Tayanch tells real photos, video and audio from AI-generated ones — and says so honestly when it isn't sure. One API call.
Status is read live from the server, so the page never claims more than the API delivers.
Detection is probabilistic. A detector that only says “real” or “fake” will eventually accuse a genuine photo. Tayanch has a third answer.
Consistent sensor noise, optics, and compression history — the fingerprint of a camera. Returned as "verdict": "real" with the confidence behind it.
Heavily re-compressed, screenshotted, or genuinely ambiguous. Instead of guessing, the API says so and lets you route the file to human review.
Missing camera physics, a reconstruction error that fits a generator, an embedding off the real-photo manifold. Returned as "verdict": "ai".
Generator-disjoint evaluation, July 2026: the held-out generators and real-photo sources were never used in training. Signed Content Credentials (C2PA) are verified before any pixel is scored — cryptographic proof beats inference.
# one call, one verdict curl -X POST "$TAYANCH_API/analyze" \ -H "X-API-Key: $TAYANCH_KEY" \ -F "image=@photo.jpg" # → 200 OK { "verdict": "real", "label": "Real", "confidence": 96.4, "p_ai": 0.018 }
REST over HTTPS, JSON in and out. Send a file, or POST a link to /analyze/url and the server fetches the media itself. The base URL and your key are on the API page. Read the API page ›
Tuned per deployment: you choose the false-positive tolerance, the threshold follows.
Drop a file in the demo, or create a free API key — 100 checks, no card required.
