KYC & onboarding
Fintechs, banks, marketplaces and gig platforms ask new users for a selfie, a photo of an ID document and sometimes a proof of address. Image generators can now produce all three. Our API scores each upload for AI generation, as an extra check next to the KYC and liveness provider you already use.
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
A selfie made by an image model shows a plausible person with natural skin, lighting and background. A reviewer comparing it to an ID photo has no obvious reason to doubt it.
Fraudsters generate photos of ID cards and passports that never existed, or ask a model to place a fake card on a desk under realistic light. The image looks like a phone snapshot of a real document.
Proof-of-address images such as utility bills and bank statements can be generated in seconds. When a manual team reviews these by eye, the volume makes careful checks hard to sustain.
A probability that the face image was generated or synthetically altered, from a model trained on current image generators and on images that were recompressed after upload.
The same check runs on the photo of a card or passport. It tells you whether the image itself looks AI-generated. It does not read the MRZ, inspect holograms or detect edits to a genuine document.
Photos and screenshots of bills, statements and letters can be scored as images, so a generated document is flagged before an agent spends time on it.
If an upload still carries a C2PA manifest, it is parsed and its signature verified. A valid declaration from the signer that the file is AI-generated is reported as certainty, not a score.
Send each selfie and document photo to the REST API alongside the call you already make to your KYC or liveness provider. The two checks answer different questions and run independently.
You decide which scores pass, which go to manual review and which block the application for now. Most teams route high scores to a reviewer instead of rejecting automatically.
Store the score and any content-credential details on the applicant record, so a reviewer and later an auditor can see why an application was escalated.
Score a selfie from an onboarding flow
curl -X POST https://www.wasitaigenerated.com/api/v1/detect/image \
-H "Authorization: Bearer YOUR_API_KEY" \
-F "file=@applicant-selfie.jpg"Images, single frames and whole clips. See the API docs for responses, limits and pricing.
Yes. Banks, lenders and payment providers use the same check on selfies, ID photos and images of supporting documents such as bank statements or payslips. It flags images that look generated or AI-edited. It does not read or validate the numbers on a document, so pair it with your existing document and AML checks.
No. We do not do liveness detection, face matching or document authenticity checks. The API answers one question: does this image look AI-generated? Use it next to your existing provider to cover synthetic uploads that other checks may not be built to catch.
A detection score is a probability, not proof. Use it to route an application to a person for review, and base the final decision on the full set of checks. Set your thresholds so that a high score leads to escalation instead of an automatic rejection.
On our October 2026 benchmark the image model caught 99.5% of AI images at a 1% false-positive rate across 19 current generators, and 96.2% after WhatsApp, Instagram and re-share compression. We recommend a paid pilot on a sample of your own uploads to set the threshold.
Uploaded files are deleted after analysis; the API keeps the result, not the file. A DPA, dedicated capacity and self-hosting on your own GPUs and servers are available for regulated teams. To try single images by hand first, use the free tool at /fake-id-detector.
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