How Lizhi Group Uses AI and Large Language Models to Enhance Minor Protection and Risk Governance

YAN Chenfan: Using AI to Drive Platform Risk Governance and Protect the Minors Behind the Screen

From keyword rules to large language models, a senior data mining engineer at Lizhi Group is redefining how digital platforms identify — and protect — their underage users.

By Lauren Hinton

When a child slips on a pair of headphones and opens a voice-based social app, how is the platform on the other side of the screen supposed to tell whether the user topping up an account and sending virtual gifts is an adult — or a teenager hiding their real age? It is a seemingly small question that goes to the compliance bottom line of an entire industry, and one that weighs on countless families. As rules such as China's Regulations on the Protection of Minors in Cyberspace continue to tighten, a platform that fails to fulfil its duty to protect minors can face regulatory consequences ranging from the suspension of its gifting features to the shutdown of its live-streaming business. Identifying minors accurately amid vast, anonymous, real-time interactions has become an unavoidable challenge for the internet sector as a whole — and YAN Chenfan, a senior data mining engineer at Lizhi Group, is among those at the forefront of solving it with artificial intelligence.

YAN holds a bachelor's degree in computer science and technology from Jinan University, together with a second bachelor's degree in business English. She began her career in software development — first at Neusoft and then at YY — where, in the later years of her time at YY, she gradually moved into risk governance and data mining. After joining NASDAQ-listed Lizhi Group (Sound Group Inc.) in 2019, she has continued to deepen her expertise in this field, driven by a desire to use technology to tackle the problems in platform operations that she describes as 'hard but right'.

Before she became involved, identifying minors on the platforms was no small challenge. Early efforts relied largely on manually configured keyword matching, with reviewers then combing through flagged content case by case — an approach that was cumbersome to maintain, prone to missed cases and heavily reliant on manual labour, making real-time screening of huge data volumes impractical. At the same time, a sizeable share of users who lodged minor-related refund complaints had, in fact, already let slip their age during interactions on the platform; with the tools then available, however, those signals were hard to catch in time. And because some minors had completed identity verification using a guardian's ID card, age based on real-name records alone could no longer be relied upon.

The capability YAN went on to lead is an end-to-end minor risk-control system spanning multiple platforms across Lizhi Group, whose products have together drawn more than 200 million users worldwide. Moving beyond the old habit of 'looking only at the age on an ID card', she redefined the objective as determining 'whether the actual user is a minor'. At its heart, the system pairs multi-source feature fusion with large-model reasoning: on one side, it distils a set of stable, codifiable risk signals from users' behavioural patterns and multidimensional data; on the other, it draws on a large language model to interpret age-related cues that surface during interactions, with frontline expert knowledge written directly into the prompts so the model can reason holistically, much like a seasoned human reviewer. By combining offline label generation with real-time interception, the system can flag risks precisely at critical moments — topping up, identity verification or going live.

The results speak for themselves. Since going into operation, the system has steadily processed vast volumes of text generated across the group's apps, identifying suspected underage users at scale and performing strongly on core metrics such as recall and accuracy. It is no longer simply a model 'running in the background': it has become an indispensable AI tool within the company. The government-affairs team relies on it to handle regulatory enquiries, markedly cutting the time involved, while the content-review team has made it a fixture of daily work, using it routinely for refund checks and evidence gathering.

The ongoing refinement has also surfaced some telling insights. Among users who described themselves as minors, a notable proportion turned out to be registered as adults under real-name verification — direct evidence that 'a verified identity is not the same as a true age', and a key input for designing detection strategies. To balance accuracy with speed, YAN iterated repeatedly — introducing concurrent processing, switching models and fine-tuning prompts across multiple rounds — to sharply reduce the time needed to assess text in bulk. Throughout, working within a data-security and privacy-compliance framework, she built in safeguards so that every deployment stays firmly within compliance boundaries.

Risk control for minors cuts across business, engineering, operations and compliance teams. As a core technical contributor, YAN not only leads development work but has long championed cross-functional collaboration, keeping risk-control capabilities aligned with both shifting business needs and changing regulatory requirements. Looking ahead, she is turning her attention to multimodal recognition — spanning audio and imagery — and to the co-evolution of large models with traditional risk-control systems. Only when technology and governance move forward together, she argues, can the industry build a more stable, transparent and sustainable ecosystem for digital safety. And protecting every child behind the screen, in her view, is where it all begins.