Level Analysis (ELA)
Run the freeimgtool Tampering Detector. It re-saves the JPEG and compares the error levels per region. Areas edited after the original save typically show different error levels and pop in the heatmap.
ELA is a guide, not proof. A photo saved many times by camera software produces false positives. Use it alongside the eyeball checks, not instead of them.
The principle behind ELA: every JPEG save loses a little quality, and the loss is roughly even across the image when the photo is saved once. When part of the image is edited and re-saved, that part has been through one more compression cycle than the rest. Re-saving the whole image at a known quality level reveals the difference — areas that were already heavily compressed show small changes, areas that were saved fewer times show big changes.
Caveats are real and significant. Phones that pre-process photos (HDR merging, computational photography, automatic enhancement) produce images where different regions already have different compression histories before any editing. ELA on these photos lights up everywhere, falsely suggesting tampering. ELA is most reliable on photos straight from a basic camera with no post-processing.
What to do when you find a manipulation
Don't share the suspect photo with the claim. Note the artifact you saw, link to the source, and let the reader inspect themselves.
Document your reasoning. A bald claim that a photo is edited isn't useful; a description of the specific anomaly (shadow direction inconsistent with the lit window, edge of inserted object shows JPEG ringing absent elsewhere in the photo) is. The goal is to give readers enough information to verify your conclusion themselves.
Check the source chain. The strongest signal that a photo is real is provenance — the chain of custody from camera to publication. If a photo first appears on an anonymous social account, treat it skeptically regardless of what your forensic tools say. If it comes from a wire service photographer with a verifiable byline, the prior is much stronger.
Beyond ELA: other forensic techniques
Metadata inspection. EXIF data records camera model, timestamp, GPS, and edit history. Modern editors often add a 'Software' tag (Photoshop, Lightroom, Snapseed). A photo claiming to be a candid news shot that has 'Photoshop' in its metadata deserves scrutiny — though absence of editing metadata isn't proof of authenticity, since editors can strip it.
Reverse image search. TinEye and Google Image Search find earlier appearances of the same photo. A photo presented as recent news that has been online for years is obviously not what it claims to be. This catches a surprising fraction of viral misinformation.
Noise pattern analysis. Camera sensors produce a unique noise pattern (PRNU — Photo Response Non-Uniformity). Forensic tools can match a photo to a specific camera based on this pattern. The flip side: edits that paste content from another photo break the noise pattern in the edited region.
Shadow and lighting analysis. Specialised tools project shadow directions back to estimate light source positions. Inconsistencies indicate composites. This is harder than ELA but more robust against re-saved photos.
Reflection geometry. The reflections in eyes, glass, water, and metal surfaces must match the scene. Tools that estimate the scene from reflections catch composites that survive ELA — including some AI-generated images, which often render reflections inconsistently.
AI-generated images: a new category
Photos generated entirely by AI (Midjourney, DALL-E, Stable Diffusion) don't have an original to compare against, so ELA is less useful. But they have their own tells.
Hands and fingers. Generative models still struggle with hands. Extra fingers, fused joints, and impossible grips are common giveaways in mid-quality AI images. High-end models have improved on this but not eliminated it.
Background text. AI-generated images often produce gibberish or near-gibberish text on signs, book covers, and logos. Real photos have readable, properly kerned typography.
Repeating patterns where they shouldn't be. AI sometimes produces wallpaper-like repetition in clothing patterns, brick walls, and crowd scenes. Real photos have natural variation.
Anatomical impossibilities. Ears at slightly different heights, eyes pointing in different directions, jewelry that connects to nothing. The model gets the overall shape right but misses small geometric consistencies.
C2PA content credentials. Some AI tools now embed cryptographic provenance metadata that identifies the generator and version. Looking for this metadata is the cleanest test when present, though it's far from universal.
When forensic analysis won't help
Heavily re-compressed images. By the time a photo has been screenshot, re-uploaded, and re-saved through three platforms, ELA can't distinguish the platform's compression from the original editing. The signal is gone.
Photos from phones with heavy computational photography. iPhone Night Mode, Pixel HDR+, and Samsung Scene Optimiser all do major post-capture editing before the file is even saved. ELA on these images is dominated by the post-processing, not by any later edits.
Cropping and color adjustment. ELA doesn't reliably detect edits that don't add new content — just cropping or shifting colors leaves no telltale compression boundary.
Sophisticated composites. A skilled editor who matches lighting, re-saves at the same quality multiple times, and flattens the result can produce a composite that defeats ELA entirely. Forensic analysis is biased toward catching careless edits, not careful ones.
Frequently asked questions
Will ELA catch every edit?
No. Edits that are flattened and re-saved as a fresh JPG can become invisible to ELA. ELA catches the easy cases.
Is the verdict 'high' a 100% accusation?
No. Treat it as a flag worth investigating, not a verdict.
Does ELA work on PNG images?
Not really. ELA depends on JPEG's compression artifacts. PNG is lossless, so re-saving produces no measurable change. For PNG forensics, look at metadata, reverse image search, and noise analysis instead.
Can ELA detect AI-generated images?
Sometimes, when the AI output has been re-saved as JPEG with regions of different compression history. More often, AI images need different techniques — looking at hands, text, reflections, and content provenance metadata.
Are there professional-grade tools?
Yes. Amped Authenticate, FotoForensics Pro, and forensic suites used by law enforcement go far beyond ELA. They're expensive and have a learning curve, but offer a much wider range of analyses.
What about deepfake video?
Different problem, different tools. Video manipulation detection focuses on temporal inconsistencies (frame-to-frame artifacts, lighting that shifts between frames) and facial micro-expressions. The single-image tools here don't cover video.