Newsrooms & fact-checking
Breaking stories bring in photos and videos from readers, eyewitnesses and social feeds. Some of them were generated. Our API scores each file for AI generation, giving verification desks one more check to run before a picture is published or a claim is rated.
99.5%
of AI images caught with only 1% of real photos flagged, across 19 current generators
96.2%
still caught after WhatsApp, Instagram and re-share compression
0.987
AUROC on AI video from 12 video models
Hidden split of our October 2026 benchmark, never used in training. Method and full results
When a story breaks, editors want the picture now. Verification has to fit into minutes, and a generated image that slips through can cost a newsroom its credibility.
Eyewitness material usually arrives as a forwarded copy or a screenshot of a post. The first uploader may be unknown and the metadata stripped.
Current models get hands, text and lighting right far more often than they used to. Spotting the old tell-tale errors is no longer a reliable way to rule out a fake.
A probability that a photo was generated or synthetically altered, from a model trained on current generators and on copies that were compressed by messengers and social networks.
Frames are sampled across the clip and scored by a dedicated video model. The result shows how many frames were analysed and how many looked AI-generated.
When you only have a still from a clip, such as a thumbnail or a grab from a broadcast, it can be checked on its own with the video model.
If a file carries a C2PA manifest, it is parsed and its signature verified. You see who signed it and, where declared, which tool generated it.
Score the file through the API or in the web tool while you run reverse image search and trace the account that first posted it.
Set the threshold your desk is comfortable with. A high score is a reason to hold the image and dig further. A low score does not replace source verification.
Keep the score, frame counts and any content credentials with your verification notes, so editors and readers can see what was checked.
Check a reader-submitted photo
curl -X POST https://www.wasitaigenerated.com/api/v1/detect/image \
-H "Authorization: Bearer YOUR_API_KEY" \
-F "file=@reader-submission.jpg"Images, single frames and whole clips. See the API docs for responses, limits and pricing.
You can report it as one finding, alongside source tracing and reverse image search. A detection score is a probability, not proof, so a verdict on an image or on the person who shared it should rest on more than the score.
Yes. The model is trained on recompressed copies. On our published benchmark it still caught 96.2% of AI images after WhatsApp, Instagram and re-share compression, at a 1% false-positive rate.
Clips up to 4MB can be uploaded directly. Longer clips up to 50MB can be passed as a URL to a publicly reachable file.
Uploaded files are deleted after analysis; the API keeps the result, not the file. For sensitive material from sources, enterprise plans offer a data processing agreement and self-hosting on your own GPUs and servers.
We run paid pilots on a sample of your images and clips, calibrate the threshold to your tolerance for false positives, and report detection and error rates back before you commit. Volume pricing and an SLA are available for production, and the models can run on your own GPUs and servers.
Talk to us about a pilot