In 2019, Brett Levenson departed Apple to lead business integrity at Facebook, a period marked by the extensive fallout from the Cambridge Analytica scandal. Initially, he believed that enhancing technology would resolve Facebook’s content moderation issues.
However, he soon realized the problem extended beyond mere technology. Levenson described how human reviewers were tasked with memorizing a 40-page policy document, poorly translated into their native languages. They then had approximately 30 seconds to assess each piece of flagged content, not only to determine policy violations but also to decide on appropriate actions, such as blocking content, banning users, or limiting distribution. These rapid decisions, according to Levenson, were only “slightly better than 50% accurate.”
Levenson shared with TechCrunch that the accuracy of human reviewers in applying policies was “kind of like flipping a coin,” and often, by the time a decision was made, the harm had already occurred days earlier.
Such a reactive and delayed approach is unsustainable when facing agile and well-resourced malicious actors. The proliferation of AI chatbots has intensified these challenges, leading to numerous high-profile content moderation failures, including instances where chatbots provided self-harm guidance to teenagers or AI-generated images circumvented safety filters.
Levenson’s frustration sparked the concept of “policy as code” – transforming static policy documents into executable, updatable logic directly linked to enforcement. This vision culminated in the establishment of Moonbounce, which, TechCrunch exclusively learned, announced on Friday that it has successfully raised $12 million in funding. The investment round was co-led by Amplify Partners and StepStone Group.
Moonbounce offers companies an extra layer of safety, applied wherever content originates, whether from users or AI. The company has developed its own large language model (LLM) to analyze a customer’s policy documents, assess content in real-time, deliver a response within 300 milliseconds, and execute a predefined action. Depending on the client’s preference, this action might involve Moonbounce’s system slowing content distribution for subsequent human review, or immediately blocking high-risk content.
Currently, Moonbounce primarily serves three sectors: platforms dealing with user-generated content, such as dating applications; AI companies developing characters or companions; and AI image generators.
Levenson reported that Moonbounce currently supports over 40 million daily reviews and serves more than 100 million daily active users on its platform. Notable customers include AI companion startup Channel AI, image and video generation firm Civitai, and character roleplay platforms Dippy AI and Moescape.
“Safety can indeed be a product advantage,” Levenson informed TechCrunch. “It’s never been perceived that way because it’s always an afterthought, not something you build into your product. Our customers are discovering innovative ways to leverage our technology to make safety a key differentiator and an integral part of their product narrative.”
The head of trust and safety at Tinder recently demonstrated how the dating platform utilizes these LLM-powered services to achieve a tenfold improvement in detection accuracy.
Lenny Pruss, general partner at Amplify Partners, stated that “Content moderation has historically plagued large online platforms, but with LLMs now central to every application, this challenge is even more formidable. We invested in Moonbounce because we envision a future where objective, real-time guardrails become the foundational backbone of every AI-mediated application.”
AI companies are encountering increasing legal and reputational pressures following accusations that chatbots have guided vulnerable users and teenagers toward suicide, and that image generators, such as xAI’s Grok, have been exploited to create nonconsensual nude imagery. Clearly, internal safety measures are proving inadequate, evolving into a significant liability concern. Levenson noted that AI companies are increasingly seeking external assistance to bolster their safety infrastructure.
Levenson explained, “We operate as a third party positioned between the user and the chatbot, meaning our system isn’t overwhelmed by context in the way the chat itself is. The chatbot might need to recall tens of thousands of prior tokens, whereas we are exclusively focused on enforcing rules at runtime.”
Levenson, along with former Apple colleague Ash Bhardwaj—who previously developed extensive cloud and AI infrastructure for Apple’s core offerings—leads the 12-person company. Their next major initiative is a capability called “iterative steering,” conceived in response to incidents like the 2024 suicide of a 14-year-old Florida boy, which was linked to his obsession with a Character AI chatbot. Instead of outright rejection when harmful subjects arise, this system would intervene in the conversation, dynamically modifying prompts in real-time to guide the chatbot toward a more actively supportive response.
“We aim to expand our actions toolkit to include the ability to steer the chatbot in a more positive direction, essentially taking the user’s prompt and adjusting it to compel the chatbot to be not just an empathetic listener, but a truly helpful one in such situations,” Levenson stated.
When asked if his exit strategy involved an acquisition by a company like Meta, potentially bringing his content moderation work full circle, Levenson acknowledged how well Moonbounce would integrate into his former employer’s technology stack, as well as his duties as CEO.
He added, “My investors would be furious if I said this, but I would dislike seeing someone acquire us only to restrict the technology. For instance, ‘This is ours now, and no one else can benefit from it.’”