Given the legal reasons for not hosting nsfw content, video is a whole different beast compared to an image based website when it comes to reaching your user base using traditional or even automatic moderation techniques because as previously mentioned moderating videos automatically with some proposed NSFW AI can be super-duper effective but on top of the law aspect having streaming volumes in terabytes make you consider how these respective companies are jerks — if they have those kind of resources and data. Analyzing video is far more resource intensive and generally must be done in real-time with frame-by-frame accuracy to detect explicit content correctly. Current nsfw ai tools are 85% accurate in image-based detection but falter when it comes to video, with the accuracy reduced by around 10-15% because of issues such as motion blur, background noise and scene complexity. In the case of identifying sexual content in videos (specifically short or otherwise obscured explicit scenes) it is non-stop processing and frequently liable for either incorrect categorization results, or simply missing things completely.
Video content moderation is extremely important for high-traffic platforms such as TikTok or YouTube, where users upload around 500 hours of video per minute! Speed in processing matters a lot. Effective nsfw ai requires fast GPUs and large amounts of memory, but real-time processing is still scarce. According to industry reports, today's AI models on average require 2.5 seconds to process each second of video, which would lead to some time lag in content identification and moderation for long videos segments making it ineffective for these use cases. Such processing delays not only impact the accuracy of moderation, but also increase operational costs as companies have to scale their infrastructure up and down along with demand.
Watching Videos, The Context Issue: Video moderation is another field where only image-based NSFW AI can be effective as the understanding context in images is much simpler. Explicit content detection in video is not merely a matter of identifying visual signals, but also the processing actions and contexts unfolding over time. E.g., flagging explicit scenes and non-explicit contexts like medical or educational videos may receive false positives. Meta's video AI had a loud and vivid demonstration of its limitations in 2022: It flagged over 20% of educational content incorrectly because it couldn't hunt down the proper context required for an accurate decision.
Another limitation is that AI tends to rely on keyframes — these are actually specific frames extracted from the videos for analysis. Doing this method can work well to manage the load on your processing as single frames are analyzed, but it will also overlook any transitions from one frame color/scale/whatever variable you've chosen because they'll likely go undetected between two successive ones. Models trained on key frames lose up to 25% of accuracy in detecting explicit scenes for high-speed videos, because skipped frame. This constraint is another reason why platforms that use nsfw ai for video splice sampling instead of uses full analysis on every frame, meaning less accurate results.
When companies use nsfw ai for video moderation, they usually do AI detection first and then follow it up with a human review of the content that is difficult to process accurately by an algorithm. If AI has a question in the results, human moderators can intervene to avoid false positives and improve content quality. But it leads to, as I mentioned above this dual approach where the operating cost grows by an average of 30% which is pretty a lot for content-heavy platforms. That being said, video content moderation is inherently difficult to do accurately at scale and in real time with the same level of nuance as a human would — AI models are getting better every day but this continues to be an active area of research.
While proficient, there are technical obstacles to the real-time operation of nsfw ai for video detection that demonstrate an ongoing necessity for some human-sourced content review and optimization in binary image on/off modes as well improvements such as artificial intelligence processing power or organic context experience; it still has a long way to go.
Looking deeper into this topic, Check out nsfw ai — Exposing all the elements behind creating a nude bot.