How to Check If an Image Is AI-Generated
Evaluate provenance records, source history, visual inconsistencies, metadata, and corroborating evidence without relying on one detector score.
Use an evidence stack, not one detector
Check provenance
- No single visual clue or detector score can reliably classify every generated image.
Test source context
- Valid Content Credentials can provide positive provenance information, while missing credentials are inconclusive.
Treat artifacts as prompts
- The strongest assessment combines source history, technical evidence, visual inspection, and external corroboration.
Compare file-level evidence with the Image Comparison Tool, then follow the photo and visual verification learning path to test source, date, place, and caption as separate claims.
The question “Is this AI?” is often too narrow. An image can be camera-captured, generated, composited, retouched, or partly synthesized. Check whether Content Credentials provide a valid provenance record, then compare that record with the source history and visible claim.
Use the authenticity evidence stack
- Source history: identify the earliest credible publisher and what it says about production.
- Provenance credentials: check for valid signed assertions about capture or editing.
- Reverse search: look for earlier versions, source assets, or the same image with a different explanation.
- Technical inspection: examine dimensions, metadata, compression, and file history.
- Visual review: inspect geometry, text, reflections, repeated texture, lighting, and continuity.
- External corroboration: compare the scene with reliable records of the claimed event, person, or place.
Interpret Content Credentials correctly
C2PA defines a technical standard for cryptographically verifiable provenance assertions. A valid credential can show that assertions have not been altered since signing and can describe tools or edits. It does not make every underlying claim true by magic; trust also depends on the signer, capture process, and assertions included. A file without credentials may simply have passed through a system that removed them or never added them.
Use visual clues as questions
Malformed text, inconsistent reflections, impossible joins, repeated texture, and mismatched lighting can justify deeper checking. They are not universal fingerprints. Generators improve, ordinary editing can create similar artifacts, and low-quality recompression can make authentic images look strange.
Be cautious with detector scores
Detector performance depends on model family, image processing, and the data used for evaluation. A percentage displayed without methodology, calibration, or known limitations should be treated as a clue. Do not use it as the sole basis for accusing a creator or rejecting important evidence.
Distinguish synthesis, editing, and false context
An image can be camera-captured, fully synthetic, partly generated, conventionally edited, or authentic but attached to a false caption. These are different questions. A strange hand may suggest closer review, but it does not establish which process created the file. Compression, panorama stitching, portrait mode, and ordinary retouching can also produce odd edges and textures.
Start with provenance. Locate the earliest credible publication, the account or organization responsible for it, a higher-quality file, neighboring images from the same event, and any available Content Credentials. Reverse search can reveal that a suspicious image predates the claimed generator or comes from a documented photo series.
Use an evidence ladder
- Verify the source, publication chain, and claim attached to the image.
- Inspect credentials and metadata while allowing for stripping or manipulation.
- Compare lighting, reflections, geometry, text, repeated texture, and object continuity at full resolution.
- Use detector scores only as one fallible signal and record the tool and file tested.
- Seek the original project, raw file, alternate frames, or creator statement when the stakes justify it.
A careful report says which evidence supports “likely synthetic,” “contains generated elements,” “conventionally edited,” or “provenance unresolved.” Avoid presenting a detector percentage as the probability that an image is fake. When the real problem is a false caption, address that claim even if the pixels are authentic.
Continue with How to Verify a Screenshot. A layer-by-layer screenshot verification workflow for social posts, messages, headlines, and image composites. For another evidence check within the same investigation, use How to Verify When an Image Was Taken or Published.