The Content Shield: How a Former Facebook Exec is Fixing AI Moderation

When Brett Levenson left Apple in 2017 to handle business integrity at Facebook, he jumped straight into the chaos of the Cambridge Analytica scandal. Back then, he believed better tech could fix the mess of content moderation. He quickly found out the problem was much deeper. Human reviewers had to memorize 40 page policy books that were poorly translated by machines. These workers had about 30 seconds to look at a flagged post and decide its fate. Levenson says the accuracy was only slightly better than a coin flip, and the damage usually happened long before a human ever saw the post.

That slow and reactive way of working just doesn’t work for the AI era. Modern AI can generate millions of images and chat logs in seconds, making human review impossible to scale. This frustration led Levenson to start Moonbounce. The startup just announced a $12 million funding round led by Amplify Partners and StepStone Group. Moonbounce aims to turn static policy books into “policy as code.” This means the rules are written directly into the software so they can be enforced instantly.

Moonbounce acts as a safety layer that sits between the user and the AI. The company trained its own large language model to read a customer’s specific policy rules and evaluate content in real time. It can make a decision in under 200 milliseconds. Depending on what the customer wants, the system can either block high risk content immediately or flag it for later review without slowing down the user experience. This speed is vital for apps that rely on instant interaction, like character roleplay bots or AI companions.

Currently, Moonbounce serves three main areas: social platforms, gaming companies, and AI app developers. They already support more than 40 million daily events and serve 100 million active users. Some of their biggest customers include the companion app Channel AI, video generator Civitai, and roleplay platforms like Dippy AI and Moescape. Even Tinder has started using these types of AI powered tools to improve how they catch bad behavior on their dating app.

AI companies are under massive pressure right now. Chatbots have been accused of pushing vulnerable users toward dangerous behavior, and image generators have been used to create nonconsensual imagery. Levenson believes that current safety filters are failing because they aren’t smart enough. Moonbounce uses a technique called “iterative steering.” If a user starts asking a chatbot about a harmful topic, the system doesn’t just give a blunt refusal. Instead, it modifies the prompts in real time to steer the conversation toward a safer and more helpful response.

Levenson runs the 12 person team with his former Apple colleague, Ash Bhardwaj. Together, they are building the infrastructure that lets companies treat safety as a product benefit rather than just a legal requirement. While Levenson knows his old employer, Meta, might want to buy a company like his, he is focused on keeping the tech available to everyone. He wants to ensure that safety tools aren’t just locked away by the biggest tech giants while the rest of the internet struggles to keep up.