The Deepfakes We Missed: We Built Detectors for a Threat That Didn't Arrive

Audio deepfakes have improved rapidly recently, yet their effect on human trust in real speech remains unstudied. We present the largest listening study on audio deepfake perception to date, collecting 35,532 judgments from 1,768 participants across 138 text-to-speech and voice conversion systems. Our central finding is a skepticism shift: compared to a 2021 baseline, human accuracy on fake samples barely changed (72.9% to 71.2%), but accuracy on real samples dropped from 72.7% to 64.1%. Participants are not worse at detecting synthesis artifacts; rather, they increasingly distrust authentic speech. Samples generated by commercial and autoregressive language model systems proved hardest to detect (61.3 - 65.9%), while those from traditional seq2seq and flow-matching models remain easier to spot (75.4 - 76.8%). An ML detector that served as a reference point maintained over 94.5% accuracy across all conditions. Our results suggest that the primary threat posed by modern deepfakes may not be mere deception, but the erosion of trust in genuine audio.

Deepfake detectors that achieve near-perfect scores on academic benchmarks collapse on real-world content: recent in-the-wild evaluations report AUC drops of 45-50% for state-of-the-art open-source models. We argue this gap is structural: static detectors are trained once against a moving generative frontier. We present BitMind Forensics (BMF), trained through Bittensor SN34, an open adversarial competition that continually refreshes the training distribution. We evaluate one dated export comprising image, general-video, and human-video checkpoints across nineteen public datasets: the canonical face-swap suites (FaceForensics++, Celeb-DF v1/v2/++, DFDC, DFD, UADFV, DF40) and recent in-the-wild and AI-generated-media benchmarks (Sumsub, Deepfake-Eval-2024, WildRF, Community Forensics, AIGCDetectBench, GenImage, AI-GenBench, AIGIBench, RAID, GenVidBench, GenVideo-100K). BMF reaches 0.936 AUC on Sumsub's original images and 0.872 pooled AUC over its full four-condition manipulation battery (1.4M images), staying robust under perturbation (0.855 JPEG, 0.799 downscaled), while GPEN enhancement improves detection (0.996). On Deepfake-Eval-2024, it matches the best commercial detector on images (0.915 vs 0.90) and exceeds it on video (0.822 vs 0.79), far above the best open-source detectors (0.56 and 0.63). It reaches 0.991 AUC on a 21-generator AI-image panel and 0.918 on GenVidBench, and exceeds the FF++-trained frontier on DFDC (0.947 vs 0.843) and Celeb-DF v2 (0.9985 vs 0.956), both contamination-audited, with statistical parity on Celeb-DF++. In a temporal study, successive dated exports improve on held-out media from generators absent from the static baseline's training (image 0.842 to 0.902; video 0.864 to 0.936). Our evaluation harness is public, and at publication the production API serves the exact evaluated snapshot for independent verification.

Deepfake detection research has largely converged on deep learning approaches that, despite strong benchmark performance, offer limited insight into what distinguishes real from manipulated facial behavior. This study presents an interpretable alternative grounded in bio-behavioral features of facial dynamics and evaluates how computational detection strategies relate to human perceptual judgments. We identify core low-dimensional patterns of facial movement, from which temporal features characterizing spatiotemporal structure were derived. Traditional machine learning classifiers trained on these features achieved modest but significant above-chance deepfake classification, driven by higher-order temporal irregularities that were more pronounced in manipulated than real facial dynamics. Notably, detection was substantially more accurate for videos containing emotive expressions than those without. An emotional valence classification analysis further indicated that emotive signals are systematically degraded in deepfakes, explaining the differential impact of emotive dynamics on detection. Furthermore, we provide an additional and often overlooked dimension of explainability by assessing the relationship between model decisions and human perceptual detection. Model and human judgments converged for emotive but diverged for non-emotive videos, and even where outputs aligned, underlying detection strategies differed. These findings demonstrate that face-swapped deepfakes carry a measurable behavioral fingerprint, most salient during emotional expression. Additionally, model-human comparisons suggest that interpretable computational features and human perception may offer complementary rather than redundant routes to detection.