Every AFIP determination is grounded in reproducible methodology. For the wider mission, see Independent Verification for the Age of Synthetic Media. We publish our detection frameworks, document false positive rates, and submit our work to peer scrutiny. Claims without methodology are not forensic science: they are marketing.
AFIP's detection methodology operates on a principle of independent corroboration. Rather than relying on a single analytical method, our framework combines multiple independent approaches, each capable of producing a standalone determination. When independent methods converge on the same conclusion, confidence is high. When they diverge, the case is flagged for deeper forensic review.
For written content, the framework applies three layers: statistical distribution analysis (token entropy, perplexity variance, n-gram frequency), linguistic pattern recognition (syntactic complexity, discourse structure, stylistic consistency), and provenance fingerprinting (metadata analysis, platform-specific markers, editing history reconstruction).
AFIP treats any single signal that text was AI-generated as a lead, not a finding. A conclusion is reported as corroborated only when independent kinds of evidence agree: the document’s history, the author’s knowledge of the content, and at least two detection methods that work in different ways.
We grade evidence on five levels. A single detector score is a lead. Agreement between two detectors using different methods is a supported lead. Text that conflicts with an author’s known writing or knowledge adds contextual support. Drafts, version history or file metadata that contradict the claimed writing process are process evidence. A finding is corroborated only when independent kinds of evidence agree and the author has had a chance to respond. The full scale, with what each grade can and cannot support, is set out in AFIP’s guide on whether AI detectors can be wrong.
A detector score cannot establish who wrote a document, whether AI use was permitted, how much a person edited a machine draft, or the writer’s intent. AFIP therefore never treats a score on its own as proof of misconduct, never reports a score without the conditions that affect it (text length, language and editing history), and never relies on one tool where a second method is available. OpenAI reached a similar conclusion about its own AI text classifier, which it withdrew on July 20, 2023, “due to its low rate of accuracy” (OpenAI, 2023). Independent testing of 14 tools likewise found them “neither accurate nor reliable” (Weber-Wulff et al., 2023).
For practical steps, see the guide to telling whether text was written by AI, why false positives add up at scale, and the research on how AI text detection works.
For audio content, the framework applies spectral analysis across multiple frequency bands, vocal biomarker identification (14 distinct biomarker categories), and temporal coherence analysis. The multi-resolution approach has proven effective against neural voice synthesis systems.
For visual media, the framework applies GAN fingerprinting (frequency-domain model attribution), facial consistency analysis (boundary, temporal, and physiological plausibility), and compression artifact analysis (encoding-specific signature detection).
All AFIP methodologies undergo internal validation before publication. We maintain chain-of-custody documentation for all analyzed samples, publish our false positive and false negative rates, and disclose known limitations of each detection approach. Peer review is not optional — it is foundational to forensic credibility.