AI-generation signals
Compare three model outputs and the overall result. A nonzero output on a real photo is possible; the number is not a probability that the photo is fake.
Check an image for AI-generation signals, compare detector results, and inspect its recorded metadata.
or choose one from your device
JPG, PNG or WebP · Up to 12 MB
Image analysis
Model outputs are signals, not the chance an image is fake.
Relative matches within the trained generator set. This does not identify a creator or prove which tool was used.
Metadata can be edited or removed. Camera fields and timestamps are not authenticated; Content Credentials signatures are not verified here.
Choose, drop, or paste a JPG, PNG, or WebP image.
Select Check image and review the detector results.
Explore the findings and available generator matches, then compare them with the original source.
A model response, a generator comparison, and recorded metadata answer different questions.
Compare three model outputs and the overall result. A nonzero output on a real photo is possible; the number is not a probability that the photo is fake.
For supported high-signal results, compare trained generator classes and families. A top match is relative to known options, not proof of which service created the image.
Inspect selected camera fields, timestamps, software markers, generation settings, and Content Credentials presence. These are observed file data, with their verification limits stated.
Swipe or scroll sideways to see every column →
| Result | Meaning | Useful next step |
|---|---|---|
| No clear AI-generation signal | The checks did not establish a strong overall AI-generation signal. | Review the source. This does not certify a photograph as authentic or unedited. |
| Inconclusive | The signals are intermediate, disagree, or do not support a clear classification. | Keep the distinction unresolved; do not choose whichever output confirms an assumption. |
| AI-generation signals found | The model comparison or recorded evidence supports further scrutiny. | Inspect the findings and conditional generator comparison, then seek independent evidence. |
Camera fields, software names, and credentials can add context to the detector results.
Swipe or scroll sideways to see every column →
| Observed data | What it supports | What it does not prove |
|---|---|---|
| Camera fields or capture timestamp | The file records those values. | That the values are genuine or that the pixels came directly from that camera. |
| Recognized AI software or workflow metadata | The metadata names a tool or contains a recognized generation-settings structure. | That the metadata is original or that every pixel was generated by that tool. |
| Content Credentials container | A recognized container is embedded in the file. | A valid signature or verified provenance chain; those are not validated here. |
| No recognized AI metadata | The inspected fields contain no recognized markers. | That AI was not used. Exports and social platforms can remove metadata. |
Use the closest available copy of the original. Screenshots add interface elements, while repeated exports can alter signals and remove metadata.
Swipe or scroll sideways to see every column →
| File | Fits the limits? | What to consider |
|---|---|---|
| JPG, 4,000 × 3,000 pixels, 4 MiB | Yes: 12 megapixels | No resizing is needed to meet these limits. |
| WebP, 8,000 × 4,000 pixels, 2 MiB | No: 32 megapixels | The pixel limit applies even when compression makes the file small. |
| PNG, 3,000 × 2,000 pixels, 14 MiB | No: over 12 MiB | Export a smaller copy if necessary; the new file may produce different outputs. |
| HEIC image from a phone | No: unsupported format | Export a JPG, PNG, or WebP copy. Keep the original for reference. |
Look for the original post or image file. Screenshots and repeated exports can remove useful context.
Read the overall signal alongside individual detector outputs and recorded metadata. Keep disagreements unresolved.
Verify the caption, date, and source separately. An image detector cannot confirm that an event happened as described.
Choose a supported image and select Check image. The tool compares three model outputs and inspects selected metadata. It reports no clear AI-generation signal, an inconclusive result, or AI-generation signals found. None of these is proof of how the image was made.
No. Model outputs are not calibrated probabilities, accuracy ratings, or percentages of edited pixels. A genuine camera photo can produce a nonzero output. The overall signal and individual outputs are shown separately so a single number is not treated as a verdict.
The available model signals do not support a clear conclusion under this tool's decision rules. Models may disagree or return intermediate outputs. Keep the result unresolved and review the original file and independent evidence.
When the overall AI-generation signal is high and attribution is available, the tool shows relative matches to the generator families and 27 generator classes represented in its attribution model. This is a conditional comparison, not verified identification. It cannot reliably name a generator outside its trained set. Other results show no reliable generator attribution.
The findings compare model outputs and report evidence recorded in the file, such as camera fields, software markers, or a Content Credentials container. They do not identify which pixels were edited or prove how the image was made.
No. EXIF camera fields, timestamps, and software names can be edited or removed. This tool detects recognized Content Credentials containers but does not verify their signatures or claims. Missing credentials or AI metadata do not establish authenticity either.
Use a still JPG/JPEG, PNG, or WebP up to 12 MiB and 24 megapixels. Images must be at least 64 × 64 pixels and no more than 16,384 pixels on either side. HEIC, GIF, SVG, animated PNG/WebP, and video files are not supported.
Supported screenshots and edited photos can be checked, but cropping, compression, overlays, and editing can affect the outputs. A separate face check uses models trained on historical face-swap and face-reenactment data when a face can be assessed. It is not a universal detector of modern face manipulation and does not affect the overall AI-generation result. Use the closest available copy of the original.