Back to Articles
Detection Technology

The Waterfall Approach to AI Detection: Why Single-Model Systems Fail

By Kim BoherDecember 22, 20258 min read

The Problem with Single-Model Detection

Most AI detection systems rely on a single model to make a binary determination: is this content AI-generated or not? This approach has fundamental limitations.

High False Positive Rates: A single model optimized for detection sensitivity will flag legitimate content. A model optimized for specificity will miss actual AI content.

Cost Inefficiency: Running expensive deep analysis on every piece of content is wasteful when 80%+ can be classified with simple metadata checks.

Lack of Explainability: Binary outputs don't provide the forensic evidence needed for legal proceedings or insurance investigations.

The Waterfall Methodology

Our proprietary waterfall approach solves these problems with tiered detection:

Tier 1: Metadata & Provenance (~$0, <1 second)

Before any visual analysis, we check:

  • EXIF metadata for AI tool signatures (Midjourney parameters, DALL-E markers)
  • C2PA Content Credentials for declared AI generation
  • File hash against known AI-generated content database
  • Device fingerprinting to verify claimed capture device
  • Early Exit: If C2PA credentials confirm AI generation, we can return a verdict without expensive visual analysis.

    Tier 2: Visual Analysis (~$0.02, 2-3 seconds)

    Using fast multimodal models, we check:

  • Physics consistency across frames
  • Temporal stability (flickering, morphing)
  • Anatomical accuracy (hands, faces, proportions)
  • Texture artifacts (over-smoothing, repeating patterns)
  • Lighting consistency
  • Escalation Trigger: If confidence is below threshold OR if the deception score is high, we escalate.

    Tier 3: Deep Contextual (~$0.20, 10-15 seconds)

    Using large multimodal models with full context:

  • Cross-reference audio transcript with lip movements
  • Verify claims against known facts
  • Analyze source credibility
  • Generate detailed forensic report
  • Results

    Our waterfall approach achieves:

  • 96% accuracy (vs. 82% for single-model)
  • $0.03 average cost per analysis (vs. $0.25 for always-deep)
  • Court-admissible output with full methodology documentation
  • Implementation

    The waterfall router is configured per use case:

  • Quick Check: Tier 1 + Tier 2, 1 frame
  • Refund Fraud: Tier 1 + Tier 2, early exit on high confidence
  • Insurance: All tiers, Truth Bundle output
  • Evidence: All tiers, 30 frames, blockchain anchoring
  • Each tier's results are preserved in the Truth Bundle, providing a complete chain of custody for legal proceedings.

    Conclusion

    Single-model detection is fundamentally limited. The waterfall approach provides better accuracy, lower cost, and explainable results suitable for regulatory compliance and litigation.

    DeFake Uses This Approach

    The waterfall methodology powers every DeFake product — from refund fraud detection to court-ready forensic reports. The same architecture serves different use cases by configuring which tiers run, how many frames are analyzed, and what output format is produced.

    Whether you need a 1-second API response for refund fraud or a 30-frame deep analysis for court, DeFake's waterfall handles it.

    See all DeFake solutions →

    Share this article:

    DeFake - AI Detection Resources