IMAGE PROVENANCE GUIDE
How can I check whether an image is AI-generated?
A practical workflow for checking original files, provenance markers and heuristic scores without confusing clues with proof.
Keep the original and its context
Ask the publisher for the original file. Keep the publication link, acquisition date and any explanation of how it was created. Forwarded images, screenshots and exports may have lost provenance information. An unusual-looking image alone does not establish AI origin.
Read provenance evidence before supporting signals
Upload an image up to 15 MB to AIGCTrace. Review C2PA markers, readable metadata and tool-name matches first, then visible-watermark clues, blind-watermark clues, screenshot features and statistical signals. Text markers can be forged; compression and ordinary textures can produce pixel anomalies. These signals are not necessarily independent evidence.
Interpret scores and missing signals carefully
The 0–100 risk score organizes heuristic evidence. It is not a calibrated probability or accuracy measurement. A high score requires checking the underlying evidence. A low score means the methods found few clues, not that the image is a camera original. Skipped or unavailable checks are not passed checks.
Example: a tool name disappears in a screenshot
This is an illustrative example, not a measured experiment: an original PNG contains a generator name in a text chunk, while a screenshot no longer contains that chunk. A negative screenshot result does not refute the original clue. The original tool name still needs corroboration from publication records. Compare the evidence in both reports rather than just their totals.
Report evidence instead of certainty
A useful conclusion might say: “The file contains a tool-name marker; no signature was verified; a separate check suggests screenshot features.” Avoid turning that into proof of AI generation. For consequential decisions, seek confirmation from the source and use independent methods appropriate to the question.
Sources and scope
Product limits reflect the current project implementation. Illustrative examples are not accuracy benchmarks.