How WebPurify optimizes video moderation

With the sheer number of user-generated videos now flooding the internet, keeping platforms safe with effective video moderation can sometimes feel like trying to hold back a tidal wave with a teaspoon. Platforms face an influx of dynamic, user-generated videos that can be anything from wholesome cooking tutorials to potentially harmful streams. The stakes are high: content moderation teams must maintain a delicate balance between speed, accuracy, and scalability to keep platforms safe while respecting user expression. But at WebPurify, we blend cutting-edge AI with human expertise to tackle the challenges of modern video moderation head-on.

WebPurify’s co-founder, Josh Buxbaum, explains, “The complexity of content has increased significantly in the 18 years we’ve been providing video moderation, and so have the guidelines and processes needed to manage it.”

In this blog, we’ll deep dive into how WebPurify optimizes video moderation, maintaining efficiency and accuracy while staying alert to new trends and remaining nimble enough to adapt our model on the fly as the nature and sophistication of digital content continues to evolve.

The biggest challenge with video moderation

Moderating video content is inherently more complex for humans and AI than moderating static images or text, simply because videos comprise thousands of frames, and each frame has the potential to violate community guidelines. In addition, some videos introduce a whole new challenging content type, audio. Josh explains, “The best approach is to utilize AI upfront to check all incoming videos, but for clients with thousands of videos being submitted daily, a significant live moderation team will be needed as well. To put it in perspective, 5k 5 minute videos per day equates to 416 hours of video requiring daily review. ”

With each project presenting its own set of risks, budget considerations, and client-specific guidelines, WebPurify customizes its approach for every client, ensuring that video moderation is both effective and scalable.

As Josh puts it, “You’ve got to remember, moderation isn’t just about volume; it’s about precision. The more we can filter out upfront, the faster we can work through what really needs human eyes.”

AI: the first line of defense

At the heart of WebPurify’s video moderation strategy is our powerful AI technology. “AI tools are our first line of defense,” Josh explains. When you have millions of videos per day, it’s physically impossible for human moderation teams to get through this volume. WebPurify’s AI tools are designed to quickly process large volumes of content by pulling frames at specific intervals — such as one frame per second, every five seconds, or even every 10 seconds — depending on the client’s needs.

As mentioned earlier, video moderation isn’t only about visual content; audio presents its own unique challenges. WebPurify tackles this by transcribing the audio and running it through our text moderation service. This approach allows us to identify harmful or inappropriate language, such as personal attacks, bigotry, sexual advances, or other sensitive content.

Our text moderation system is adaptable, enabling clients to specify words or phrases for customized blocklists that trigger flags. “Effectively, we’re operating two processes simultaneously,” Josh explains. “The AI scans video frames while audio is reviewed separately. Any potential offensive content detected in either stream prompts an escalation for human review to ensure comprehensive moderation of both visual and auditory elements.”

This flexibility allows our teams to adjust the level of detail in the review based on the platform’s risk tolerance and budget constraints. For example, a more cautious client might request a higher frame rate for greater accuracy, while others may opt for a lower rate to reduce costs without compromising too much on safety.

Our AI is trained to recognize a wide variety of content violations, such as violence, nudity, and hate symbols. Optical character recognition (OCR) technology also plays an important role by identifying text that appears in videos.

“OCR essentially pulls text from video frames, allowing us to analyze even the written elements in a video — things like on-screen profanity or contact info,” Josh explains. This is particularly useful for content where harmful messages aren’t explicitly spoken but are visually embedded within a video, such as signs, banners, or text on clothing.

Human moderation for more nuanced cases

While AI technology is vital for processing high volumes of video content, there’s still no substitute for human judgment when it comes to understanding context. AI can flag potential violations, but it’s up to WebPurify’s experienced human moderators to ensure that decisions are accurate and aligned with client expectations. This human layer is critical for making more nuanced decisions, especially when content falls into gray areas that automated systems can’t easily interpret.

The escalation process is key to maintaining the balance between speed and thoroughness in WebPurify’s moderation pipeline. When AI flags a piece of content, or if a human moderator encounters something particularly ambiguous or complex, it is escalated to a subject matter expert for further review. These experts bring specialized knowledge to the table, allowing them to handle sensitive or high-stakes content that requires more than a quick decision.

