Synthetic Data Detection
How Crowdee scores the likelihood that content was AI-generated, and where that signal appears in a verification run.
Alongside manipulation and authenticity checks, several verification pipelines screen content for evidence that it was generated (or substantially rewritten) by AI, rather than captured or written by a human. This is not a separate product — it's a stage inside the multimedia verification pipelines that produces a structured syntheticLikelihood score alongside the standard verdict.
Synthetic-content screening is a stage-level capability, not a standalone pipeline you select. It appears as one stage within verify-image-deep, verify-video-full, verify-audio-technical, and verify-text-coherence.
Where it runs
| Pipeline | Stage | What it screens for |
|---|---|---|
verify-image-deep | image-synthetic-screening | Diffusion/GAN generation artifacts, missing camera EXIF fields, content-credential (C2PA) tags |
verify-video-full | video-visual-deepfake-screen | Face-swap, background replacement, and AI video-generation encoding signatures |
verify-audio-technical | audio-prosody-analysis | TTS/voice-clone indicators — pitch uniformity, missing breath sounds, vocoder artifacts |
verify-text-coherence | text-synthetic-screening | LLM stylometric tells — sentence-length uniformity, generic discourse markers, lack of personal voice |
Each of these stages runs automatically as part of its pipeline — there is nothing extra to configure to enable synthetic screening.
Output shape
A synthetic-screening stage produces the standard stage result fields (verdict, confidence, stageReport) plus two additional fields in artifacts:
| Field | Type | Description |
|---|---|---|
syntheticLikelihood | integer (0–100) | How likely the content is fully or partially AI-generated |
syntheticIndicators | string array | The specific pieces of evidence the stage cited |
{
"slug": "image-synthetic-screening",
"type": "ai",
"verdict": "synthetic",
"confidence": 82,
"stageReport": {
"summary": "Strong evidence of AI generation.",
"reasoning": "No camera EXIF fields present; C2PA generator tag found."
},
"artifacts": {
"syntheticLikelihood": 91,
"syntheticIndicators": [
"Missing lens/exposure/ISO EXIF fields expected for a genuine photograph",
"C2PA content-credential tag identifies a diffusion-model generator"
]
}
}Verdict mapping
When a synthetic-screening stage finds strong evidence, it sets its stage-level verdict to synthetic — see Verdicts & Confidence Scores for how this fits alongside authentic, manipulated, inconclusive, and unverified. The pipeline's final synthesis stage weighs this alongside every other stage's findings, so a high syntheticLikelihood on one stage does not automatically make synthetic the run's overall verdict — it's one input among several.
A synthetic verdict means the pipeline found strong evidence of AI generation — it does not by itself mean the content is malicious or should be rejected. Route the decision (label, reject, or allow) through your own editorial policy.
How is this guide?
Verdicts & Confidence Scores
Understanding the five verdict types, confidence scoring, and audit trails.
Web Research
How Crowdee performs live web search and crawling as part of source-tracing pipelines.