Agency manages social for several brands. were processing ~200 user-generated content submissions weekly. need to verify nothing is AI generated before we repost. manual review doesnt scale. existing video detectors are mostly built for long-form forensic analysis (slow per-video). is anyone using a tool that can handle batches of 15-60 second clips at scale
from a technical angle, the short-form challenge is real because most video deepfake detection looks for temporal inconsistencies across many frames. 15-30 second clips dont give the detectors enough data. Hive Moderation has a video API thats designed for short clips and processes at near real-time speeds. Their accuracy on short-form is around 87% from independent tests Ive seen. Sensity AI has a similar API. Both have batch endpoints. For volume work: feed your batch through Hive first as a fast pass, then run anything flagged through Sensity as a second opinion. About 80% of submissions clear on the first pass. The remaining 20% need 30 seconds of API time which is workable for 200/week volume.
we use a similar pipeline for ecommerce UGC. hive first, sensity for ambiguous results, manual review only for the 2-3% that flag on both. cost is around 4 cents per video which is fine at our volume.
Editorial use case here too. The two-pass approach with two different vendors is the right pattern. Single-vendor pipelines fail when the vendor has a model update that introduces a blind spot. Two vendors with different training data is robustness insurance.
@ChloeCipher this is exactly what i needed. starting the hive integration today. ill report back on accuracy after 4 weeks of production data.