— For digital forensics specialists
Built for evidentiary standards.
Most detection tools are optimized to impress in a clean-image demo. Yours has to survive cross-examination on a screenshot of a WhatsApp image. That gap is exactly what Vestige Forensics is built around.
— The problem
Naive thresholds collapse under cross-examination
Run a public detector at its default 0.5 cutoff and it will flag a fifth of genuine messenger photos as AI, treating compression artifacts as evidence of synthesis. One inflated false-positive rate is all it takes to lose a court's trust in the method.
Measured on our benchmark (run full_v4)
27%
real photos flagged by a public detector @ 0.5 under a messenger re-encode
≤5%
our standard cap, by construction
≤1%
our strict cap for evidentiary work
Vestige verdicts never use 0.5. Thresholds are set on a named calibration run so the false-positive rate on real photographs is capped, and the cap is printed on the report.
— Capabilities
What the lab needs
Every feature below exists because a report has to hold up months later, in front of someone paid to break it.
FPR-capped verdicts
Standard (≤5%) for triage, strict (≤1%) for evidence. tpr@fpr1 is the number a defense expert will ask about, and it is already on the report.
Chain of custody
SHA-256 of the exhibit bytes, a hash of the report document, pinned pipeline versions, and deterministic re-analysis of identical bytes.
Condition-aware power
Detection rates are quoted under the image's estimated condition, including social_chain, the screenshot-of-a-WhatsApp-image that walks into your lab.
Expert-witness support
Detector provenance from peer-reviewed venues, calibration disclosures, methodology pack, and a limits section on every report.
A named checkpoint, not a black box du jour
The primary detector is a pinned fine-tuned checkpoint (sha-256 on record) calibrated on a named run. Nothing retrained between your intake and your testimony.
Zero retention, verifiable custody
Exhibits are analyzed in memory, never viewed by a human, and removed once the report is written, so there is nothing to subpoena, breach, or leak. Detector versions stay pinned by hash, so the same exhibit can be re-analyzed identically if the matter reopens.
— Workflow
Intake to testimony
01
Intake
The exhibit is SHA-256-hashed on arrival; every downstream artifact references that hash. Identical bytes always produce the same report id.
02
Analyze
A fine-tuned primary detector, pinned by checkpoint hash, scores under a pinned pipeline version; independent supporting signals corroborate. No tunable knobs to defend.
03
Report
Verdict at your chosen FPR cap, detection power under the estimated condition, limits, and a hash of the report document itself.
04
Testify
Methodology pack, detector provenance (CVPR 2023–2025), calibration disclosures, and the exact numbers a defense expert will probe, already on the record.
“At a 1% false-positive standard, the calibrated detection rate under this exhibit’s condition is 43%. A flag is strong evidence; the absence of one is not.”
— the kind of sentence our reports put on the record, with the calibration run named
— Further reading
Guides from our benchmark work
Written from published results, limits included.
How to tell if an image is AI-generated
Reverse image search, visual tells, provenance, then calibrated detection: the order that works, what each step misses, and the real numbers from our benchmark.
How accurate are AI image detectors? The measured answer
Vendors say 99%. Our benchmark measures 55 to 64% detection at a 5% false-positive cap, and 36 to 45% at the strict setting. Where the gap comes from.
Can AI-generated images be used as evidence?
Not legal advice. What a detector's verdict proves, the error rate a cross-examination will reach for, and the steps that keep a disputed image defensible.
AI image detector false positives: why real photos get flagged
A public detector flagged 27% of real messenger photos as AI at its default threshold. Why false alarms dominate when fakes are rare.
Does compression affect AI image detection?
Measured: detection falls from 64% on clean files to 55% after the WhatsApp-and-screenshot chain, at the same capped false-positive rate.
Is that AI image detector accurate? Judging a vendor's claim
A checklist for any AI detection tool: the four disclosures a real accuracy claim makes, and how to test one in twenty minutes with photos you already have.
AI image detection glossary: forensics terms, plainly defined
Thirty-five terms from AI-image detection and digital forensics in plain language: false positive rate, calibration, base rate, chain of custody.
Bring a defensible number to your next case
Run one exhibit free and read the full report: thresholds, limits, hashes and all. Then talk to us about lab plans and the methodology pack.