T
TrueFarmAI
Log inTry Free →
METHODOLOGY

How TrueFarm AI
detects synthetic content

Every analysis runs through three independent layers — two commercial AI detectors and our own forensic engine. No single model decides. The verdict is a weighted consensus.

DETECTION ARCHITECTURE

📁
Input
video / image
→
HIVE AI
Score 0–1
DETECTOR24
Score 0–1
→
FUSION
50% Hive
50% D24
+
FORENSICS
6 signals
weighted
→ Final verdict: verified_ai / strong_ai / mixed_signals / no_strong_signals
①
LAYER 1 — EXTERNAL DETECTOR

Hive AI

Hive Moderation is one of the most established AI content detection APIs in the industry. Their model is trained on millions of verified authentic and synthetic media samples across all major AI generation tools.

For videos, Hive analyzes frame-level artifacts, temporal consistency, and deepfake signatures. For images, it detects GAN fingerprints, diffusion model artifacts, and known generator patterns.

API endpointv3/ai-generated-and-deepfake
OutputScore 0–1 + model attribution
Video analysisFrame-by-frame + temporal
Image analysisPixel-level + metadata
Weight in fusion50%
②
LAYER 2 — EXTERNAL DETECTOR

Detector24

Detector24 (by Bynn Intelligence) is a specialized content moderation platform with 68+ AI models. We use their dedicated image AI detection and deepfake detection models in parallel.

This second detector is especially strong on AI-generated product photography, food images, and synthetic scenes where Hive alone has weaker coverage — exactly the type of content that's hard to detect.

Models usedai-generated-image + deepfake
Run modeParallel (both simultaneously)
Score takenMax of both models
Accuracy98.3% on image AI
Weight in fusion50%
③
LAYER 3 — PROPRIETARY ENGINE

Our forensic signals

Beyond the commercial detectors, we run six independent forensic analyzers. Each signal is classified into one of four classes with defined weights. They can confirm, strengthen, or even contradict the AI scores.

🔬Forensic analysis
→AI tool fingerprint in metadata
→Missing audio stream
→Re-encoding detected
→Empty camera metadata
→Unusual frame rate (8/16fps)
→Creation tool (Lavf, FFmpeg)
🖼Frame analysis
→Hive run on 5 individual frames
→AI frame count vs. total frames
→Max frame score tracked
→Mixed verdict detection
→Frame-level confidence scoring
⏱Temporal consistency
→SSIM between consecutive frames
→AI loop detection (avg SSIM > 0.99)
→Freeze frame detection
→Window-based consistency check
→Artifact window counting
🌊Optical flow
→Natural vs. unnatural motion
→Scene change frequency
→Motion level classification
→Suspicious jitter patterns
→Flow smoothness analysis
🏷Context analysis
→Explicit AI hashtags (#aiart etc.)
→AI keywords in title/description
→Suspicious account patterns
→Upload date correlation
→Known AI creator profiles
⚖Signal weighting
→HARD: 3.0 pts (one = strong_ai)
→MEDIUM: 1.5 pts
→SOFT: 0.4 pts (only if hard/medium)
→EXCULPATORY: −1.5 pts
→rawScore → final verdict threshold

Verdict thresholds

The final verdict is determined by counting hard/medium signal hits and the total rawScore, not just the AI percentage alone.

verified_ai≥ 2 hard signals OR rawScore ≥ 6
strong_ai≥ 1 hard signal OR rawScore ≥ 3 OR ≥ 2 medium signals
mixed_signals≥ 1 medium signal (no hard)
no_strong_signals≥ 2 exculpatory signals, hive score ≤ 0.25
not_reliableVideo < 3 seconds or quality warning

See it in action

3 free analyses — no account needed.

Try the detector →