AI-generated image detection API

What it detects

The API detects AI-generated and AI-manipulated images, with a dedicated face-forensics model for deepfakes. Each response includes per-model explanations: which detector fired and the face box.

Measured catch and false-flag rates, each with its dataset and default thresholds, will be published here once the current test round is finished.

Where teams use it

• Trust and safety: flag AI-generated or swapped profile photos and uploads for review.
• Marketplaces and listings: catch AI-made product, property and review photos.
• Dating and social platforms: check profile photos at sign-up or on report.
• Newsrooms and fact-checking tools: add a second opinion to a verification workflow.

The request

• POST /v1/check with a multipart file, a JSON image_url, or JSON image_b64.
• Authorization: Bearer with a cb_live_ API key.
• Original bytes are scored as sent; nothing is resized or re-encoded on the way.
• POST /v1/check/batch takes up to 8 images per call (API Growth and above). GET /v1/usage returns checks used this period.

The response

• id: the check ID, which also links to its result page.
• band and is_ai: the verdict.
• fired: which detectors crossed their threshold (v0 for the whole-image analysis, face_specialist for the face analysis).
• reason: a one-line, human-readable explanation.
• scores and face_box: both detector scores and the location of the face that was checked.
• model and bands_version: the exact model and thresholds used, so every decision can be audited later.

Plans

• API Starter: $49 a month for 5,000 checks, then $0.012 per extra check.
• API Growth: $149 a month for 20,000 checks, plus the batch endpoint, then $0.009 per extra check.
• API Scale: $499 a month for 100,000 checks, then $0.006 per extra check.
• Extra checks are billed monthly. Enterprise plans start at $1,500 a month.

Data handling

Images are processed in memory and not stored, unless you turn on Store images for a key or send store=true with a request; then the image is kept, encrypted, for your plan's 30-day history. Each check keeps a SHA-256 hash of the file, the verdict and the scores, so results can be audited later.

Limits

• Images under 256 pixels on the short side, or JPEGs below quality 50, are refused with image_too_small or image_too_compressed. Refused images aren't billed.
• Maximum 20 MB per image. JPEG, PNG, WebP, AVIF, GIF, BMP and TIFF are accepted; HEIC isn't yet.
• Talking-head avatars from new engines are a weak spot: many are missed.
• Not for documents or ID forgery, audio, or liveness.

Questions