witness
ILLUSTRATIVE PREVIEW OF PLANNED CAPABILITIES

See Witness in action

Try one of our examples to see how a Witness forensic report will look — multi-layer analysis from L0 verdict to L4 verifiable claims. Results shown are illustrative.

Try an example

Drop a file here or click to upload

Image (JPG, PNG) or video (MP4, MOV) up to 50 MB

DETECTION PIPELINE

Media in, forensic report out

Every scan runs an orchestration pipeline of independent detection methods. The output is a standardized forensic report with five layers of analysis.

Input
Image, video, or audio
Detection ensemble
Parallel model analysis
Signal aggregation
Weighted consensus
Narrative layer
Plain-language explanation
Forensic report
L0–L4 structured output
L0

Summary

Single verdict with confidence score and one-paragraph explanation.

L1

Overlays

Attention heatmaps and artifact density maps highlighting suspicious regions.

L2

Attribution

Likely generation method and model family identification with forensic fingerprints.

L3

Narrative

AI-generated plain-language explanation of findings, with hallucination caveat.

L4

Verifiable claims

Juror-checkable factual claims with measurable thresholds and evidence.

MULTI-LAYER DETECTION

How Witness analyzes content

Every scan runs multiple independent detection methods in parallel, then combines their signals into a single verdict with a transparent breakdown.

01

Diffusion fingerprinting

Detects noise residual patterns left by AI image generators like Stable Diffusion, Midjourney, DALL-E, and Flux.

02

Generative artifact scan

Identifies spectral artifacts and frequency anomalies from GANs, autoencoders, and neural network upsampling.

03

Pixel-level analysis

Examines neighboring pixel relationships and upsampling traces invisible to the human eye.

04

Temporal consistency

Analyzes frame-to-frame motion, facial dynamics, and temporal coherence across video clips.

05

Audio & lip-sync

Checks voice authenticity and phoneme-level lip-audio alignment to detect voice clones and lip-sync manipulation.

06

Provenance & metadata

Reads C2PA content credentials and EXIF metadata. A valid cryptographic signature beats any probabilistic score.

MULTIMODAL

Images, video, and audio

Images

Diffusion fingerprinting, generative artifact analysis, pixel-level forensics, face integrity checks, content provenance, and metadata analysis.

Video

AI video generation detection, face-deepfake analysis, temporal consistency, lip-sync verification, audio voice analysis, and per-frame scanning.

Audio

Voice clone detection, spectral analysis, phoneme-level speech verification, transcript coherence, and micro-prosody analysis.

DEVELOPER & ENTERPRISE

Integrate detection everywhere

Witness is the verification layer for the AI era. A simple API call or MCP tool returns a real-or-AI verdict that your agents, pipelines, and applications can act on.

REST APIv1
curl -X POST https://api.witness.vision/v1/detect \
  -H "Authorization: Bearer wt_live_..." \
  -F "file=@suspect-image.jpg"

{
  "verdict": "synthetic",
  "confidence": 0.94,
  "modality": "image",
  "report": {
    "l0_summary": "Synthetic — 94% confidence",
    "l2_attribution": "Latent diffusion model",
    "l4_claims": [
      "Diffusion noise residual detected (KL: 0.85)",
      "No EXIF camera metadata present",
      "NPR score 0.91 (natural range: 0.00–0.25)"
    ]
  },
  "report_pdf": "https://api.witness.vision/reports/..."
}
MCP SERVERModel Context Protocol
// AI agents call Witness as an MCP tool
// Claude, GPT, or any MCP-compatible agent

> "Check if this image is AI-generated"

Tool call: witness.detect({
  url: "https://example.com/photo.jpg"
})

→ {
    verdict: "real",
    confidence: 0.96,
    summary: "Authentic — camera sensor noise
      profile confirmed, EXIF metadata present,
      no synthetic generation indicators.",
    report_url: "https://..."
  }

REST API

Single endpoint. Send any image, video, or audio file. Get a structured verdict with confidence score and full forensic report.

MCP Server

Give AI agents the ability to verify content. Any MCP-compatible model can call Witness as a tool and act on the result.

Python & JS SDKs

Native client libraries with type-safe responses, automatic retries, and streaming support for batch processing.

Webhooks

Asynchronous processing for large files. Submit media, get a webhook callback with the full L0–L4 forensic report when analysis completes.

Built for integration

Content moderation

Flag AI-generated uploads before they reach your platform.

Newsroom verification

Verify source media authenticity before publication.

Legal & compliance

Generate forensic reports with verifiable claims for evidence review.

Insurance & KYC

Detect manipulated documents and identity photos in submission pipelines.

AI agent workflows

Let autonomous agents verify content they encounter via MCP tool calls.

Social platforms

Label or flag synthetic content at upload time with real-time API checks.