Error level analysis explained: what ELA does, and why it is far weaker evidence than it looks
Error level analysis, usually shortened to ELA, is a technique that re-saves a JPEG image at a known compression quality and maps the difference between the original and the re-saved version. The theory is that regions edited and re-saved at a different point in the file's history will show a different error level from the rest of the frame.
Neal Krawetz introduced the method in 2007 and it became widely known through free web tools. It is probably the most misused technique in amateur image forensics, and if a claim file arrives with an ELA map attached as proof, that map is not proof.
How does error level analysis work?
JPEG is lossy. Every save discards information, and after several saves at the same quality an untouched region reaches a stable state where further saves change it very little. A region introduced later has been through fewer compression cycles, so in principle it changes more.
ELA renders that difference as a brightness map. Bright areas mean high error, dark areas mean low error, and the analyst is invited to read bright patches as candidate edits.
- Input: a JPEG. ELA cannot work on PNG, TIFF or any lossless format, because there is no compression error to measure.
- Process: re-save at a fixed quality, subtract, amplify the difference.
- Output: a false-colour map, interpreted by eye.
- Interpretation: entirely subjective. There is no threshold, no score, and no pass or fail.
Why ELA produces false positives routinely
Almost everything that happens to an image in normal life changes its error levels, and none of it involves fraud.
- Texture drives the result: high-detail areas such as text, edges, grass or gravel light up brightly on an ELA map because they always carry more compression error. Flat areas such as sky, painted panels or a plain wall stay dark. That is a property of the content, not evidence of tampering.
- Re-encoding flattens the signal: an image forwarded through WhatsApp, uploaded to a portal, or saved twice by a claims system has been recompressed. After a few cycles genuine edits stop standing out at all.
- Global edits leave no boundary: brightening, a filter, a crop and re-save affect the whole frame uniformly, so ELA shows nothing.
- Skilled edits can be flattened deliberately: re-saving the composite repeatedly brings the pasted region's error level in line with its surroundings.
- Different cameras and quality settings shift the baseline: two genuine photos from two phones will not produce comparable maps.
What practitioners say about it
In 2013 Krawetz used ELA to argue that a World Press Photo award winner was a composite. The organisers commissioned independent examination of the original files and confirmed their integrity.
Hany Farid, one of the field's most cited researchers, told Wired UK in May 2013 that error level analysis "incorrectly labels altered images as original and incorrectly labels original images as altered with the same likelihood". In 2015, after Bellingcat used ELA on satellite imagery relating to MH17, image forensics specialist Jens Kriese described the method as subjective, not based entirely on science, and a technique used by hobbyists. Krawetz's own position has consistently been that ELA produces an artefact map and that interpreting it is the user's responsibility.
ELA in a claims file: a worked example
A motor claim arrives with a photo of a cracked bumper. Someone runs ELA and the registration plate and the crack both glow white against a dark body panel. It looks damning.
It is not. Plates and cracks are the highest-detail regions in that frame, and detail is exactly what ELA brightens. Running the same analysis on an undisputed photo of the same car from the same phone will produce the same glow. That control test takes two minutes and it is the step almost nobody runs.
When is ELA worth running at all?
As a free, fast, first-pass prompt to look harder at a specific region, in the hands of someone who understands what it is measuring. A sharp rectangular boundary in the error map, in an area with no matching change in texture, is worth investigating.
It is not worth anything as a written finding. Never put an ELA map in a decision letter, a referral rationale, or anything that might reach an ombudsman or a court. It will not survive contact with a competent expert, and staking a decline on it turns a defensible position into an indefensible one. Under a claimant's right to explanation, "the compression map looked odd" is a poor reason to have to give.
What to use instead
- Cross-reference the file: does the image agree with the FNOL account, the damage estimate, the weather on the stated date, and the other photos supplied.
- Check for reuse: reverse image search catches recycled photos, which are far more common than skilled composites.
- Ask about provenance: how the file reached you is a more answerable question than whether it looks edited, and it is one a claimant can address directly.
- Request a retake: a guided recapture of the same damage, made inside your process, resolves more ambiguity than any analysis of the original upload.
- Escalate rather than decide: send it to someone with genuine forensic training, and treat everything before that as a reason to look. The full picture is in photo evidence in insurance claims and in shallowfake.