“The last thing we want is for a moderator to get hung up on one difficult video for 10 minutes, slowing down the entire workflow,” Josh explains. “Escalating it to an internal expert keeps things moving, while also ensuring the content is thoroughly reviewed by someone with the right expertise.”

Escalation not only ensures that tricky content is handled by the right people, but it also creates a feedback loop that benefits the whole team. When a piece of content is escalated, the moderator is later informed about the final decision and reasoning behind it, helping them learn for future cases. This continuous learning process improves the overall accuracy of WebPurify’s video moderation, allowing moderators to handle more complex content as they gain experience.

“It’s not just about moderating millions of images or videos — it’s about being smart with how you handle that volume,” Josh explains. “By filtering out easy decisions upfront, and escalating the challenging cases to experts, we ensure no time is wasted and the trickiest content gets the attention it needs.”

This system allows WebPurify to handle a vast amount of content without sacrificing quality or speed, ensuring every item flagged is given the appropriate level of review, while maintaining momentum across the entire moderation team.

Storyboarding: the power of visual summarization

One of our most effective methods for moderating videos is the use of storyboarding. This technique converts videos into a sequence of static frames, creating a visual summary that moderators can quickly scan. “Storyboarding turns a five-minute video into a series of images. You look at the storyboarded frames to determine the content and then quickly move on,” Josh describes.

The advantage of this approach is twofold: it saves time but also reduces the psychological toll on moderators. By avoiding real-time viewing, particularly for disturbing content, the team can maintain efficiency while protecting their mental health. Josh points out, “When you have a particularly disturbing video and it’s converted to just still images, it helps reduce the impact. You know what’s there, but it doesn’t affect you the same way as it does when seeing it in real-time.

“It’s not just about efficiency; it’s about reducing the impact of harmful content.”

Optimizing speed without sacrificing accuracy

WebPurify’s moderators will also adjust playback speeds as a time-saving measure. Moderators can view videos at accelerated speeds, such as 2x or 3x, allowing for quicker scanning. “You can process a 10-minute video in just over three minutes this way,” says Josh. “If something looks off, you can always pause and take a closer look.”

This technique, combined with storyboarding, ensures that WebPurify can handle the massive volume of video content efficiently without compromising quality. Moderators can switch between modes depending on the type of content, balancing speed with the necessary level of scrutiny.

Looping multiple videos for efficiency

At the heart of our video moderation process is our proprietary tool, Montage, which streamlines workflows for both AI and human moderators. “Montage was built for accuracy and speed,” Josh explains. Its design allows moderators to efficiently process content without leaving the interface, making quick decisions while handling vast amounts of video.

Key features include multi-video viewing, where moderators can see several short clips looping on a single screen, allowing them to accept or reject content with a single click. This is particularly useful for high-volume, short-form videos like animated GIFs. For longer videos, moderators can review at accelerated speeds or use storyboarding to quickly scan through frames as discussed earlier.

Montage also provides customizable moderation tools. The moderation screen can provide different labeling methods, such as radio buttons or direct clicks, depending on project needs. Moderators have immediate access to community guidelines and can use built-in text moderation tools to verify flagged words or phrases, even in different languages. This ensures accuracy while keeping the process fast.

Additionally, Montage includes a feedback loop that allows moderators to get insight from subject matter experts on how and why they made their decisions on content that was escalated. The platform also has what Josh calls a “two-ahead” feature that allows moderators to quickly revisit content they’ve just reviewed if they have second thoughts. The reason for this being that after clicking “accept” on a piece of content moderators oftentimes have second thoughts. By having to click two pieces of content ahead in the queue, it gives that moderator time to (often subconsciously)ruminate and ensures decisions can be revised without slowing down the queue.

As video content continues to grow in volume and complexity, WebPurify remains committed to staying ahead. “We’re always looking at what’s coming next,” Josh says. “From AI-generated images to GenAI video, the landscape is shifting. We’ve built our tools to adapt.”