Deepfake signals

/v1/deepfake_detection first checks the file for C2PA Content Credentials, then runs two detectors and reports everything in a signals object:

Signal	Looks at	Catches	Runs when
`ai_generated`	The whole image	Images made or edited by generative AI: diffusion models, GPT-style image models, AI inpainting. Works on any content: faces, documents, products, art	Always
`face_manipulation`	The largest face, cropped with some margin	Face swaps, AI face edits, lip-sync and reenactment frames, talking-head avatars	A face is found and it’s at least 64 px

The image is "real" only if neither signal flags it. The top-level genuine_score is the lower of the two, so decision and threshold work exactly as on the other endpoints.

Deepfake response

{
  "product": "deepfake",
  "decision": "spoof",
  "genuine_score": 0.141725,
  "threshold": 0.5,
  "spoof_type": null,
  "latency_ms": 640,
  "credits_remaining": 1238,
  "session_id": "a1b2c3d4e5f60718293a4b5c6d7e8f90",
  "request_id": "0f1e2d3c4b5a69788766554433221100",
  "signals": {
    "content_credentials": { "decision": null, "genuine_score": null, "status": "not_found" },
    "ai_generated":      { "decision": "real",  "genuine_score": 0.936734 },
    "face_manipulation": { "decision": "spoof", "genuine_score": 0.141725,
                           "status": "checked", "face_box": [71, 100, 136, 136] }
  }
}

That example is a face swap: the photo as a whole looks camera-made, but the face was replaced.

Fields

Field	Type	Meaning
`signals.*.decision`	`"real"` \| `"spoof"` \| `null`	This signal at your threshold. `null` when it didn’t run
`signals.*.genuine_score`	number \| `null`	Calibrated like the top-level score: `0.5` is the boundary, higher is more genuine
`face_manipulation.status`	`"checked"` \| `"no_face"` \| `"face_too_small"` \| `"skipped"`	Whether the face check ran. `"skipped"` (also on `ai_generated`) means Content Credentials already decided
`face_manipulation.face_box`	`[x, y, w, h]`	The face that was judged, in pixels of the image as sent (no EXIF rotation). Absent when no face was found

Content Credentials (C2PA)

ChatGPT, Gemini (Nano Banana), Adobe Firefly and a growing list of other generators sign the images they make with C2PA Content Credentials: a cryptographically signed record of how the file was made. We check for one before running the models. If a valid credential says the image was generated or edited with AI, that settles it: the image is flagged with genuine_score: 0 and the models don’t run.

Decided by Content Credentials

"genuine_score": 0.0,
"signals": {
  "content_credentials": { "decision": "spoof", "genuine_score": 0.0, "status": "ai_generated",
                           "generator": "Gemini", "signed_by": "Google LLC" },
  "ai_generated":        { "decision": null, "genuine_score": null, "status": "skipped" },
  "face_manipulation":   { "decision": null, "genuine_score": null, "status": "skipped" }
}

`content_credentials.status`	Meaning	Decides?
`ai_generated`	Signed credential says the image was generated by AI	Yes, flagged
`ai_edited`	Signed credential says AI was used to edit or composite it	Yes, flagged
`no_ai_claim`	Credential present (e.g. from a camera or editor), no AI involved	No, the models decide
`invalid`	Credential present but the image was altered after signing, or it can’t be verified	No, the models decide
`not_found`	No credential in the file	No, the models decide

generator and signed_by appear only when the signer chains to the official C2PA trust list (or the earlier Content Authenticity Initiative list), so they can’t be spoofed with a self-made certificate.
An AI image that was later cropped or adjusted in a non-AI tool still counts: the AI step stays in its signed history.
A credential can prove an image is AI-made, but never that it’s real, so content_credentials.decision is only ever "spoof" or null.
Credentials are fragile. Screenshots, re-saves, messaging apps and most social platforms strip them, so a missing credential tells you nothing. That’s why the models still run on everything else.

Images without a face

A face isn’t required. If there’s none, or the largest one is under 64 px, the face check is skipped and ai_generated decides alone:

"signals": {
  "ai_generated":      { "decision": "spoof", "genuine_score": 0.0417 },
  "face_manipulation": { "decision": null, "genuine_score": null, "status": "face_too_small", "face_box": [210, 144, 62, 62] }
}

If your flow needs a face (selfie onboarding, for example), check face_manipulation.status === "checked" on your side, or use /v1/unified_detection, whose liveness check rejects images without a usable face.

Thresholds

Your org’s deepfake threshold (default 0.5) applies to both signals at once. At 0.5, each signal sits at its identity-verification operating point: about 0.5% of genuine selfies rejected per signal, roughly 0.9% combined. A higher threshold catches more and rejects more genuine images. See Thresholds & scores.

Getting the best results

Send the original file. Resizing, re-compressing, screenshotting or rotating removes the fine traces the AI-image check reads, and strips Content Credentials.
JPEG, PNG, WebP, AVIF, BMP, TIFF and GIF work. HEIC doesn’t, so convert iPhone photos first.
Video: sample frames (one per second is plenty) and send each one. There’s no calibrated clip-level rule yet; flagging a clip when any frame is "spoof" is a reasonable start.

Known limits

Talking-head avatars and lip-sync from tools the model hasn’t seen are the hardest case.
Heavily compressed or downscaled real photos get flagged a little more often.
Faces under 64 px aren’t checked by face_manipulation. ai_generated still runs.

Getting started

Guides

Verification sessions

Fields

Content Credentials (C2PA)

Images without a face

Thresholds

Getting the best results

Known limits

​Fields

​Content Credentials (C2PA)

​Images without a face

​Thresholds

​Getting the best results

​Known limits

Fields

Content Credentials (C2PA)

Images without a face

Thresholds

Getting the best results

Known limits