How our detection works

The scoring pipeline

Every scan runs through independent engines whose outputs are blended with confidence-based weights:

• ✓Commercial ML classifiers — transformer models trained on millions of human and AI samples, retrained continuously for new model generations (GPT-5-class, Claude, Gemini, Llama).
• ✓Statistical engine — our in-house stylometric analysis: sentence-length variation (burstiness), AI-typical phrasing density, formulaic transitions, contraction rate, paragraph uniformity, structural repetition and more. Zero-cost, fully explainable, runs on every scan.
• ✓Provenance inspector (images/video) — C2PA Content Credentials, IPTC digitalSourceType declarations, EXIF camera data, and generation parameters embedded by tools like Stable Diffusion and ComfyUI. Hard provenance evidence dominates the blend when found.
• ✓Frame analysis (video) — frames sampled evenly across the timeline, each scored by the image engines, aggregated with a peak-sensitive formula so short AI segments aren't averaged away.

What the score means

The 0–100% figure is a calibrated likelihood estimate that the content is AI-generated — not a measurement, and never a verdict. Verdicts map from score ranges: below 15% "very likely human", 15–40% "likely human", 40–60% "mixed signals", 60–85% "possibly AI-generated", above 85% "likely AI-generated".

Every result carries a confidence level (low / medium / high) computed from text length, engine agreement and language fit. A high score at low confidence is weak evidence — we say so explicitly rather than hiding it.

Known limitations (read this before acting on a score)

• ✓False positives are real. Independent research — including a 2023 Stanford study in Patterns — shows detectors disproportionately flag non-native English writers (61% average false-positive rate on TOEFL essays across seven tools tested).
• ✓Short texts are unreliable. Below ~120 words we mark results low-confidence; below ~30 words we won't score at all.
• ✓Edited and paraphrased AI text is harder to detect — the RAID benchmark (ACL 2024) documents sharp accuracy drops under paraphrase attacks for every tool tested.
• ✓Compressed images lose detection signal. Social-media re-encoding destroys both classifier features and provenance metadata.
• ✓Absence of metadata proves nothing. Platforms strip C2PA/EXIF routinely; only its presence carries evidential weight.

Our fair-use commitments

• ✓We never claim "99% accuracy" — no honest vendor can. Published accuracy numbers without fixed false-positive rates are marketing, not science.
• ✓Anonymous free scans are not stored; account scans are private, deletable, and never used to train third-party models.
• ✓Reports are designed for due process: confidence front and center, per-sentence/per-frame evidence, and explicit fair-use warnings.
• ✓We publicly document score interpretation so institutions can set fair thresholds. A detection score should start a conversation, not end one.