Inpainting lets anyone select part of a photo and have AI fill it in: remove a person, add a dent, change a sign. The result can look seamless. Error level analysis (ELA) is one of the oldest tools for finding those edits, and it still helps — if you read it correctly.
How ELA works
JPEG images lose a little quality each time they are saved. ELA re-saves the image at a known quality and measures how much each area changes. Areas that were saved the same number of times change by similar amounts. An area pasted in or regenerated later often stands out as brighter or darker.
How to read the heatmap
- Edges and fine texture are always brighter. Hair, text and sharp edges light up in real photos too.
- Look for blocks that don't follow the content — a flat wall with one bright rectangle is suspicious.
- Compare similar surfaces. Two parts of the same sky should look alike.
Where ELA fails
Heavily recompressed images (social media, messaging apps), PNG files and fully generated images give weak or misleading ELA results. A fully AI-generated image has no "original" to compare against, so it can look uniform. That's why VerifyAI combines ELA with metadata checks, provenance (see what C2PA is) and an AI review.
Try it
Drop a photo into the checker below. In VerifyAI, ELA runs in your browser and boxes the regions that stand out, next to the other evidence. Read the short definition in our ELA glossary entry.