Original research in
AI content authenticity
Benchmark studies, platform audits, and methodology analyses. Published independently — no vendor relationships, no sponsored findings.
Published Studies
AI Humanizer Bypass Rates: 2025 Annual Survey
We tested 14 AI humanizer tools against 6 major AI detectors across 1,800 text samples. Bypass rates ranged from 23% to 91% depending on the pairing. No detector was robust against all humanizers. Average accuracy dropped 31 percentage points on humanized text versus raw AI output. Detectors relying primarily on perplexity showed the largest degradation. The study includes a breakdown by content category, showing that marketing copy and casual writing were easiest to humanize successfully, while academic and technical writing retained higher detectability.
Domain-Specific False Positive Rates in AI Detection
False positive rates — the rate at which AI detectors incorrectly flag human writing as AI-generated — vary dramatically by content domain. This study measured FPRs across eight writing domains for four major detection tools. STEM academic writing showed the highest FPRs (14-31%), while news journalism showed the lowest (4-9%). Non-native English speaker writing produced FPRs 1.8-2.4 times higher than native speaker writing across all tools tested. These findings have significant implications for academic integrity policy and content moderation decisions.
C2PA Adoption Across Major Platforms: 2025 Audit
The Coalition for Content Provenance and Authenticity (C2PA) standard provides a cryptographic mechanism for attaching verifiable provenance metadata to media files. This audit examined C2PA adoption across 23 major content platforms and tools — AI image generators, social networks, news organizations, and stock media libraries. As of Q4 2025, 11 of 23 audited platforms had deployed C2PA signing. Image generators led adoption (Adobe Firefly, DALL-E 3, Stability AI), while social platforms showed the lowest uptake. The audit includes a technical assessment of implementation quality across all adopters.
Voice Deepfake Detection: Benchmark 2025
This benchmark measured the accuracy of five voice authenticity detection tools across 600 audio clips spanning eight TTS systems and four voice cloning frameworks. Tools were tested on clips from ElevenLabs, Suno, Sora, Bark, RVC, and three professional voice cloning systems. Hive Moderation led with 88% accuracy and a 9% false positive rate. Resemble Detect achieved 79% accuracy. All tools showed degraded performance on emotional or character-acted synthetic voice, suggesting current detectors rely heavily on spectral artifacts that expressive TTS systems are beginning to eliminate. The study also measured detection latency, relevant for real-time moderation applications.
Burstiness Calibration Across 12 Text Genres
Burstiness — the variance in sentence-level perplexity — is a widely-used AI detection signal, but its calibration differs significantly across writing genres. This analysis measured burstiness distributions in 3,600 human-written text samples across 12 genres: literary fiction, genre fiction, academic STEM, academic humanities, journalism, marketing, technical documentation, legal writing, personal essays, business email, social media, and poetry. Burstiness thresholds optimal for academic writing produce 22% false positive rates when applied to technical documentation. The analysis proposes genre-adaptive calibration as a path to materially lower false positive rates across the board.
Contribute Research
We welcome collaboration with academic researchers working on AI content authenticity, detection methodology, and provenance systems. If you have findings relevant to this field that you would like to discuss, contact us via the glossary or our research email below.
research [at] aicontentauthenticity [dot] com