V50-MASTER+ID · Version 50 · Updated 2026-05-07
How Turing Verify reaches a verdict.
Every verification on Turing Verify runs through 50 forensic patterns organized into five categories. The verdict is a confidence-scored aggregate; the per-pattern breakdown is always returned alongside it, so reviewers audit the decision instead of receiving a black-box answer. Authored by Jeff Hu and the Turing Verify forensic team.
1. Typography Analysis
- — Font consistency and period-accuracy checks — fonts must match the era of the document.
- — Character spacing, kerning, and baseline regularity — modern forgeries often misalign these.
- — Font rendering artifacts — AI-generated documents typically leave detectable rendering traces.
- — OCR confidence scoring per text region — low scores in critical fields (name, date, GPA) flag tampering.
2. Visual Forensics
- — Seal and stamp analysis — shape regularity, ink distribution, pressure pattern.
- — Watermark integrity and placement — authentic watermarks are difficult to replicate.
- — Logo fidelity vs. known institutional assets — vector and raster comparison.
- — Background pattern consistency, color space, and print profile.
- — JPEG compression artifact analysis — detects splicing.
- — Error Level Analysis (ELA) — detects pixel-level edits.
3. Data Cross-Reference
- — Institution name verification against accreditation registries (CHEA, UNESCO WHED, Office for Students, MOE).
- — Signatory name and title validation — must match historical institutional leadership at the award date.
- — Date format and calendar consistency.
- — Credential numbering and serial format checks.
- — Grade/GPA scale validation specific to the issuing institution.
- — Accreditation status of the issuing body itself (catches accreditation-mill chains).
4. Metadata & Structural Analysis
- — PDF metadata inspection — creation tool, timestamps, modification history.
- — EXIF data analysis for photographed documents.
- — Document structure and layer analysis.
- — Digital signature and certificate chain validation.
- — File hash consistency checks.
5. AI-Generation Detection
- — GAN/diffusion artifact detection — signatures of generative models.
- — Synthetic text pattern recognition — fingerprints LLM-generated copy.
- — AI-generated image fingerprint analysis.
- — Statistical anomaly detection in pixel distributions.
Verdict taxonomy
Every verification returns one of five verdicts. The taxonomy is stable — if it changes, the change is published as a methodology revision and dated.
- VERIFIED
- High confidence the document is authentic — all forensic patterns align with the issuing institution's known templates.
- LIKELY_AUTHENTIC
- Forensic patterns align but one or more signals are inconclusive (low-resolution scan, unusual but not anomalous template).
- MANUAL_REVIEW
- Forensic signals are mixed; the document is escalated to a human forensic examiner with a per-pattern confidence breakdown.
- LIKELY_FAKE
- Forensic patterns indicate manipulation, forgery, or AI generation, but the verdict is not unambiguous.
- REJECTED
- High confidence the document is forged, altered, or AI-generated. The specific failed patterns are listed in the verdict report.
Read further
- Technical white paper → — full methodology including pattern-by-pattern detail.
- Glossary → — definitions for ELA, KYC, KYB, GAN, diffusion, and 35+ other terms.
- Direct answers → — atomic Q&A on document verification.